正在加载视频...

视频加载失败

返回首页

Vijay Kedia ne kaunse 3 AI Stocks liye? 🤯 Data Centre Hidden Code Revealed #VijayKedia #AIStocks #DataCentre #StockMarketIndia #MultibaggerStocks

drmukulagrawal's profile picture

Dr. Mukul Agrawal

17,610 subscribers

14,363 次观看 • 4 个月前 •via X (Twitter)

#VijayKedia#AIStocks#DataCentre#StockMarketIndia#MultibaggerStocks
LIVE

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

Donald Trump Just Revealed One of the Biggest Investments Right Now🚨 AI and data centers are already crowded. The next bottleneck? Rare earth minerals & magnets. “I’ll tell you how to make money… do magnets.” The 3 stocks I’m watching: 🧲 $UUUU 🧲 $USAR 🧲 $MP Every EV, robot, AI server, missile, and fighter jet needs these materials. This supply chain could become one of the biggest winners of the decade. 👀 $MPP $XST $XRP
3:48

Donald Trump Just Revealed One of the Biggest Investments Right Now🚨 AI and data centers are already crowded. The next bottleneck? Rare earth minerals & magnets. “I’ll tell you how to make money… do magnets.” The 3 stocks I’m watching: 🧲 $UUUU 🧲 $USAR 🧲 $MP Every EV, robot, AI server, missile, and fighter jet needs these materials. This supply chain could become one of the biggest winners of the decade. 👀 $MPP $XST $XRP

Joshua Jake

48,646 次观看 • 20 天前

you're not trading 500 stocks you're trading 5 hidden forces, recycled on repeat quants cracked this decades ago using principal component analysis strip correlations across every ticker and what's left isn't 500 independent bets it's 5 latent signals: > broad market direction > growth vs value tilt > rate sensitivity > size premium > volatility regime every stock is just a weighted combo of those 5 so when a citadel desk builds a "diversified" book, they're not thinking in tickers they're thinking in factor exposures - "how much rate risk am I carrying, how much of my book is a value bet" Bookmark asap before it buried retail manages 500 variables. quants manage 5 PCA lives in every stats textbook, scipy builds it in 3 lines of code, data is public and always has been they kept you counting tickers while they were reading 5 forces underneath them
0:29

you're not trading 500 stocks you're trading 5 hidden forces, recycled on repeat quants cracked this decades ago using principal component analysis strip correlations across every ticker and what's left isn't 500 independent bets it's 5 latent signals: > broad market direction > growth vs value tilt > rate sensitivity > size premium > volatility regime every stock is just a weighted combo of those 5 so when a citadel desk builds a "diversified" book, they're not thinking in tickers they're thinking in factor exposures - "how much rate risk am I carrying, how much of my book is a value bet" Bookmark asap before it buried retail manages 500 variables. quants manage 5 PCA lives in every stats textbook, scipy builds it in 3 lines of code, data is public and always has been they kept you counting tickers while they were reading 5 forces underneath them

Livsun

23,865 次观看 • 1 个月前

Anthropic CEO Dario Amodei just revealed the hidden bottleneck that will kill most AI companies in the next 18 months (Save this). The insight comes from a principle in computer science called Amdahl's Law. Dario's argument is simple when something starts working really well inside an organization, you have to immediately ask what isn't working well around it. Amdahl's Law states that the maximum speedup of any system is capped by the fraction you haven't improved and that applies to companies just as brutally as it applies to processors. If you can suddenly write three or four times as many pull requests as before, you don't get three or four times the output but you rather get a pile of code no one can review, verify, or trust. The data makes this impossible to ignore. Teams with heavy AI coding adoption are merging 98% more pull requests but PR review time has ballooned 91%, deployment velocity is effectively flat and 96% of developers don't fully trust AI-generated code reaching production. AI generated code produces 1.7x more issues per pull request than human written code, 0.83 issues per PR versus 6.45. Veracode's 2026 State of Software Security report found that 82% of organizations now carry security debt, up 11% year over year, with critical security debt surging 36% in a single year driven directly by AI-generated code reaching production faster than security teams can handle. What Dario is describing is a systems problem, not a software problem and coding is roughly 20% of the software delivery cycle. Even at infinite coding speed, you're still bottlenecked by review, security, verification, testing, and deployment which make up the other 80%. The enterprises that win are the ones that identify which part of their system is the new constraint after AI accelerates the old one and fix that next. This is why Anthropic's Claude Code focuses on the full development loop, not just generation, and why the verification and security layer of the AI stack is where the next wave of enterprise value gets created. This is also why Anthropic as a company is positioned differently than most people realize. Anthropic's 2026 Agentic Coding Trends Report found that organizations using full-loop agentic coding workflows where AI handles not just generation but testing, review, and deployment validation reduced their software defect rates by 43% while increasing velocity by 2.8x. Claude Code now authors 4% of all GitHub commits and is on track to hit 20%+ by year-end, with the full-loop use case growing 3x faster than pure code generation. Dario has been building Anthropic around the exact insight he's describing publicly ,the constraint isn't writing code but rather everything that has to happen after.
0:48

Anthropic CEO Dario Amodei just revealed the hidden bottleneck that will kill most AI companies in the next 18 months (Save this). The insight comes from a principle in computer science called Amdahl's Law. Dario's argument is simple when something starts working really well inside an organization, you have to immediately ask what isn't working well around it. Amdahl's Law states that the maximum speedup of any system is capped by the fraction you haven't improved and that applies to companies just as brutally as it applies to processors. If you can suddenly write three or four times as many pull requests as before, you don't get three or four times the output but you rather get a pile of code no one can review, verify, or trust. The data makes this impossible to ignore. Teams with heavy AI coding adoption are merging 98% more pull requests but PR review time has ballooned 91%, deployment velocity is effectively flat and 96% of developers don't fully trust AI-generated code reaching production. AI generated code produces 1.7x more issues per pull request than human written code, 0.83 issues per PR versus 6.45. Veracode's 2026 State of Software Security report found that 82% of organizations now carry security debt, up 11% year over year, with critical security debt surging 36% in a single year driven directly by AI-generated code reaching production faster than security teams can handle. What Dario is describing is a systems problem, not a software problem and coding is roughly 20% of the software delivery cycle. Even at infinite coding speed, you're still bottlenecked by review, security, verification, testing, and deployment which make up the other 80%. The enterprises that win are the ones that identify which part of their system is the new constraint after AI accelerates the old one and fix that next. This is why Anthropic's Claude Code focuses on the full development loop, not just generation, and why the verification and security layer of the AI stack is where the next wave of enterprise value gets created. This is also why Anthropic as a company is positioned differently than most people realize. Anthropic's 2026 Agentic Coding Trends Report found that organizations using full-loop agentic coding workflows where AI handles not just generation but testing, review, and deployment validation reduced their software defect rates by 43% while increasing velocity by 2.8x. Claude Code now authors 4% of all GitHub commits and is on track to hit 20%+ by year-end, with the full-loop use case growing 3x faster than pure code generation. Dario has been building Anthropic around the exact insight he's describing publicly ,the constraint isn't writing code but rather everything that has to happen after.

Milk Road AI

52,190 次观看 • 3 个月前

I just built one of the greatest insider buying tracker tools of all time with Perplexity Computer. I wanted to find out one question: Which stocks actually have a high correlation of share price appreciation and insider buying? What Perplexity Computer built to answer this was truly amazing. Here's what it did: It pulled 1,301 real SEC Form 4 insider purchases across 184 S&P 500 stocks over the last 5 years. Then it tracked what happened to each stock AFTER insiders bought, measuring forward returns, win rates, and purchase frequency. It combined all of that into a single "Alpha Score" that ranks every stock by how reliably it goes up after insiders buy. The results? $COIN: 100% win rate. +187% average return after insider buys. $ET: 100% win rate. 52 purchases. $514M in total insider buying. $VST, $LUV, $CAT, $LLY, all 100% win rates. 73% of ALL insider purchases across the S&P 500 led to share price gains. But it didn't stop there. It also built: - A live purchase feed tracking every new SEC filing - Cluster buy detection (multiple insiders buying the same stock within 14 days) - A sector heatmap showing where insiders are putting their money - Top conviction buys (Berkshire's $2.1B OXY position, Musk's $1B TSLA buy) The whole thing looks like a Bloomberg terminal. Dark theme. Real time data. Fully interactive. I didn't write a single line of code. I just told Perplexity Computer what I wanted, and it researched the data, ran the analysis, and built the entire dashboard from scratch. This is the future of building things with AI. The best part was that it created a proprietary "Alpha Score" for every stock. This was a composite ranking from 0 to 100 that weighs four factors: 1. How often insiders bought (frequency) 2. How much money they put in (total value) 3. What percentage of buys led to gains (win rate) 4. How big those gains were (average forward return) The higher the Alpha Score, the stronger the correlation between insider buying and share price appreciation. It revealed insights you would've never expected. The energy sector ended up having the highest Alpha Score, meaning insiders buying energy stocks is very correlated with energy stocks surging. I have paid for data on insider buying in the past. I literally built a better tool with Perplexity Computer than the insider buying tools I have been using for the last 3 years.
0:24

I just built one of the greatest insider buying tracker tools of all time with Perplexity Computer. I wanted to find out one question: Which stocks actually have a high correlation of share price appreciation and insider buying? What Perplexity Computer built to answer this was truly amazing. Here's what it did: It pulled 1,301 real SEC Form 4 insider purchases across 184 S&P 500 stocks over the last 5 years. Then it tracked what happened to each stock AFTER insiders bought, measuring forward returns, win rates, and purchase frequency. It combined all of that into a single "Alpha Score" that ranks every stock by how reliably it goes up after insiders buy. The results? $COIN: 100% win rate. +187% average return after insider buys. $ET: 100% win rate. 52 purchases. $514M in total insider buying. $VST, $LUV, $CAT, $LLY, all 100% win rates. 73% of ALL insider purchases across the S&P 500 led to share price gains. But it didn't stop there. It also built: - A live purchase feed tracking every new SEC filing - Cluster buy detection (multiple insiders buying the same stock within 14 days) - A sector heatmap showing where insiders are putting their money - Top conviction buys (Berkshire's $2.1B OXY position, Musk's $1B TSLA buy) The whole thing looks like a Bloomberg terminal. Dark theme. Real time data. Fully interactive. I didn't write a single line of code. I just told Perplexity Computer what I wanted, and it researched the data, ran the analysis, and built the entire dashboard from scratch. This is the future of building things with AI. The best part was that it created a proprietary "Alpha Score" for every stock. This was a composite ranking from 0 to 100 that weighs four factors: 1. How often insiders bought (frequency) 2. How much money they put in (total value) 3. What percentage of buys led to gains (win rate) 4. How big those gains were (average forward return) The higher the Alpha Score, the stronger the correlation between insider buying and share price appreciation. It revealed insights you would've never expected. The energy sector ended up having the highest Alpha Score, meaning insiders buying energy stocks is very correlated with energy stocks surging. I have paid for data on insider buying in the past. I literally built a better tool with Perplexity Computer than the insider buying tools I have been using for the last 3 years.

Dividendology

184,326 次观看 • 6 个月前

I just built a self-improving second brain in Claude Code 🤯 A brain that runs your brand: every tool reads from it, it's wired to your live data, and it gets smarter every week. All running on the Claude Agent SDK. Perfect for DTC brands and agencies whose AI output sounds generic because every new chat starts from zero. If you're re-explaining your brand to AI every single time — re-pasting the voice guidelines, re-describing the customer you've described a hundred times, re-uploading the same positioning doc you uploaded yesterday, and still editing for an hour to strip out the generic phrasing... A brand second brain fixes the entire loop: → Build 3 foundation files once: brand DNA, voice, and customer → Every skill you create reads from them automatically → Wire in live data — your ad account, competitor ads, customer reviews → A weekly routine refreshes the brain with what's actually working → Every output comes back on-brand on the first pass No re-briefing AI on every chat. No hour of editing to undo generic phrasing. No brain that goes stale the week after you build it. What a second brain gives you: → The exact 3-file foundation that runs the whole system → The skill structure that makes every tool brand-aware by default → The live-data wiring that keeps it grounded in reality → The weekly self-improvement loop that keeps it sharp → The cold-start sequence to stand it all up from zero Built 100% in Claude Code. I put together the full playbook with the file structure, the wiring, and the exact setup. Want it for free? > Like this post > Comment "BRAIN" And I'll send it over (must be following so I can DM)
0:24

I just built a self-improving second brain in Claude Code 🤯 A brain that runs your brand: every tool reads from it, it's wired to your live data, and it gets smarter every week. All running on the Claude Agent SDK. Perfect for DTC brands and agencies whose AI output sounds generic because every new chat starts from zero. If you're re-explaining your brand to AI every single time — re-pasting the voice guidelines, re-describing the customer you've described a hundred times, re-uploading the same positioning doc you uploaded yesterday, and still editing for an hour to strip out the generic phrasing... A brand second brain fixes the entire loop: → Build 3 foundation files once: brand DNA, voice, and customer → Every skill you create reads from them automatically → Wire in live data — your ad account, competitor ads, customer reviews → A weekly routine refreshes the brain with what's actually working → Every output comes back on-brand on the first pass No re-briefing AI on every chat. No hour of editing to undo generic phrasing. No brain that goes stale the week after you build it. What a second brain gives you: → The exact 3-file foundation that runs the whole system → The skill structure that makes every tool brand-aware by default → The live-data wiring that keeps it grounded in reality → The weekly self-improvement loop that keeps it sharp → The cold-start sequence to stand it all up from zero Built 100% in Claude Code. I put together the full playbook with the file structure, the wiring, and the exact setup. Want it for free? > Like this post > Comment "BRAIN" And I'll send it over (must be following so I can DM)

Mike Futia

17,964 次观看 • 2 个月前

Man Group, the world's largest publicly listed hedge fund, $150B+ AUM, 40+ years of systematic trading, just revealed exactly how they're using AI agents to build trading strategies. this is their actual pipeline: agent 1: reads academic papers, analyzes datasets, reasons about whether there's an economic hypothesis worth testing. agent 2: writes the code to build the signal. agent 3: pulls the market data and runs the backtests. agent 4: evaluates the output of the prior agents. "there've been 15 to 20 models at last count that started as ideas generated by AI, went all the way through signal construction, the validation process, were reviewed and validated by a human investment committee, and deemed fit and proper to manage client assets." "this is very real. it's not make-believe at this point." a $150B hedge fund running a multi-agent pipeline that goes from reading a paper to trading live capital, with humans only entering the loop at the final approval stage. and they've been doing it for over a year and a half.
2:11

Man Group, the world's largest publicly listed hedge fund, $150B+ AUM, 40+ years of systematic trading, just revealed exactly how they're using AI agents to build trading strategies. this is their actual pipeline: agent 1: reads academic papers, analyzes datasets, reasons about whether there's an economic hypothesis worth testing. agent 2: writes the code to build the signal. agent 3: pulls the market data and runs the backtests. agent 4: evaluates the output of the prior agents. "there've been 15 to 20 models at last count that started as ideas generated by AI, went all the way through signal construction, the validation process, were reviewed and validated by a human investment committee, and deemed fit and proper to manage client assets." "this is very real. it's not make-believe at this point." a $150B hedge fund running a multi-agent pipeline that goes from reading a paper to trading live capital, with humans only entering the loop at the final approval stage. and they've been doing it for over a year and a half.

Goshawk Trades

38,079 次观看 • 1 个月前

A single tweet just vaporized BILLIONS from cybersecurity stocks. CrowdStrike down 8%. Cloudflare down 8.1%. Okta down 9.2%. SailPoint down 9.4%. The Global X Cybersecurity ETF just hit its lowest level since November 2023. What happened? Anthropic dropped Claude Code Security. It scans your entire codebase for vulnerabilities and suggests patches. Sounds boring. Until you read what it actually did: In testing, Claude found over 500 HIGH-SEVERITY BUGS in production open-source codebases. Bugs that had been sitting there for DECADES. Despite years of expert review. Despite fuzzing campaigns. Despite penetration testing. Despite million-dollar security audits. Nobody found them. An AI did. In hours. Here's the terrifying part: Traditional security tools work by pattern matching. They look for known vulnerabilities in a database. Claude doesn't do that. It READS code the way a human security researcher would. Traces data flows. Understands how components interact. Catches logic flaws that rule-based tools can't see. And it just outperformed every cybersecurity tool on the market. Combined. Wall Street figured this out fast. If an AI can find what your $500k/year security team missed... Why do you need the team? If an AI catches bugs that CrowdStrike, Okta, and Cloudflare couldn't... Why are you paying those subscriptions? Barclays came out saying the selloff was "illogical" and Claude "doesn't compete" with these companies. But here's what Barclays missed: It's not about what Claude competes with TODAY. It's about what it replaces TOMORROW. The SaaS apocalypse hit legal software 3 weeks ago. Thomson Reuters dropped 18% in one day. $285 billion wiped from software stocks. Now it's cybersecurity's turn. The pattern is obvious: Every industry that sells "expertise as a service" is about to get repriced. Legal research? Done. Code vulnerability scanning? Done. Compliance checking? Coming soon. Financial analysis? On deck. Companies that spent 20 years building "moats" around specialized knowledge are watching AI swim right over them. CrowdStrike is worth $95 billion. They have 30,000 customers. Claude just found 500 bugs their tools missed. Do the math. The smart money already sees what's happening. The iShares Expanded Tech-Software Sector ETF is down 23% YTD. Heading for its largest quarterly decline since the 2008 financial crisis. Not because software is dying... But because software companies that charge per-seat subscriptions for AI-replicable work are dying. Anthropic just proved that a general-purpose AI can outperform DECADES of specialized cybersecurity infrastructure. In a single product release. While still in "limited research preview." It's not even fully launched yet. What happens when it scales? We're about to find out.
0:50

A single tweet just vaporized BILLIONS from cybersecurity stocks. CrowdStrike down 8%. Cloudflare down 8.1%. Okta down 9.2%. SailPoint down 9.4%. The Global X Cybersecurity ETF just hit its lowest level since November 2023. What happened? Anthropic dropped Claude Code Security. It scans your entire codebase for vulnerabilities and suggests patches. Sounds boring. Until you read what it actually did: In testing, Claude found over 500 HIGH-SEVERITY BUGS in production open-source codebases. Bugs that had been sitting there for DECADES. Despite years of expert review. Despite fuzzing campaigns. Despite penetration testing. Despite million-dollar security audits. Nobody found them. An AI did. In hours. Here's the terrifying part: Traditional security tools work by pattern matching. They look for known vulnerabilities in a database. Claude doesn't do that. It READS code the way a human security researcher would. Traces data flows. Understands how components interact. Catches logic flaws that rule-based tools can't see. And it just outperformed every cybersecurity tool on the market. Combined. Wall Street figured this out fast. If an AI can find what your $500k/year security team missed... Why do you need the team? If an AI catches bugs that CrowdStrike, Okta, and Cloudflare couldn't... Why are you paying those subscriptions? Barclays came out saying the selloff was "illogical" and Claude "doesn't compete" with these companies. But here's what Barclays missed: It's not about what Claude competes with TODAY. It's about what it replaces TOMORROW. The SaaS apocalypse hit legal software 3 weeks ago. Thomson Reuters dropped 18% in one day. $285 billion wiped from software stocks. Now it's cybersecurity's turn. The pattern is obvious: Every industry that sells "expertise as a service" is about to get repriced. Legal research? Done. Code vulnerability scanning? Done. Compliance checking? Coming soon. Financial analysis? On deck. Companies that spent 20 years building "moats" around specialized knowledge are watching AI swim right over them. CrowdStrike is worth $95 billion. They have 30,000 customers. Claude just found 500 bugs their tools missed. Do the math. The smart money already sees what's happening. The iShares Expanded Tech-Software Sector ETF is down 23% YTD. Heading for its largest quarterly decline since the 2008 financial crisis. Not because software is dying... But because software companies that charge per-seat subscriptions for AI-replicable work are dying. Anthropic just proved that a general-purpose AI can outperform DECADES of specialized cybersecurity infrastructure. In a single product release. While still in "limited research preview." It's not even fully launched yet. What happens when it scales? We're about to find out.

Ricardo

110,899 次观看 • 6 个月前

.bolt.new is the second fastest-growing product in history—only behind ChatGPT. They hit $20M ARR just 60 days after launching the product, and ~$40M ARR (and 1 million DAU) just five months in 🤯 The craziest part is that they almost shut down the company. After seven years of building and iterating, they weren't getting anywhere. It turned out that what they'd been building was exactly what you need to build AI apps in the browser at scale. So they decided to give it one last shot. An overnight success, seven years in the making. In my conversation with Eric Simons (founder and CEO), we discuss: 🔸 Why Anthropic’s 3.5 Sonnet model was the critical breakthrough that made AI-generated code production-ready and unlocked the entire text-to-app market 🔸 How Bolt leverages WebContainer technology—a browser-based operating system developed over seven years—to create a dramatically faster, more reliable AI coding experience than competitors 🔸 How Bolt reached nearly $40M ARR and 3 million registered users in just five months with a team of only 15 to 20 people 🔸 Why PMs may be better positioned than engineers in the AI era 🔸 How AI will dramatically reshape company org charts 🔸 Why Eric lived at AOL’s HQ for many months 🔸 Much more Listen now 👇 • YouTube: • Spotify: • Apple: Thank you to our wonderful sponsors for supporting the podcast: 🏆 @Get_Eppo — Run reliable, impactful experiments: 🏆 Fundrise Flagship Fund — Invest in $1.1 billion of real estate: 🏆 OneSchema — Import CSV data 10x faster:
1:05

.bolt.new is the second fastest-growing product in history—only behind ChatGPT. They hit $20M ARR just 60 days after launching the product, and ~$40M ARR (and 1 million DAU) just five months in 🤯 The craziest part is that they almost shut down the company. After seven years of building and iterating, they weren't getting anywhere. It turned out that what they'd been building was exactly what you need to build AI apps in the browser at scale. So they decided to give it one last shot. An overnight success, seven years in the making. In my conversation with Eric Simons (founder and CEO), we discuss: 🔸 Why Anthropic’s 3.5 Sonnet model was the critical breakthrough that made AI-generated code production-ready and unlocked the entire text-to-app market 🔸 How Bolt leverages WebContainer technology—a browser-based operating system developed over seven years—to create a dramatically faster, more reliable AI coding experience than competitors 🔸 How Bolt reached nearly $40M ARR and 3 million registered users in just five months with a team of only 15 to 20 people 🔸 Why PMs may be better positioned than engineers in the AI era 🔸 How AI will dramatically reshape company org charts 🔸 Why Eric lived at AOL’s HQ for many months 🔸 Much more Listen now 👇 • YouTube: • Spotify: • Apple: Thank you to our wonderful sponsors for supporting the podcast: 🏆 @Get_Eppo — Run reliable, impactful experiments: 🏆 Fundrise Flagship Fund — Invest in $1.1 billion of real estate: 🏆 OneSchema — Import CSV data 10x faster:

Lenny Rachitsky

162,790 次观看 • 1 年前

Ex-Balyasny PM Ying Hua (Ying Hua) on why automation will increase demand for hedge fund talent, the quant/fundamental convergence, & why quant is blackjack but fundamental is poker. Ying Hua (PM @ Balyasny — built & led a quantamental team covering US insurance, capital markets & fintech | ~5 yrs @ Citadel running a long/short insurance book | Equity research @ Goldman Sachs | MS in Data Science @ UC Berkeley | Now founder & CEO of Implied Implied) "One of the best-kept secrets: fundamental investors are not good at sizing. Quant funds are really good at sizing." We cover: - The only real line between quant and fundamental: historical pattern matching vs. "how is this time different" — and the alpha neither group is looking at - Why she rebuilt her process so every model updated within 2 minutes of a print - Scraping highway patrol data from 15 states to track auto insurance losses live, every single day - The Malibu wildfire: mapping burned mansions from celebrity tweets to estimate losses before any industry consultant published a number - Her automation math: data gathering ~100% automatable, processing ~80%, judgment still 100% human - AI is quant for words — next-token prediction is pattern matching, which makes this just the next automation wave after quant and indexing - The proof differentiated views pay more: insurance stocks moved 2-3% on earnings in 2010; by the time she left, 15-20% intraday - Why "hook Claude Code up to data and let it rip" fails: BloombergGPT losing to a smaller open-source model, & why horizontal models are college grads - Quant is blackjack with card counting; multi-manager investing is poker — your hand, others' perception of it, your seat, everyone's stack - Most PMs are playing the wrong game: the positioning game hiding inside "fundamental" sectors with no new money coming in - Her hiring bar at BAM: every fundamental analyst learns Python — and the one skill she says can't be trained - The only two truly meritocratic jobs: hedge fund PM & sales Highlights: (00:00) Intro (00:40) How a quantamental PM actually puts on a position (02:05) The only real line between quant and fundamental (04:45) Why quantamental lowers the burden on your brain (06:40) Scraping 15 states of highway patrol data to nowcast insurance losses (09:25) The Malibu wildfire: estimating losses from celebrity tweets (11:55) How much of fundamental investing can be automated (13:45) Quantifying intuition: when a CFO's filler words jump 8% to 20% (16:25) The contrarian case: automation expands demand for talent (18:15) Earnings vol exploded — differentiated views pay more (20:25) Why Claude Code can't run your book (23:40) Horizontal models are college grads with no domain knowledge (30:50) Why chat is the wrong interface for investors (36:20) Will AI make markets more or less efficient? (39:10) Two things every fundamental PM should do today (41:50) The moat that expands: talent, redefined (45:50) Sometimes the game is positioning, not fundamentals (48:25) Blackjack vs. poker vs. surfing: matching the game to your horizon (52:00) Should young analysts chase the hottest sector? (57:55) Munger vs. Musk: two philosophies of wealth (1:01:05) Self-awareness in investing is bimodal (1:07:05) The only two truly meritocratic jobs: hedge funds & sales (1:08:50) The one skill for every regime: reconstruct the narrative
1:11:24

Ex-Balyasny PM Ying Hua (Ying Hua) on why automation will increase demand for hedge fund talent, the quant/fundamental convergence, & why quant is blackjack but fundamental is poker. Ying Hua (PM @ Balyasny — built & led a quantamental team covering US insurance, capital markets & fintech | ~5 yrs @ Citadel running a long/short insurance book | Equity research @ Goldman Sachs | MS in Data Science @ UC Berkeley | Now founder & CEO of Implied Implied) "One of the best-kept secrets: fundamental investors are not good at sizing. Quant funds are really good at sizing." We cover: - The only real line between quant and fundamental: historical pattern matching vs. "how is this time different" — and the alpha neither group is looking at - Why she rebuilt her process so every model updated within 2 minutes of a print - Scraping highway patrol data from 15 states to track auto insurance losses live, every single day - The Malibu wildfire: mapping burned mansions from celebrity tweets to estimate losses before any industry consultant published a number - Her automation math: data gathering ~100% automatable, processing ~80%, judgment still 100% human - AI is quant for words — next-token prediction is pattern matching, which makes this just the next automation wave after quant and indexing - The proof differentiated views pay more: insurance stocks moved 2-3% on earnings in 2010; by the time she left, 15-20% intraday - Why "hook Claude Code up to data and let it rip" fails: BloombergGPT losing to a smaller open-source model, & why horizontal models are college grads - Quant is blackjack with card counting; multi-manager investing is poker — your hand, others' perception of it, your seat, everyone's stack - Most PMs are playing the wrong game: the positioning game hiding inside "fundamental" sectors with no new money coming in - Her hiring bar at BAM: every fundamental analyst learns Python — and the one skill she says can't be trained - The only two truly meritocratic jobs: hedge fund PM & sales Highlights: (00:00) Intro (00:40) How a quantamental PM actually puts on a position (02:05) The only real line between quant and fundamental (04:45) Why quantamental lowers the burden on your brain (06:40) Scraping 15 states of highway patrol data to nowcast insurance losses (09:25) The Malibu wildfire: estimating losses from celebrity tweets (11:55) How much of fundamental investing can be automated (13:45) Quantifying intuition: when a CFO's filler words jump 8% to 20% (16:25) The contrarian case: automation expands demand for talent (18:15) Earnings vol exploded — differentiated views pay more (20:25) Why Claude Code can't run your book (23:40) Horizontal models are college grads with no domain knowledge (30:50) Why chat is the wrong interface for investors (36:20) Will AI make markets more or less efficient? (39:10) Two things every fundamental PM should do today (41:50) The moat that expands: talent, redefined (45:50) Sometimes the game is positioning, not fundamentals (48:25) Blackjack vs. poker vs. surfing: matching the game to your horizon (52:00) Should young analysts chase the hottest sector? (57:55) Munger vs. Musk: two philosophies of wealth (1:01:05) Self-awareness in investing is bimodal (1:07:05) The only two truly meritocratic jobs: hedge funds & sales (1:08:50) The one skill for every regime: reconstruct the narrative

Ethan Kho

244,449 次观看 • 1 个月前

🚀 The Segment Anything Model (SAM) has been upgraded to SAM2, featuring an efficient image encoder for segmenting images and videos. But does SAM2 outperform SAM1 in medical image and video segmentation? We're thrilled to present our paper "Segment Anything in Medical Images and Videos: Benchmark and Deployment"! We comprehensively benchmark SAM2 across 11 medical image modalities and videos. 📄 Paper: 💻 Code: **Highlights:** 1. SAM2 doesn’t always outperform SAM1 in 2D medical images, but excels in video segmentation, making it more accurate and efficient for 3D images, such as CT and MR scans. 2. MedSAM still outperforms SAM2 on most 2D modalities, but SAM2 surpasses MedSAM for 3D image segmentation in a slice-by-slice approach. 3. Segmentation performance varies with model size; sometimes the smallest model outperforms larger ones. 4. Fine-tuning SAM2 significantly boosts its performance for medical image segmentation. While SAM2 may struggle with challenging objects that have unclear boundaries or low contrast, it excels in generating good initial segmentation masks for common medical images and videos. However, the official interface doesn’t support medical data formats and has limitations on video length. To address this, we've developed a 3D Slicer Plugin and Gradio API for efficient 3D medical image and video segmentation. We invite you to try them out and provide feedback! 🔧 Deployment: - 3D Slicer Plugin: - Gradio API: (Note: Due to GPU limitations, the online API is available for only 12 hours and may be slow. We highly recommend deploying the Gradio API with your own computing resources: A big shoutout to Jun Ma (JunMa) who recently joined our UHN AI hub (UHN AI Hub) as Machine Learning Lead, and kudos to all co-authors: Sumin Kim, Feifei Li, Mohammed Baharoon (Mohammed Baharoon), Reza Asakereh, and Hongwei Lyu! This is true teamwork! Looking forward to collaborating with the community to advance 3D medical image and video segmentation foundation models! University Health Network U of T Department of Computer Science Department of Laboratory Medicine & Pathobiology Temerty Centre for AI in Medicine (T-CAIREM) Vector Institute #MedTech #AIinHealthcare #DeepLearning #MedicalImaging #SAM2 #MedSAM #AIResearch
0:22

🚀 The Segment Anything Model (SAM) has been upgraded to SAM2, featuring an efficient image encoder for segmenting images and videos. But does SAM2 outperform SAM1 in medical image and video segmentation? We're thrilled to present our paper "Segment Anything in Medical Images and Videos: Benchmark and Deployment"! We comprehensively benchmark SAM2 across 11 medical image modalities and videos. 📄 Paper: 💻 Code: **Highlights:** 1. SAM2 doesn’t always outperform SAM1 in 2D medical images, but excels in video segmentation, making it more accurate and efficient for 3D images, such as CT and MR scans. 2. MedSAM still outperforms SAM2 on most 2D modalities, but SAM2 surpasses MedSAM for 3D image segmentation in a slice-by-slice approach. 3. Segmentation performance varies with model size; sometimes the smallest model outperforms larger ones. 4. Fine-tuning SAM2 significantly boosts its performance for medical image segmentation. While SAM2 may struggle with challenging objects that have unclear boundaries or low contrast, it excels in generating good initial segmentation masks for common medical images and videos. However, the official interface doesn’t support medical data formats and has limitations on video length. To address this, we've developed a 3D Slicer Plugin and Gradio API for efficient 3D medical image and video segmentation. We invite you to try them out and provide feedback! 🔧 Deployment: - 3D Slicer Plugin: - Gradio API: (Note: Due to GPU limitations, the online API is available for only 12 hours and may be slow. We highly recommend deploying the Gradio API with your own computing resources: A big shoutout to Jun Ma (JunMa) who recently joined our UHN AI hub (UHN AI Hub) as Machine Learning Lead, and kudos to all co-authors: Sumin Kim, Feifei Li, Mohammed Baharoon (Mohammed Baharoon), Reza Asakereh, and Hongwei Lyu! This is true teamwork! Looking forward to collaborating with the community to advance 3D medical image and video segmentation foundation models! University Health Network U of T Department of Computer Science Department of Laboratory Medicine & Pathobiology Temerty Centre for AI in Medicine (T-CAIREM) Vector Institute #MedTech #AIinHealthcare #DeepLearning #MedicalImaging #SAM2 #MedSAM #AIResearch

Bo Wang

178,579 次观看 • 2 年前

Dear ICP community, the Internet Computer has now been running strong for 5 years 👏👏👏 Here is a celebratory preview of ICP "cloud engines," the sovereign frontier cloud technology the network shall soon provide from Main points: — Cloud engines enable anyone to spin up their own sovereign frontier cloud. The technology involves an extraordinary inventive step, in which cloud is created from a mathematically secure network of nodes. The nodes run as part of the Internet Computer network ( but are selected and configured by the cloud engine's owner. — The frontier cloud provided by engines is strongly focused on enabling AI agents to build and update online applications and services for us. The world is changing fast, and nearly all new online apps and services are already being built with the help of AI, and thus cloud engines target the future of cloud. — Software hosted on cloud engines is tamperproof, which means that it is immune to infrastructure hacks, because it runs inside a mathematically secure network protocol, rather than on computers directly. This means that AI agents, and those building with them, don't need to have a security team in the loop, or to trust someone else's security team. This is crucial, because in the future, non technical people will demand the freedom to build with full automation — where they just need to issue instructions to AI about what to build, and don't need to worry about anything or anyone else. Of course, apps and services running on engines are also vastly safer from the new breed of hacker being enabled by frontier AI. (The cloud engines themselves are also "tamperproof." Even if a hacker gains physical access to some portion of a cloud engine's nodes, and can make arbitrary changes, the computations and data of the hosted apps and services cannot be corrupted or interrupted so long as the network's fault bounds aren't exceeded. The recent hack of Vercel, a major cloud platform, which gave hackers access to the apps it hosted, provides additional perspective on the importance of this advantage.) — Software hosted on cloud engines is guaranteed to run, so long as a sufficient number of the engine's nodes are running. This means that AI can build applications and services without the need to have a human systems admin team constantly tinkering with the underlying platform to keep it running, which is again crucial, because in the future, non technical people will expect the freedom to use AI to build without the support of others. — New frontier programming language technology, in the form of the Motoko language developed by Caffeine Labs, leverages seminal "orthogonal persistence" technology that unifies program logic and data to deliver further unlocks for AI (Motoko is the first computer language being developed that targets agents that are writing software rather than humans engineers per se). Nowadays, AI can build and update production apps at a prodigious rate, even at the speed of conversation. But it can also make mistakes, and there's a risk that an update it creates might be "lossy" in the sense it causes some transformed data to be lost. Again, in this new world, it's both undesirable and impractical for everyone to have to have a systems admin team on-hand to detect lossy updates and roll them back, but Motoko provides a solution: it can detect new software updates are lossy before they are applied, reducing potentially catastrophic errors by AI to harmless coding retries. — Software hosted on cloud engines is "serverless" but unlike traditional serverless software, directly it directly incorporates data through "orthogonal persistence." Another key purpose is simplify backend software logic and fuel the modeling power of AI by increasing abstraction (sorry for the technical language!!!). Put simply, this enables AI to produce more sophisticated backends, faster, and at dramatically lower costs, as measured by the number AI API tokens consumed during coding. (Tip for the technical: orthogonal persistence is a new paradigm where "the program is the database," and data lives inside program variables, which is possible because it's as if hosted software runs forever in persistent memory). — An expanding database of skills at shall make it possible to develop and directly deploy apps and services to your cloud engines directly from Claude Code, Perplexity, Codex and other AI platforms. Further, your account on can be connected, so that new apps and updates created through conversation automatically appear hosted from your cloud engine. In the future, R&D is going to be very seamless. You converse with AI, and your secure and unstoppable apps or services are created or updated. Cloud engines are designed to directly support this "self-writing cloud" future where we can work hands-free. — Tech sovereignty is becoming a huge issue worldwide, with governments and corporations seeking to create sovereign tech stacks owing to geopolitical tensions. Increasingly, people are realizing that tech provided by foreign nations can come with hidden backdoors and kills switches, from the base platform, right up through hosted apps and services. ICP technology is open source, and those building on ICP using AI own their own source code. When you have the source code, you can verify that there are no backdoors, and when you own the source code thanks to AI, you can update it at will, freeing you from vendor lock-in. But cloud engines take sovereignty much further... — You create a cloud engine by selecting the nodes that will be combined. You can choose the class of nodes used, and their number, but more importantly, you can choose who operates the nodes, and where they are located. Almost any configuration is possible, because the Internet Computer scales the security privileges afforded to hosted software within the network according to configuration (software hosted on cloud engines can directly interoperate with software on other engines and traditional subnets, but base restrictions are applied according to security rules). A cloud engine can be created within a region such as Europe, to comply with regs such as GDPR, or completely within a sovereign state like Switzerland or Pakistan. But cloud engines go further still... — Sovereignty is also about freedom from vendor lock-in. Cloud engines are essentially ICP (Internet Computer Protocol) network configurations, and this means the underlying compute nodes they combine can be swapped out without interrupting their hosted apps and services. This is a big deal. In addition, cloud engines now support nodes that are instances running on Big Tech's clouds, in addition to nodes that are dedicated specialized hardware, as per the Gen I and Gen II nodes that dominate the Internet Computer today. For example, it is possible to have an engine running across different AWS data centers, say, and then reconfigure the engine to run across a mixture of AWS, Google, Azure and Hetzner for even more resilience, without the users of hosted apps and services noticing a thing. That's true freedom. — Sovereign AI is becoming increasingly important too, and cloud engines allow special "AI nodes" to be added to them, so that hosted software can perform inference on hardware provisioned by the owner from a location the owner has selected. Even though the AI nodes are only accessible within the cloud engine, they can still benefit from the forthcoming Internet Intelligence Gateway (IG), which will make it possible to validate inference performed on key frontier open weights LLMs, even when the inference is performed on completely independent AI clouds. When the results of inference are received, this technology can verify that neither the prompt+context (input) nor the inference result (output) have been modified, and that the results were produced by the precise LLM expected. This ensures that AI clouds don't cheat by running inference on cheaper models than are being paid for, and bad actors aren't modifying the inputs or outputs to surreptitiously insert advertising into results, say, or change facts, or insert malware when code is being generated. What's super cool about this technology is the cost of the verification is scalable. A very valuable additional security can be achieved with only 1-2% of extra cost. — Scaling apps and services when they hit capacity limits is another thorny problem that cloud engines help the world address. Engines make scaling possible without rewriting or reconfiguring software. The query workload capacity of hosted software can be horizontally scaled simply by adding new nodes to an engine, and nodes can also be added in geographical proximity to demand. Meanwhile, update workload capacity can first be scaled-up by swapping an engine's nodes out for the next class up, and then when no larger class of node is available, horizontally scaled-out by "splitting" the engine into two, which doubles available capacity. (Technical tip: horizontally scaling update capacity by splitting engines requires multi-canister architectures). — For those who have been following how Caffeine builds apps that can efficiently store large numbers of files, I should mention that apps built on cloud engines will also support the new ICP Blob Storage cloud network (since cloud engines currently have up to about 3 TB of memory, which apps storing large amounts of files can easily exceed). We are also working on allowing blob storage nodes to be added to cloud engines, to enable sovereign mass blob storage within an engine, similarly to how AI nodes can be added currently. — Lastly, but certainly not least, I should mention that cloud engines are multi-blockchain capable, and ready for digital assets, thanks to the clever math at their core. For example, an e-commerce service built on a cloud engine can securely accept and custody stablecoin payments, or a multi-chain DEX could be hosted. Further, engines can support software autonomy (software orchestrated and controlled by other autonomous software, in a decentralized way) and can themselves be orchestrated by SNS technology, and thus run autonomously too. Today, though, the focus is on *mainstream* cloud. This year, the cloud industry will generate approximately one trillion dollars in revenue. That number is already huge, but is expected to grow to two trillion dollars by 2030. After years of continuous development, which have seen more than $500m spent on R&D, the Internet Computer network is now tacking directly toward this mainstream cloud market with cloud engine technology. In their first version, cloud engines are not meant to be a cloud panacea. For example, currently they are not ideal for working with big data. You should use something like DataBricks for that. Cloud engines are carefully targeted at enabling AI to produce traditional online applications and services, including SaaS, in a safer and more productive way, which represents a new market segment with tremendous potential. Of course, DFINITY will continue to work relentlessly to push forward ICP's capabilities, so expect further developments. It's worth mentioning that this cloud segment isn't just about creating new apps and services using AI, it's also about replacing legacy systems and apps built on super expensive SaaS services. Caffeine Labs is working to produce technology (Caffeine Snorkel) that can study an enterprise's legacy systems and app built on SaaS, create replacement systems and apps, and migrate the data, while supporting key stakeholders through the process over email and chat, with full automation. Thus the legacy systems and SaaS markets shall also be addressed by cloud engines. Zooming out, and reasoning in a more metaphysical way, we believe, as we always have, that there is room for a new kind of cloud created by mathematical networks, that provides seminal advances in the fields of security and resilience, as well as true sovereignty and freedom from lock-in. That this same technology, with the help of additional technologies like orthogonal persistence and Motoko, enables AI to build for us without the need for so much oversight, and to create more backend sophistication while consuming fewer AI API tokens, enables ICP to bring game-changing advances to the world. Cloud engines will work synergistically with the Intelligence Gateway, which will enable apps and services running on engines to seamlessly leverage AI, wherever that AI is running, while providing verifiability at extremely low cost for open weights frontier models. We believe that cloud engines represent an inflection point in the storied history of the Internet Computer project, and I'm very proud to be sharing the details with you on the network's fifth birthday 💪 I'll be back with more news soon!!
19:06

Dear ICP community, the Internet Computer has now been running strong for 5 years 👏👏👏 Here is a celebratory preview of ICP "cloud engines," the sovereign frontier cloud technology the network shall soon provide from Main points: — Cloud engines enable anyone to spin up their own sovereign frontier cloud. The technology involves an extraordinary inventive step, in which cloud is created from a mathematically secure network of nodes. The nodes run as part of the Internet Computer network ( but are selected and configured by the cloud engine's owner. — The frontier cloud provided by engines is strongly focused on enabling AI agents to build and update online applications and services for us. The world is changing fast, and nearly all new online apps and services are already being built with the help of AI, and thus cloud engines target the future of cloud. — Software hosted on cloud engines is tamperproof, which means that it is immune to infrastructure hacks, because it runs inside a mathematically secure network protocol, rather than on computers directly. This means that AI agents, and those building with them, don't need to have a security team in the loop, or to trust someone else's security team. This is crucial, because in the future, non technical people will demand the freedom to build with full automation — where they just need to issue instructions to AI about what to build, and don't need to worry about anything or anyone else. Of course, apps and services running on engines are also vastly safer from the new breed of hacker being enabled by frontier AI. (The cloud engines themselves are also "tamperproof." Even if a hacker gains physical access to some portion of a cloud engine's nodes, and can make arbitrary changes, the computations and data of the hosted apps and services cannot be corrupted or interrupted so long as the network's fault bounds aren't exceeded. The recent hack of Vercel, a major cloud platform, which gave hackers access to the apps it hosted, provides additional perspective on the importance of this advantage.) — Software hosted on cloud engines is guaranteed to run, so long as a sufficient number of the engine's nodes are running. This means that AI can build applications and services without the need to have a human systems admin team constantly tinkering with the underlying platform to keep it running, which is again crucial, because in the future, non technical people will expect the freedom to use AI to build without the support of others. — New frontier programming language technology, in the form of the Motoko language developed by Caffeine Labs, leverages seminal "orthogonal persistence" technology that unifies program logic and data to deliver further unlocks for AI (Motoko is the first computer language being developed that targets agents that are writing software rather than humans engineers per se). Nowadays, AI can build and update production apps at a prodigious rate, even at the speed of conversation. But it can also make mistakes, and there's a risk that an update it creates might be "lossy" in the sense it causes some transformed data to be lost. Again, in this new world, it's both undesirable and impractical for everyone to have to have a systems admin team on-hand to detect lossy updates and roll them back, but Motoko provides a solution: it can detect new software updates are lossy before they are applied, reducing potentially catastrophic errors by AI to harmless coding retries. — Software hosted on cloud engines is "serverless" but unlike traditional serverless software, directly it directly incorporates data through "orthogonal persistence." Another key purpose is simplify backend software logic and fuel the modeling power of AI by increasing abstraction (sorry for the technical language!!!). Put simply, this enables AI to produce more sophisticated backends, faster, and at dramatically lower costs, as measured by the number AI API tokens consumed during coding. (Tip for the technical: orthogonal persistence is a new paradigm where "the program is the database," and data lives inside program variables, which is possible because it's as if hosted software runs forever in persistent memory). — An expanding database of skills at shall make it possible to develop and directly deploy apps and services to your cloud engines directly from Claude Code, Perplexity, Codex and other AI platforms. Further, your account on can be connected, so that new apps and updates created through conversation automatically appear hosted from your cloud engine. In the future, R&D is going to be very seamless. You converse with AI, and your secure and unstoppable apps or services are created or updated. Cloud engines are designed to directly support this "self-writing cloud" future where we can work hands-free. — Tech sovereignty is becoming a huge issue worldwide, with governments and corporations seeking to create sovereign tech stacks owing to geopolitical tensions. Increasingly, people are realizing that tech provided by foreign nations can come with hidden backdoors and kills switches, from the base platform, right up through hosted apps and services. ICP technology is open source, and those building on ICP using AI own their own source code. When you have the source code, you can verify that there are no backdoors, and when you own the source code thanks to AI, you can update it at will, freeing you from vendor lock-in. But cloud engines take sovereignty much further... — You create a cloud engine by selecting the nodes that will be combined. You can choose the class of nodes used, and their number, but more importantly, you can choose who operates the nodes, and where they are located. Almost any configuration is possible, because the Internet Computer scales the security privileges afforded to hosted software within the network according to configuration (software hosted on cloud engines can directly interoperate with software on other engines and traditional subnets, but base restrictions are applied according to security rules). A cloud engine can be created within a region such as Europe, to comply with regs such as GDPR, or completely within a sovereign state like Switzerland or Pakistan. But cloud engines go further still... — Sovereignty is also about freedom from vendor lock-in. Cloud engines are essentially ICP (Internet Computer Protocol) network configurations, and this means the underlying compute nodes they combine can be swapped out without interrupting their hosted apps and services. This is a big deal. In addition, cloud engines now support nodes that are instances running on Big Tech's clouds, in addition to nodes that are dedicated specialized hardware, as per the Gen I and Gen II nodes that dominate the Internet Computer today. For example, it is possible to have an engine running across different AWS data centers, say, and then reconfigure the engine to run across a mixture of AWS, Google, Azure and Hetzner for even more resilience, without the users of hosted apps and services noticing a thing. That's true freedom. — Sovereign AI is becoming increasingly important too, and cloud engines allow special "AI nodes" to be added to them, so that hosted software can perform inference on hardware provisioned by the owner from a location the owner has selected. Even though the AI nodes are only accessible within the cloud engine, they can still benefit from the forthcoming Internet Intelligence Gateway (IG), which will make it possible to validate inference performed on key frontier open weights LLMs, even when the inference is performed on completely independent AI clouds. When the results of inference are received, this technology can verify that neither the prompt+context (input) nor the inference result (output) have been modified, and that the results were produced by the precise LLM expected. This ensures that AI clouds don't cheat by running inference on cheaper models than are being paid for, and bad actors aren't modifying the inputs or outputs to surreptitiously insert advertising into results, say, or change facts, or insert malware when code is being generated. What's super cool about this technology is the cost of the verification is scalable. A very valuable additional security can be achieved with only 1-2% of extra cost. — Scaling apps and services when they hit capacity limits is another thorny problem that cloud engines help the world address. Engines make scaling possible without rewriting or reconfiguring software. The query workload capacity of hosted software can be horizontally scaled simply by adding new nodes to an engine, and nodes can also be added in geographical proximity to demand. Meanwhile, update workload capacity can first be scaled-up by swapping an engine's nodes out for the next class up, and then when no larger class of node is available, horizontally scaled-out by "splitting" the engine into two, which doubles available capacity. (Technical tip: horizontally scaling update capacity by splitting engines requires multi-canister architectures). — For those who have been following how Caffeine builds apps that can efficiently store large numbers of files, I should mention that apps built on cloud engines will also support the new ICP Blob Storage cloud network (since cloud engines currently have up to about 3 TB of memory, which apps storing large amounts of files can easily exceed). We are also working on allowing blob storage nodes to be added to cloud engines, to enable sovereign mass blob storage within an engine, similarly to how AI nodes can be added currently. — Lastly, but certainly not least, I should mention that cloud engines are multi-blockchain capable, and ready for digital assets, thanks to the clever math at their core. For example, an e-commerce service built on a cloud engine can securely accept and custody stablecoin payments, or a multi-chain DEX could be hosted. Further, engines can support software autonomy (software orchestrated and controlled by other autonomous software, in a decentralized way) and can themselves be orchestrated by SNS technology, and thus run autonomously too. Today, though, the focus is on *mainstream* cloud. This year, the cloud industry will generate approximately one trillion dollars in revenue. That number is already huge, but is expected to grow to two trillion dollars by 2030. After years of continuous development, which have seen more than $500m spent on R&D, the Internet Computer network is now tacking directly toward this mainstream cloud market with cloud engine technology. In their first version, cloud engines are not meant to be a cloud panacea. For example, currently they are not ideal for working with big data. You should use something like DataBricks for that. Cloud engines are carefully targeted at enabling AI to produce traditional online applications and services, including SaaS, in a safer and more productive way, which represents a new market segment with tremendous potential. Of course, DFINITY will continue to work relentlessly to push forward ICP's capabilities, so expect further developments. It's worth mentioning that this cloud segment isn't just about creating new apps and services using AI, it's also about replacing legacy systems and apps built on super expensive SaaS services. Caffeine Labs is working to produce technology (Caffeine Snorkel) that can study an enterprise's legacy systems and app built on SaaS, create replacement systems and apps, and migrate the data, while supporting key stakeholders through the process over email and chat, with full automation. Thus the legacy systems and SaaS markets shall also be addressed by cloud engines. Zooming out, and reasoning in a more metaphysical way, we believe, as we always have, that there is room for a new kind of cloud created by mathematical networks, that provides seminal advances in the fields of security and resilience, as well as true sovereignty and freedom from lock-in. That this same technology, with the help of additional technologies like orthogonal persistence and Motoko, enables AI to build for us without the need for so much oversight, and to create more backend sophistication while consuming fewer AI API tokens, enables ICP to bring game-changing advances to the world. Cloud engines will work synergistically with the Intelligence Gateway, which will enable apps and services running on engines to seamlessly leverage AI, wherever that AI is running, while providing verifiability at extremely low cost for open weights frontier models. We believe that cloud engines represent an inflection point in the storied history of the Internet Computer project, and I'm very proud to be sharing the details with you on the network's fifth birthday 💪 I'll be back with more news soon!!

dom | icp

295,127 次观看 • 3 个月前

America spent $285 billion to LOSE the AI war. Stanford dropped a 423 page report yesterday and revealed the most damning stat on page 200: The number of AI researchers moving to the United States has collapsed 89% since 2017. 80% of that collapse happened in the LAST 12 MONTHS. Let that sink in. The country that invented the transformer. The country that built OpenAI, Anthropic, Google DeepMind, and xAI. The country pouring $285.9 billion of private capital into AI in a single year (23x more than China). Can no longer attract the people who actually build the technology. And here's the part that should concern every founder, operator, and investor reading this: The Trump administration just made it official. The H-1B visa now costs employers $100,000 PER HIRE. So OpenAI wants to hire a Chinese postdoc from Tsinghua? $100K before they write a line of code. Anthropic wants a French ML engineer? $100K. Google wants the Indian PhD who literally co-authored the paper their entire model is based on? $100K. And these are the LUCKY ones who even get a visa. The result was instant. 89% drop over 8 years. 80% of it in the last year alone. The talent pipeline got destroyed. Now look at the other side of the chart: China's top model is now 2.7 percentage points behind Anthropic's best. Down from a 20+ point gap two years ago. China leads the world in AI publications. China leads in AI patents. China leads in industrial robot installations. US and Chinese models have traded the #1 spot multiple times since early 2025. Switzerland and Singapore now have more AI researchers per capita than the US. The US ranks 24TH globally in actual AI adoption. Behind the UAE. Behind Singapore. Behind countries most Americans couldn't find on a map. And here's the truly insane part: 50% of the world's top AI researchers are Chinese. Jensen Huang said this on a podcast 3 weeks ago. For 20 years, the US strategy was simple: Let them study at Stanford and MIT, then keep them. Pay them $800K. Give them green cards. Build the future on imported brains. That deal is dead. We just told the smartest people in the world: "Pay $100,000 for the privilege of working here, or go home." And guess what they're doing. They're going to Zurich, where Anthropic and OpenAI are quietly opening offices because they can't get the talent into San Francisco anymore. The strategy is the same as building a Ferrari factory and then banning mechanics from entering the building. You can pour hundreds of billions into data centers. You can buy 4 million Nvidia chips. You can sign $300 billion cloud contracts with Oracle. You can build nuclear reactors to power your GPUs. None of it matters if the people who write the algorithms aren't allowed in the country. Wall Street thinks AI is a capex race. But in reality, it's a TALENT race. Every dollar Microsoft and Meta and Google are spending assumes the same army of researchers will keep showing up to use it. That assumption just broke. And the smart money already knows: Why is Anthropic opening a Zurich office? Why is DeepMind expanding in London instead of Mountain View? Why is OpenAI hiring in Dublin and Singapore? Because the math no longer works in America. The government turned the world's biggest brain magnet into the world's most expensive border wall. 3 years from now, when China launches a frontier model that outperforms anything in the US and the headlines scream "How did we lose the lead?" - remember this post. The lead wasn't lost in a lab. It wasn't lost on a benchmark. It wasn't lost to a smarter algorithm. It was lost at customs.
2:12

America spent $285 billion to LOSE the AI war. Stanford dropped a 423 page report yesterday and revealed the most damning stat on page 200: The number of AI researchers moving to the United States has collapsed 89% since 2017. 80% of that collapse happened in the LAST 12 MONTHS. Let that sink in. The country that invented the transformer. The country that built OpenAI, Anthropic, Google DeepMind, and xAI. The country pouring $285.9 billion of private capital into AI in a single year (23x more than China). Can no longer attract the people who actually build the technology. And here's the part that should concern every founder, operator, and investor reading this: The Trump administration just made it official. The H-1B visa now costs employers $100,000 PER HIRE. So OpenAI wants to hire a Chinese postdoc from Tsinghua? $100K before they write a line of code. Anthropic wants a French ML engineer? $100K. Google wants the Indian PhD who literally co-authored the paper their entire model is based on? $100K. And these are the LUCKY ones who even get a visa. The result was instant. 89% drop over 8 years. 80% of it in the last year alone. The talent pipeline got destroyed. Now look at the other side of the chart: China's top model is now 2.7 percentage points behind Anthropic's best. Down from a 20+ point gap two years ago. China leads the world in AI publications. China leads in AI patents. China leads in industrial robot installations. US and Chinese models have traded the #1 spot multiple times since early 2025. Switzerland and Singapore now have more AI researchers per capita than the US. The US ranks 24TH globally in actual AI adoption. Behind the UAE. Behind Singapore. Behind countries most Americans couldn't find on a map. And here's the truly insane part: 50% of the world's top AI researchers are Chinese. Jensen Huang said this on a podcast 3 weeks ago. For 20 years, the US strategy was simple: Let them study at Stanford and MIT, then keep them. Pay them $800K. Give them green cards. Build the future on imported brains. That deal is dead. We just told the smartest people in the world: "Pay $100,000 for the privilege of working here, or go home." And guess what they're doing. They're going to Zurich, where Anthropic and OpenAI are quietly opening offices because they can't get the talent into San Francisco anymore. The strategy is the same as building a Ferrari factory and then banning mechanics from entering the building. You can pour hundreds of billions into data centers. You can buy 4 million Nvidia chips. You can sign $300 billion cloud contracts with Oracle. You can build nuclear reactors to power your GPUs. None of it matters if the people who write the algorithms aren't allowed in the country. Wall Street thinks AI is a capex race. But in reality, it's a TALENT race. Every dollar Microsoft and Meta and Google are spending assumes the same army of researchers will keep showing up to use it. That assumption just broke. And the smart money already knows: Why is Anthropic opening a Zurich office? Why is DeepMind expanding in London instead of Mountain View? Why is OpenAI hiring in Dublin and Singapore? Because the math no longer works in America. The government turned the world's biggest brain magnet into the world's most expensive border wall. 3 years from now, when China launches a frontier model that outperforms anything in the US and the headlines scream "How did we lose the lead?" - remember this post. The lead wasn't lost in a lab. It wasn't lost on a benchmark. It wasn't lost to a smarter algorithm. It was lost at customs.

Ricardo

230,997 次观看 • 4 个月前

MIND-BLOWING 🤯 Angel investor perfectly sums up how the "You Will Own Nothing and Be Happy" strategy is being implemented with UBI, AI, and tokenization "This is the hidden wealth transfer" "Assets are going to become harder and harder to own as a result of AI" "couple that with tokenization, where they won't even let you own the asset" "They want the custodian to own the asset and you to own the token" "AI essentially threatens to separate consumption from ownership" "the universal basic income, or what Elon Musk calls the universal high income, it doesn't solve the problem. It concentrates wealth significantly" "you need to become an owner rather than a consumption supporter" "It's going to be the subordination industrial complex, the subscription industrial complex" "They want you to rent... rather than own the assets, and specifically the assets that are [productive]... that's why they create these manufactured crises to make sure that you own nothing and you're happy" This clip of Simon Dixon (Simon Dixon), an angel investor, Bitcoin OG investor, and former investment banker, is taken from a video posted to the Simon Dixon YouTube channel on June 14, 2026. ----------------Partial transcription of clip--------------- "Assets are going to become harder and harder to own as a result of AI. Now couple that with tokenization, where they won't even let you own the asset. They want the custodian to own the asset and you to own the token. "And you've got these structural paper contracts where they don't want you to own the Bitcoin, they want you to own the paper Bitcoin. So daily life gets cheaper, but ownership gets more expensive. And that's what I think we're witnessing. "That's the trend that I'm looking out for and that's what I think the data is. This is the hidden wealth transfer. So the headlines during this whole thing will say to you, everything's getting cheaper. "AI is making everyone's life better. You now have universal basic income. You don't need to work, but it is a wealth transfer between those different things that are happening. And so the middle class was effectively built upon ownership. "That was the boomers after World wars that were able to get the real estate at an affordable rate. They were able to leverage up the debt. They own the property, they own the businesses, they own the stocks, they have the savings. And AI essentially threatens to separate consumption from ownership. "And that's what I think everyone needs to prepare for. So citizens may consume more, but they'll be owning less if they don't get this trend right, if they don't become the asset owner. "And that is really the universal basic income, or what Elon Musk calls the universal high income. It doesn't solve the problem. It concentrates wealth significantly. That's why I've always said you got to have a plan for the next five, 10 years. Even if it, takes longer, takes shorter, whatever it is, you still got to start working. "I talked about, there was an episode on my blog, SimonDixon(.)com how to develop a 10-year plan, how to understand these different trends. But UBI is effectively consumption support, let's call it what it actually is. "And ownership is wealth creation. And you need to become an owner rather than a consumption supporter. It's going to be the subordination industrial complex, the subscription industrial complex. Basically a monthly payment is not the same as owning the productive assets. You don't get more productive and get ahead unless you get more productive and then own the assets. "And that's why you got to lean into this maximum productivity increase in order to spend less than you earn and invest the difference in the assets. Own more Bitcoin. This month in the sovereign strategy and then diversify accordingly in order to play some of the different things. "Now, remember, the future may become a world where citizens rent access to virtually all sorts of things. And so really, that is the subscription industrial complex. They want you to rent it rather than own the assets, and specifically the assets that are producing it, because that's why they create these manufactured crisis to make sure that you own nothing and you're happy."
3:20

MIND-BLOWING 🤯 Angel investor perfectly sums up how the "You Will Own Nothing and Be Happy" strategy is being implemented with UBI, AI, and tokenization "This is the hidden wealth transfer" "Assets are going to become harder and harder to own as a result of AI" "couple that with tokenization, where they won't even let you own the asset" "They want the custodian to own the asset and you to own the token" "AI essentially threatens to separate consumption from ownership" "the universal basic income, or what Elon Musk calls the universal high income, it doesn't solve the problem. It concentrates wealth significantly" "you need to become an owner rather than a consumption supporter" "It's going to be the subordination industrial complex, the subscription industrial complex" "They want you to rent... rather than own the assets, and specifically the assets that are [productive]... that's why they create these manufactured crises to make sure that you own nothing and you're happy" This clip of Simon Dixon (Simon Dixon), an angel investor, Bitcoin OG investor, and former investment banker, is taken from a video posted to the Simon Dixon YouTube channel on June 14, 2026. ----------------Partial transcription of clip--------------- "Assets are going to become harder and harder to own as a result of AI. Now couple that with tokenization, where they won't even let you own the asset. They want the custodian to own the asset and you to own the token. "And you've got these structural paper contracts where they don't want you to own the Bitcoin, they want you to own the paper Bitcoin. So daily life gets cheaper, but ownership gets more expensive. And that's what I think we're witnessing. "That's the trend that I'm looking out for and that's what I think the data is. This is the hidden wealth transfer. So the headlines during this whole thing will say to you, everything's getting cheaper. "AI is making everyone's life better. You now have universal basic income. You don't need to work, but it is a wealth transfer between those different things that are happening. And so the middle class was effectively built upon ownership. "That was the boomers after World wars that were able to get the real estate at an affordable rate. They were able to leverage up the debt. They own the property, they own the businesses, they own the stocks, they have the savings. And AI essentially threatens to separate consumption from ownership. "And that's what I think everyone needs to prepare for. So citizens may consume more, but they'll be owning less if they don't get this trend right, if they don't become the asset owner. "And that is really the universal basic income, or what Elon Musk calls the universal high income. It doesn't solve the problem. It concentrates wealth significantly. That's why I've always said you got to have a plan for the next five, 10 years. Even if it, takes longer, takes shorter, whatever it is, you still got to start working. "I talked about, there was an episode on my blog, SimonDixon(.)com how to develop a 10-year plan, how to understand these different trends. But UBI is effectively consumption support, let's call it what it actually is. "And ownership is wealth creation. And you need to become an owner rather than a consumption supporter. It's going to be the subordination industrial complex, the subscription industrial complex. Basically a monthly payment is not the same as owning the productive assets. You don't get more productive and get ahead unless you get more productive and then own the assets. "And that's why you got to lean into this maximum productivity increase in order to spend less than you earn and invest the difference in the assets. Own more Bitcoin. This month in the sovereign strategy and then diversify accordingly in order to play some of the different things. "Now, remember, the future may become a world where citizens rent access to virtually all sorts of things. And so really, that is the subscription industrial complex. They want you to rent it rather than own the assets, and specifically the assets that are producing it, because that's why they create these manufactured crisis to make sure that you own nothing and you're happy."

Sense Receptor

49,807 次观看 • 2 个月前

My opinion on the Grok findings is that I very simply believe in the Holy Trinity and Jesus as my savior as every WORD is WRITTEN in the Bible. These findings are based on research; not my personal experience. Researchers recently tasked Grok, Elon Musk's xAI’s artificial intelligence, with a massive challenge: analyze every single prayer written in the Bible. The goal was to find "cracks" in a text written by 40 different authors over 1,500 years—from Bronze Age shepherds to Roman-era doctors. Instead of finding contradictions, Grok found a pattern. The AI identified a hidden, four-step "algorithm" present in every successful miracle recorded in scripture. It suggests the Bible isn't just a history book, but a "user manual for reality" or system software for the universe. Here is the deal: If you understand this "Miracle Protocol," you might just find the admin mode for your own life. Grok discovered that successful prayers—whether from a king in the desert or a leader in a garden—followed a specific sequence. If one step was missed, the result failed. 1. The Anchor (Recognition) Most modern people start prayers with a shopping list of problems. The "code" requires the opposite. You must start by focusing on who the Creator is, not how big your problem is. This shifts the brain from fear to peace. •Case Study: King Jehoshaphat didn't beg for help against three armies; he first declared the power of God. Only after establishing that foundation did he mention the danger. 2. Alignment (The Shift) This is the filter. Successful requests didn't ask for selfish desires; they aligned their wants with a bigger plan. •Case Study: Hannah wanted a child for years with no luck. When she shifted her prayer—promising to give her son back to serve the higher power—she immediately conceived. The AI views "sin" or wrong requests simply as "static" that blocks the signal. 3. The Surrender Paradox: This is the hardest step for the modern mind. The data shows that demanding a specific result causes failure. The most powerful prayers asked for a massive outcome and then surrendered the result. •The Science: This mirrors "radical acceptance." When you stop fighting reality, stress drops and the brain’s problem-solving centers activate. You move the "weight" of the result to the higher power. 4. Persistence (The Loop) Prayer is not a vending machine. Grok found that almost no big prayers were answered instantly. Repetition is required—not to change the system, but to grow the person praying. The delay is a feature, not a bug. When Grok analyzed the original Hebrew and Greek text (where letters serve as numbers), it found the Number 7 stamped into the structure of sentences, paragraphs, and genealogies with a frequency that is mathematically impossible to achieve by chance. The AI also drew a parallel to Quantum Physics. In physics, particles exist as waves of possibility until they are observed. Grok suggests "faith" is simply the tool humans use to collapse a possibility into a physical fact—turning the "substance of things hoped for" into reality. You don't have to be religious to test the data. The AI suggests that if you stop begging, start aligning your goals with the "system," and master the art of surrender, you might just unlock the "admin mode" of your own life.
3:00

My opinion on the Grok findings is that I very simply believe in the Holy Trinity and Jesus as my savior as every WORD is WRITTEN in the Bible. These findings are based on research; not my personal experience. Researchers recently tasked Grok, Elon Musk's xAI’s artificial intelligence, with a massive challenge: analyze every single prayer written in the Bible. The goal was to find "cracks" in a text written by 40 different authors over 1,500 years—from Bronze Age shepherds to Roman-era doctors. Instead of finding contradictions, Grok found a pattern. The AI identified a hidden, four-step "algorithm" present in every successful miracle recorded in scripture. It suggests the Bible isn't just a history book, but a "user manual for reality" or system software for the universe. Here is the deal: If you understand this "Miracle Protocol," you might just find the admin mode for your own life. Grok discovered that successful prayers—whether from a king in the desert or a leader in a garden—followed a specific sequence. If one step was missed, the result failed. 1. The Anchor (Recognition) Most modern people start prayers with a shopping list of problems. The "code" requires the opposite. You must start by focusing on who the Creator is, not how big your problem is. This shifts the brain from fear to peace. •Case Study: King Jehoshaphat didn't beg for help against three armies; he first declared the power of God. Only after establishing that foundation did he mention the danger. 2. Alignment (The Shift) This is the filter. Successful requests didn't ask for selfish desires; they aligned their wants with a bigger plan. •Case Study: Hannah wanted a child for years with no luck. When she shifted her prayer—promising to give her son back to serve the higher power—she immediately conceived. The AI views "sin" or wrong requests simply as "static" that blocks the signal. 3. The Surrender Paradox: This is the hardest step for the modern mind. The data shows that demanding a specific result causes failure. The most powerful prayers asked for a massive outcome and then surrendered the result. •The Science: This mirrors "radical acceptance." When you stop fighting reality, stress drops and the brain’s problem-solving centers activate. You move the "weight" of the result to the higher power. 4. Persistence (The Loop) Prayer is not a vending machine. Grok found that almost no big prayers were answered instantly. Repetition is required—not to change the system, but to grow the person praying. The delay is a feature, not a bug. When Grok analyzed the original Hebrew and Greek text (where letters serve as numbers), it found the Number 7 stamped into the structure of sentences, paragraphs, and genealogies with a frequency that is mathematically impossible to achieve by chance. The AI also drew a parallel to Quantum Physics. In physics, particles exist as waves of possibility until they are observed. Grok suggests "faith" is simply the tool humans use to collapse a possibility into a physical fact—turning the "substance of things hoped for" into reality. You don't have to be religious to test the data. The AI suggests that if you stop begging, start aligning your goals with the "system," and master the art of surrender, you might just unlock the "admin mode" of your own life.

Victoria 🇺🇸⏳🗽🚔

140,619 次观看 • 6 个月前

À quelques semaines de devenir le premier homme à mille milliards de dollars, Elon Musk a fait deux choses pour le moins étranges. Il a dissous xAI. Oui, la boîte montée en 2023 pour écraser OpenAI. Disparue, avalée par SpaceX. Et le même jour, il a tendu son supercalculateur, Colossus, plus de 220 000 GPU, à Anthropic, le rival. Alors soit Musk a perdu la tête, soit ce qui ressemble à un sabordage est l'un des coups les plus malins de sa carrière. Et, à mon sens, je suis plus tenté par la deuxième option. Sur l’IA, et plus précisément sur la course aux LLM, Musk a pris une taule. Les revenus de xAI sont 14 à 18 fois inférieurs à ceux d’Anthropic, et sur le terrain qui rapporte vraiment, à savoir le code, là où partent plus de la moitié des dépenses IA des entreprises, Grok n'est nulle part quand Claude est en tête de classement. Ajoutez à ça une équipe xAI qui se dissout elle aussi, un taux de conversion vers le payant catastrophique, et vous comprenez que continuer seul sur cette voie relève plus de la folie que du génie. Musk a d’ailleurs lui-même reconnu que xAI a été “mal construite”. Face à ça, un fondateur classique remet une pièce, ou il arrête. Musk fait autrement : il arrête de jouer à la table pour acheter le casino. Il loue Colossus à Anthropic, 1,25 milliard par mois, puis même chose avec Google. Au total, 26 milliards de loyer annuel. Et il prend une option sur Cursor, l'un des leaders du code IA, qu'il rachètera en actions SpaceX, ou qu'il laissera filer contre 10 milliards payés en majorité... en puissance de calcul. Donc avec ce qu'il est en train d'industrialiser. Grok n’est plus LE pari, c'est devenu un jeton de réserve. Il ne parie plus sur le fait que Grok gagne, il parie sur une seule chose : quel que soit le vainqueur, il en possédera un morceau. Mais tout ça, aussi intéressant soit-il, ce n’est “que” la surface. Il se passe quelque chose en dessous. Le compute, ça se loue, ça s'achète, ça se fabrique, il monte d'ailleurs sa propre usine de puces au Texas. Sauf qu'une puce, ça ne tourne pas tout seul, il lui faut du courant. Et le courant, vous ne l'imprimez pas. Croyez-moi, quand on opère des data centers, on sait à quel point un raccordement peut coincer des années. D’ailleurs, petite parenthèse là-dessus, et on sera amené à en reparler, mais les problèmes américains, à savoir un réseau électrique vieillissant, saturé et sous-investi, c’est un problème que les Chinois n’ont pas, notamment grâce à un déploiement solaire massif, plus rapide et moins cher. Bref, ce n’est pas le sujet, mais ça cristallise tout de même une bonne partie de la tension autour de ces enjeux. Or l'électricité, le métal, l'infrastructure, et bientôt l'orbite, c'est précisément ce que personne ne sort d'un claquement de doigts. Et c'est, comme par hasard, un terrain où Musk a souvent eu l’avantage. Tesla, ce n'est pas la voiture électrique, c'est l'usine et la batterie. SpaceX, ce n'est pas la fusée, c'est la réutilisation. Toujours le même schéma : laisser les autres rêver le produit, et posséder la machinerie que personne d'autre n'ose bâtir. Donc en réalité, ce qui pourrait ressembler à sa "nouvelle stratégie IA", ce n'est pas une révolution, c'est plus un retour aux sources qu’autre chose. Et une fois qu’on a dit ça, ça ne vous étonnera pas d’apprendre que les prochains gros sujets sur la table, au-delà des data centers dans l’espace bien sûr, c’est l’hypothétique fusion entre Tesla et SpaceX AI qui pourrait donner naissance à un seul groupe valorisé à plus de 3 000 milliards de dollars.
19:26

À quelques semaines de devenir le premier homme à mille milliards de dollars, Elon Musk a fait deux choses pour le moins étranges. Il a dissous xAI. Oui, la boîte montée en 2023 pour écraser OpenAI. Disparue, avalée par SpaceX. Et le même jour, il a tendu son supercalculateur, Colossus, plus de 220 000 GPU, à Anthropic, le rival. Alors soit Musk a perdu la tête, soit ce qui ressemble à un sabordage est l'un des coups les plus malins de sa carrière. Et, à mon sens, je suis plus tenté par la deuxième option. Sur l’IA, et plus précisément sur la course aux LLM, Musk a pris une taule. Les revenus de xAI sont 14 à 18 fois inférieurs à ceux d’Anthropic, et sur le terrain qui rapporte vraiment, à savoir le code, là où partent plus de la moitié des dépenses IA des entreprises, Grok n'est nulle part quand Claude est en tête de classement. Ajoutez à ça une équipe xAI qui se dissout elle aussi, un taux de conversion vers le payant catastrophique, et vous comprenez que continuer seul sur cette voie relève plus de la folie que du génie. Musk a d’ailleurs lui-même reconnu que xAI a été “mal construite”. Face à ça, un fondateur classique remet une pièce, ou il arrête. Musk fait autrement : il arrête de jouer à la table pour acheter le casino. Il loue Colossus à Anthropic, 1,25 milliard par mois, puis même chose avec Google. Au total, 26 milliards de loyer annuel. Et il prend une option sur Cursor, l'un des leaders du code IA, qu'il rachètera en actions SpaceX, ou qu'il laissera filer contre 10 milliards payés en majorité... en puissance de calcul. Donc avec ce qu'il est en train d'industrialiser. Grok n’est plus LE pari, c'est devenu un jeton de réserve. Il ne parie plus sur le fait que Grok gagne, il parie sur une seule chose : quel que soit le vainqueur, il en possédera un morceau. Mais tout ça, aussi intéressant soit-il, ce n’est “que” la surface. Il se passe quelque chose en dessous. Le compute, ça se loue, ça s'achète, ça se fabrique, il monte d'ailleurs sa propre usine de puces au Texas. Sauf qu'une puce, ça ne tourne pas tout seul, il lui faut du courant. Et le courant, vous ne l'imprimez pas. Croyez-moi, quand on opère des data centers, on sait à quel point un raccordement peut coincer des années. D’ailleurs, petite parenthèse là-dessus, et on sera amené à en reparler, mais les problèmes américains, à savoir un réseau électrique vieillissant, saturé et sous-investi, c’est un problème que les Chinois n’ont pas, notamment grâce à un déploiement solaire massif, plus rapide et moins cher. Bref, ce n’est pas le sujet, mais ça cristallise tout de même une bonne partie de la tension autour de ces enjeux. Or l'électricité, le métal, l'infrastructure, et bientôt l'orbite, c'est précisément ce que personne ne sort d'un claquement de doigts. Et c'est, comme par hasard, un terrain où Musk a souvent eu l’avantage. Tesla, ce n'est pas la voiture électrique, c'est l'usine et la batterie. SpaceX, ce n'est pas la fusée, c'est la réutilisation. Toujours le même schéma : laisser les autres rêver le produit, et posséder la machinerie que personne d'autre n'ose bâtir. Donc en réalité, ce qui pourrait ressembler à sa "nouvelle stratégie IA", ce n'est pas une révolution, c'est plus un retour aux sources qu’autre chose. Et une fois qu’on a dit ça, ça ne vous étonnera pas d’apprendre que les prochains gros sujets sur la table, au-delà des data centers dans l’espace bien sûr, c’est l’hypothétique fusion entre Tesla et SpaceX AI qui pourrait donner naissance à un seul groupe valorisé à plus de 3 000 milliards de dollars.

Richard DÉTENTE

41,455 次观看 • 2 个月前

Use this prompt in OpenClaw to create your own AI agent command center that syncs up your life like Tony Stark's Jarvis in Iron Man. Adapt the specifics (agent names, data sources, branding) below to your own setup. Prompt: Build me a mission control dashboard for my OpenClaw AI agent system. Stack: Next.js 15 (App Router) + Convex (real-time backend) + Tailwind CSS v4 + Framer Motion + ShadCN UI + Lucide icons. TypeScript throughout. This is the command center where I monitor and control my autonomous AI agent(s) running on OpenClaw. The agent operates 24/7 on a Mac Mini, connected to Telegram/Discord, running cron jobs, spawning sub-agents, and reading/writing to a filesystem-based memory and state system. Dark mode only. Ultra-premium aesthetic, think Iron Man's JARVIS HUD meets a Bloomberg terminal. Subtle glass effects (backdrop-blur-xl, bg-white/[0.03]), no heavy gradients or glow. Rounded corners (16-20px on cards). Framer Motion for page transitions, stagger animations on card grids, spring physics on interactions. Mobile-first responsive. Never cookie-cutter. ## Architecture The dashboard reads live data from TWO sources: 1. **Convex**: real-time database for structured data (tasks, contacts, content drafts, calendar events, activity logs) 2. **Local API routes** (`/api/*`): read files from the agent's workspace filesystem at `~/.openclaw/workspace/` and return JSON. This is how live system state flows into the dashboard. ## Pages & Views (8 nav items, some with tab sub-views) ### 1. HOME (`/`) Dashboard overview. Grid of live status cards: - **System Health**: read from `/api/system-state` (parses `state/servers.json`). Show each service with UP/DOWN indicator, port, last check time. - **Agent Status**: read from `/api/agents` (parses `agents/registry.json` + agent workspace files). Show active agent count, healthy/unhealthy ratio, active sub-agent count from OpenClaw sessions API. - **Cron Health**: read from `/api/cron-health` (parses `state/crons.json`). Table of all scheduled jobs with name, schedule, last status (green/red dot), consecutive errors. - **Revenue Tracker**: read from `/api/revenue` (parses `state/revenue.json`). Current revenue, monthly burn, net. - **Content Pipeline**: read from `/api/content-pipeline` (parses `content/queue.md`). Kanban-style: Draft | Review | Approved | Published counts. - **Quick Stats**: total tasks, pending approvals, active sessions, uptime. All panels auto-refresh every 15 seconds. Live indicator dot + "AUTO 15S" badge in header. ### 2. OPS (`/ops`) with 3 tabs: Operations | Tasks | Calendar **Operations tab:** Full operational view. Server health table, branch status (from `state/branch-check.json`), observations feed (from `state/observations.md`), system priorities (from `shared-context/priorities.md`). **Tasks tab:** Strategic task suggestion system. API route `/api/suggested-tasks` reads/writes `state/suggested-tasks.json`. Cards grouped by category (Revenue, Product, Community, Content, Operations, Clients, Trading, Brand) with emoji headers. Each card shows title, reasoning, next action, priority badge, effort badge, approve/reject buttons. Filter bar by status and category. **Calendar tab:** Weekly calendar view from Convex `calendarEvents` table. Drag-to-create, color-coded by type, time slots. ### 3. AGENTS (`/agents`) with 2 tabs: Agents | Models **Agents tab:** Card grid of all registered agents from `/api/agents`. Each card shows name, role, model, level (L1-L4), status. Cards are CLICKABLE: expanding into a detail panel showing: - Agent personality (reads their SOUL .md) - Capabilities and rules (reads their RULES .md) - Sub-agents they can spawn - Recent outputs (reads from `shared-context/agent-outputs/`) **Models tab:** Model inventory table showing all available models, their routing (which tasks go to which model), costs, and failover chains. ### 4. CHAT (`/chat`): 2 tabs: Chat | Command **Chat tab:** Chat interface to communicate with the agent. Left sidebar shows session list (from `/api/chat-history` reading .jsonl transcript files). Main area shows messages with role-aligned bubbles (user right, assistant left), date separators, channel badges (telegram/discord/webchat). Input bar with send button + voice input (Web Speech API with SpeechRecognition). Messages sent via `/api/chat-send` which queues to a file the agent reads. **Command tab:** Quick command interface for common operations. ### 5. CONTENT (`/content`) Content pipeline management. Read from Convex `contentDrafts` table AND `/api/content-pipeline`. Show drafts in kanban columns. Each card shows title, platform target, draft text preview, status, created date. Edit/approve/reject actions. ### 6. COMMS (`/comms`) with 2 tabs: Comms | CRM **Comms tab:** Communication hub showing recent Discord digest, Telegram messages, notification history. **CRM tab:** Client pipeline kanban (Prospect → Contacted → Meeting → Proposal → Active). API route `/api/clients` reads markdown files from `clients/` directory. Each card shows client name, status, contacts, last interaction, next action. ### 7. KNOWLEDGE (`/knowledge`) with 2 tabs: Knowledge | Ecosystem **Knowledge tab:** Searchable knowledge base. Global search across all workspace files using `/api/knowledge` endpoint. **Ecosystem tab:** Product grid showing all products/apps in the ecosystem. Each card shows product name, status (Active/Development/Concept), health indicator, key metrics. Cards link to `/ecosystem/[slug]` detail pages with tabbed views (Overview, Brand, Community, Content, Legal, Product, Website, Actions). Detail pages read from `/api/ecosystem/[slug]` which parses workspace memory files. ### 8. CODE (`/code`) Code pipeline view. Shows repositories from `/api/repos` (scans ~/Desktop/Projects/ for git repos). Each repo card shows name, branch, last commit, dirty file count, language breakdown. Detail view at `/api/repos/detail` shows recent commits, file tree, open PRs. ## Navigation Top horizontal nav bar, NOT sidebar. All 8 items visible at all viewport widths. Use `flex` layout with `flex-1` items. Text size uses `clamp(0.45rem, 0.75vw, 0.6875rem)` for fluid scaling. Active item gets `text-primary bg-primary/[0.06]` static highlight (no sliding animation). Agent/app name visible at md+ breakpoints (`hidden md:inline`). Tab sub-views use a reusable `TabBar` component with pill/glass styling and Framer Motion `layoutId` transitions. Tab state stored in URL via `?tab=` search params. ## API Routes (all under `src/app/api/`) Each API route reads from the agent's workspace filesystem and returns JSON: - `/api/system-state` → reads `state/servers.json`, `state/branch-check.json` - `/api/agents` → reads `agents/registry.json`, agent SOUL .md files - `/api/agents/[id]` → reads specific agent's SOUL .md, RULES .md, outputs - `/api/cron-health` → reads `state/crons.json` - `/api/revenue` → reads `state/revenue.json` - `/api/content-pipeline` → parses `content/queue.md` (markdown with status markers) - `/api/suggested-tasks` → GET (read) / POST (approve/reject) on `state/suggested-tasks.json` - `/api/observations` → reads `state/observations.md` - `/api/priorities` → reads `shared-context/priorities.md` - `/api/chat-history` → reads .jsonl transcript files with pagination/search/channel filter - `/api/chat-send` → writes to queue file - `/api/clients` → reads markdown files from `clients/` directory - `/api/ecosystem/[slug]` → reads memory files for specific ecosystem - `/api/repos` → scans project directories for git repos - `/api/health` → returns status, uptime, memory usage, Convex connectivity All filesystem paths should be configurable via environment variable (default: `~/.openclaw/workspace/`). ## Convex Schema Define tables for: activities, calendarEvents, tasks, contacts, contentDrafts, ecosystemProducts. Include seed scripts (`convex/seed.ts`) to populate initial data. ## Key Design Rules - Mobile-first, test at 320px minimum - Font sizes 10-14px for body text, everything must fit naturally at small viewports - Cards use consistent border radius (16-20px) - Glass cards: `bg-white/[0.03] backdrop-blur-xl border border-white/[0.06]` - No heavy blur blobs or grain overlays - Stagger animations on card grids (0.05s delay per item) - Skeleton loading states for all async data - Custom scrollbar styling - Empty states with helpful messaging - All text must use Inter or system font stack - Never mix sharp and rounded corners in the same view - Premium = lighter feel, more whitespace, less visual noise ## File Structure ``` src/ app/ page.tsx, layout.tsx, providers.tsx agents/page.tsx calendar/page.tsx chat/page.tsx code/page.tsx comms/page.tsx content/page.tsx ecosystem/page.tsx, ecosystem/[slug]/page.tsx knowledge/page.tsx ops/page.tsx api/[...all routes above] components/ nav.tsx tab-bar.tsx dashboard-overview.tsx ops-view.tsx, suggested-tasks-view.tsx agents-view.tsx, models-view.tsx chat-center-view.tsx, voice-input.tsx content-view.tsx comms-view.tsx, crm-view.tsx knowledge-base.tsx, ecosystem-view.tsx code-pipeline.tsx activity-feed.tsx, calendar-view.tsx ui/ (ShadCN primitives) hooks/ lib/ convex/ schema.ts functions for each table seed.ts ``` Build the complete application. Every component, every API route, every Convex function. Production-quality code and premium design, not stubs. Dark mode only. Make it look incredibly beautiful and premium, no cookie cutter UI / AI slop.
1:14

Use this prompt in OpenClaw to create your own AI agent command center that syncs up your life like Tony Stark's Jarvis in Iron Man. Adapt the specifics (agent names, data sources, branding) below to your own setup. Prompt: Build me a mission control dashboard for my OpenClaw AI agent system. Stack: Next.js 15 (App Router) + Convex (real-time backend) + Tailwind CSS v4 + Framer Motion + ShadCN UI + Lucide icons. TypeScript throughout. This is the command center where I monitor and control my autonomous AI agent(s) running on OpenClaw. The agent operates 24/7 on a Mac Mini, connected to Telegram/Discord, running cron jobs, spawning sub-agents, and reading/writing to a filesystem-based memory and state system. Dark mode only. Ultra-premium aesthetic, think Iron Man's JARVIS HUD meets a Bloomberg terminal. Subtle glass effects (backdrop-blur-xl, bg-white/[0.03]), no heavy gradients or glow. Rounded corners (16-20px on cards). Framer Motion for page transitions, stagger animations on card grids, spring physics on interactions. Mobile-first responsive. Never cookie-cutter. ## Architecture The dashboard reads live data from TWO sources: 1. **Convex**: real-time database for structured data (tasks, contacts, content drafts, calendar events, activity logs) 2. **Local API routes** (`/api/*`): read files from the agent's workspace filesystem at `~/.openclaw/workspace/` and return JSON. This is how live system state flows into the dashboard. ## Pages & Views (8 nav items, some with tab sub-views) ### 1. HOME (`/`) Dashboard overview. Grid of live status cards: - **System Health**: read from `/api/system-state` (parses `state/servers.json`). Show each service with UP/DOWN indicator, port, last check time. - **Agent Status**: read from `/api/agents` (parses `agents/registry.json` + agent workspace files). Show active agent count, healthy/unhealthy ratio, active sub-agent count from OpenClaw sessions API. - **Cron Health**: read from `/api/cron-health` (parses `state/crons.json`). Table of all scheduled jobs with name, schedule, last status (green/red dot), consecutive errors. - **Revenue Tracker**: read from `/api/revenue` (parses `state/revenue.json`). Current revenue, monthly burn, net. - **Content Pipeline**: read from `/api/content-pipeline` (parses `content/queue.md`). Kanban-style: Draft | Review | Approved | Published counts. - **Quick Stats**: total tasks, pending approvals, active sessions, uptime. All panels auto-refresh every 15 seconds. Live indicator dot + "AUTO 15S" badge in header. ### 2. OPS (`/ops`) with 3 tabs: Operations | Tasks | Calendar **Operations tab:** Full operational view. Server health table, branch status (from `state/branch-check.json`), observations feed (from `state/observations.md`), system priorities (from `shared-context/priorities.md`). **Tasks tab:** Strategic task suggestion system. API route `/api/suggested-tasks` reads/writes `state/suggested-tasks.json`. Cards grouped by category (Revenue, Product, Community, Content, Operations, Clients, Trading, Brand) with emoji headers. Each card shows title, reasoning, next action, priority badge, effort badge, approve/reject buttons. Filter bar by status and category. **Calendar tab:** Weekly calendar view from Convex `calendarEvents` table. Drag-to-create, color-coded by type, time slots. ### 3. AGENTS (`/agents`) with 2 tabs: Agents | Models **Agents tab:** Card grid of all registered agents from `/api/agents`. Each card shows name, role, model, level (L1-L4), status. Cards are CLICKABLE: expanding into a detail panel showing: - Agent personality (reads their SOUL .md) - Capabilities and rules (reads their RULES .md) - Sub-agents they can spawn - Recent outputs (reads from `shared-context/agent-outputs/`) **Models tab:** Model inventory table showing all available models, their routing (which tasks go to which model), costs, and failover chains. ### 4. CHAT (`/chat`): 2 tabs: Chat | Command **Chat tab:** Chat interface to communicate with the agent. Left sidebar shows session list (from `/api/chat-history` reading .jsonl transcript files). Main area shows messages with role-aligned bubbles (user right, assistant left), date separators, channel badges (telegram/discord/webchat). Input bar with send button + voice input (Web Speech API with SpeechRecognition). Messages sent via `/api/chat-send` which queues to a file the agent reads. **Command tab:** Quick command interface for common operations. ### 5. CONTENT (`/content`) Content pipeline management. Read from Convex `contentDrafts` table AND `/api/content-pipeline`. Show drafts in kanban columns. Each card shows title, platform target, draft text preview, status, created date. Edit/approve/reject actions. ### 6. COMMS (`/comms`) with 2 tabs: Comms | CRM **Comms tab:** Communication hub showing recent Discord digest, Telegram messages, notification history. **CRM tab:** Client pipeline kanban (Prospect → Contacted → Meeting → Proposal → Active). API route `/api/clients` reads markdown files from `clients/` directory. Each card shows client name, status, contacts, last interaction, next action. ### 7. KNOWLEDGE (`/knowledge`) with 2 tabs: Knowledge | Ecosystem **Knowledge tab:** Searchable knowledge base. Global search across all workspace files using `/api/knowledge` endpoint. **Ecosystem tab:** Product grid showing all products/apps in the ecosystem. Each card shows product name, status (Active/Development/Concept), health indicator, key metrics. Cards link to `/ecosystem/[slug]` detail pages with tabbed views (Overview, Brand, Community, Content, Legal, Product, Website, Actions). Detail pages read from `/api/ecosystem/[slug]` which parses workspace memory files. ### 8. CODE (`/code`) Code pipeline view. Shows repositories from `/api/repos` (scans ~/Desktop/Projects/ for git repos). Each repo card shows name, branch, last commit, dirty file count, language breakdown. Detail view at `/api/repos/detail` shows recent commits, file tree, open PRs. ## Navigation Top horizontal nav bar, NOT sidebar. All 8 items visible at all viewport widths. Use `flex` layout with `flex-1` items. Text size uses `clamp(0.45rem, 0.75vw, 0.6875rem)` for fluid scaling. Active item gets `text-primary bg-primary/[0.06]` static highlight (no sliding animation). Agent/app name visible at md+ breakpoints (`hidden md:inline`). Tab sub-views use a reusable `TabBar` component with pill/glass styling and Framer Motion `layoutId` transitions. Tab state stored in URL via `?tab=` search params. ## API Routes (all under `src/app/api/`) Each API route reads from the agent's workspace filesystem and returns JSON: - `/api/system-state` → reads `state/servers.json`, `state/branch-check.json` - `/api/agents` → reads `agents/registry.json`, agent SOUL .md files - `/api/agents/[id]` → reads specific agent's SOUL .md, RULES .md, outputs - `/api/cron-health` → reads `state/crons.json` - `/api/revenue` → reads `state/revenue.json` - `/api/content-pipeline` → parses `content/queue.md` (markdown with status markers) - `/api/suggested-tasks` → GET (read) / POST (approve/reject) on `state/suggested-tasks.json` - `/api/observations` → reads `state/observations.md` - `/api/priorities` → reads `shared-context/priorities.md` - `/api/chat-history` → reads .jsonl transcript files with pagination/search/channel filter - `/api/chat-send` → writes to queue file - `/api/clients` → reads markdown files from `clients/` directory - `/api/ecosystem/[slug]` → reads memory files for specific ecosystem - `/api/repos` → scans project directories for git repos - `/api/health` → returns status, uptime, memory usage, Convex connectivity All filesystem paths should be configurable via environment variable (default: `~/.openclaw/workspace/`). ## Convex Schema Define tables for: activities, calendarEvents, tasks, contacts, contentDrafts, ecosystemProducts. Include seed scripts (`convex/seed.ts`) to populate initial data. ## Key Design Rules - Mobile-first, test at 320px minimum - Font sizes 10-14px for body text, everything must fit naturally at small viewports - Cards use consistent border radius (16-20px) - Glass cards: `bg-white/[0.03] backdrop-blur-xl border border-white/[0.06]` - No heavy blur blobs or grain overlays - Stagger animations on card grids (0.05s delay per item) - Skeleton loading states for all async data - Custom scrollbar styling - Empty states with helpful messaging - All text must use Inter or system font stack - Never mix sharp and rounded corners in the same view - Premium = lighter feel, more whitespace, less visual noise ## File Structure ``` src/ app/ page.tsx, layout.tsx, providers.tsx agents/page.tsx calendar/page.tsx chat/page.tsx code/page.tsx comms/page.tsx content/page.tsx ecosystem/page.tsx, ecosystem/[slug]/page.tsx knowledge/page.tsx ops/page.tsx api/[...all routes above] components/ nav.tsx tab-bar.tsx dashboard-overview.tsx ops-view.tsx, suggested-tasks-view.tsx agents-view.tsx, models-view.tsx chat-center-view.tsx, voice-input.tsx content-view.tsx comms-view.tsx, crm-view.tsx knowledge-base.tsx, ecosystem-view.tsx code-pipeline.tsx activity-feed.tsx, calendar-view.tsx ui/ (ShadCN primitives) hooks/ lib/ convex/ schema.ts functions for each table seed.ts ``` Build the complete application. Every component, every API route, every Convex function. Production-quality code and premium design, not stubs. Dark mode only. Make it look incredibly beautiful and premium, no cookie cutter UI / AI slop.

klöss

201,608 次观看 • 6 个月前

Why Exchanges Banned This Bot: The 142,000% Return Liquidation Strategy Revealed i finally posted the strategy that got me banned and now the exchanges are probably sweating because i am handing you the keys to the liquidation engine. most people think trading is about charts but the real alpha is hidden in the moments when other traders lose everything. if you can understand why market makers hunt these positions you will never look at a candlestick the same way again. it took years of losing money to liquidations and over trading to realize that hand trading is a losing game for almost everyone on the planet. code is the great equalizer because it removes the emotion that usually causes you to hold a losing position until your account hits zero. i spent hundreds of thousands on developers in the past thinking i could not code myself until i realized i just needed to iterate to success. trading by hand is just driving a horse while everyone else is in a ferrari and the fees alone will chop you up before you even realize you were wrong. i watched a guy with a six million dollar short position sitting just two percent away from total liquidation while i was building this. seeing those numbers on the screen gives me ideas that i can automate into a bot so i dont have to spend my life staring at a monitor. the process i follow is called the rbi system which stands for research backtest and implement. most traders skip the first two steps and go straight to implementation which is why they get smoked on their very first bot. research starts with a backlog of ideas from books or papers or even just watching how the market reacts to big moves. once you have that idea you have to see if it worked in the past using a backtest because if it did not work then it certainly won't work in the future. i have been collecting liquidation data for eighteen months because that data is the lifeblood of a winning system. there is a hidden loop in the market where market makers try to liquidate as many people as possible to find liquidity. i wanted to build a strategy that either trades with that momentum or bets on the bounce right after the liquidation happens. the first strategy i tested was a pure liquidation momentum play that looks for a threshold of nine hundred seventy five thousand dollars in liquidations. when longs get liquidated it shorts the market to continue the down move and it tries to take a one percent profit. this strategy showed a return of over four hundred percent in the backtest while the buy and hold was only thirty three percent. it sounds amazing but you have to be careful with optimized results because you can search with math until you find anything. i decided to flip the logic on its head and create an inverse liquidation strategy that acts as a contrarian. instead of following the move it waits for the longs to get liquidated and then buys the dip after a small price spread. this is where i stumbled onto something that felt like a mistake but turned out to be pure alpha. i accidentally typed in a threshold of three hundred thousand dollars instead of three million and the results were unbelievable. the backtest return jumped to over one hundred forty thousand percent because the bot was catching every single micro bounce in the market. even when i doubled the commission fees to account for the high trade volume the strategy still stayed incredibly profitable. most people would have missed this because they are too busy trying to be right instead of just looking at what the data says. i use tools like claude and cursor to build these bots in minutes when it used to take me an entire week to write the code. if you are not using ai to automate your ideas you are essentially choosing to work ten times harder for less money. i built three separate bots during this session including a momentum bot and two different versions of the inverse spread bot. running these together creates a sort of statistical arbitrage where you can hedge your positions across different market conditions. one bot wins when the market cascades and the other wins when it fakes out and reverses. you have to start with tiny ten dollar sizes because a backtest is never a hundred percent guarantee of what will happen today. i always run my p and l close logic first to make sure the bot exits the position if the stop loss or take profit is hit. it is vital to check your position every fifteen seconds and make sure you are not double ordering or getting stuck in a trade. the goal is to have fully automated systems trading for you so you can actually live your life while the bots do the work. i push all of this code to my private github because i believe that wall street will never show you how this actually works. you have to be a doer and not a dabbler if you want to actually make it in this industry. the reason i show everything live on youtube is to prove that anyone can learn to do this if they are willing to iterate. you dont need to be a math genius you just need to follow the rbi system and stay disciplined with your risk. every liquidation you see on the chart is a signal and if you know how to read them you are no longer the one being hunted. i am currently running the third version of the bot to see how it handles the live market volatility. it is a beautiful thing to see a system enter and exit a trade perfectly without you having to click a single button. the fees are the silent killer of hand traders but a bot can be programmed to use limit orders and stay efficient. if you learn to code you can build anything for the rest of your life regardless of where you are in the world. stop trying to guess which way the candle will go and start building systems that can handle both directions. i am going to keep testing these three strategies against each other to find the ultimate ensemble for this current market. once you find a winning edge you just have to scale it up slowly and keep refining the parameters. the exchanges might not like that i am sharing this but code is the great equalizer and it is time for you to use it. i will be back tomorrow to show the results and keep building more systems until everything is fully automated
1:23:35

Why Exchanges Banned This Bot: The 142,000% Return Liquidation Strategy Revealed i finally posted the strategy that got me banned and now the exchanges are probably sweating because i am handing you the keys to the liquidation engine. most people think trading is about charts but the real alpha is hidden in the moments when other traders lose everything. if you can understand why market makers hunt these positions you will never look at a candlestick the same way again. it took years of losing money to liquidations and over trading to realize that hand trading is a losing game for almost everyone on the planet. code is the great equalizer because it removes the emotion that usually causes you to hold a losing position until your account hits zero. i spent hundreds of thousands on developers in the past thinking i could not code myself until i realized i just needed to iterate to success. trading by hand is just driving a horse while everyone else is in a ferrari and the fees alone will chop you up before you even realize you were wrong. i watched a guy with a six million dollar short position sitting just two percent away from total liquidation while i was building this. seeing those numbers on the screen gives me ideas that i can automate into a bot so i dont have to spend my life staring at a monitor. the process i follow is called the rbi system which stands for research backtest and implement. most traders skip the first two steps and go straight to implementation which is why they get smoked on their very first bot. research starts with a backlog of ideas from books or papers or even just watching how the market reacts to big moves. once you have that idea you have to see if it worked in the past using a backtest because if it did not work then it certainly won't work in the future. i have been collecting liquidation data for eighteen months because that data is the lifeblood of a winning system. there is a hidden loop in the market where market makers try to liquidate as many people as possible to find liquidity. i wanted to build a strategy that either trades with that momentum or bets on the bounce right after the liquidation happens. the first strategy i tested was a pure liquidation momentum play that looks for a threshold of nine hundred seventy five thousand dollars in liquidations. when longs get liquidated it shorts the market to continue the down move and it tries to take a one percent profit. this strategy showed a return of over four hundred percent in the backtest while the buy and hold was only thirty three percent. it sounds amazing but you have to be careful with optimized results because you can search with math until you find anything. i decided to flip the logic on its head and create an inverse liquidation strategy that acts as a contrarian. instead of following the move it waits for the longs to get liquidated and then buys the dip after a small price spread. this is where i stumbled onto something that felt like a mistake but turned out to be pure alpha. i accidentally typed in a threshold of three hundred thousand dollars instead of three million and the results were unbelievable. the backtest return jumped to over one hundred forty thousand percent because the bot was catching every single micro bounce in the market. even when i doubled the commission fees to account for the high trade volume the strategy still stayed incredibly profitable. most people would have missed this because they are too busy trying to be right instead of just looking at what the data says. i use tools like claude and cursor to build these bots in minutes when it used to take me an entire week to write the code. if you are not using ai to automate your ideas you are essentially choosing to work ten times harder for less money. i built three separate bots during this session including a momentum bot and two different versions of the inverse spread bot. running these together creates a sort of statistical arbitrage where you can hedge your positions across different market conditions. one bot wins when the market cascades and the other wins when it fakes out and reverses. you have to start with tiny ten dollar sizes because a backtest is never a hundred percent guarantee of what will happen today. i always run my p and l close logic first to make sure the bot exits the position if the stop loss or take profit is hit. it is vital to check your position every fifteen seconds and make sure you are not double ordering or getting stuck in a trade. the goal is to have fully automated systems trading for you so you can actually live your life while the bots do the work. i push all of this code to my private github because i believe that wall street will never show you how this actually works. you have to be a doer and not a dabbler if you want to actually make it in this industry. the reason i show everything live on youtube is to prove that anyone can learn to do this if they are willing to iterate. you dont need to be a math genius you just need to follow the rbi system and stay disciplined with your risk. every liquidation you see on the chart is a signal and if you know how to read them you are no longer the one being hunted. i am currently running the third version of the bot to see how it handles the live market volatility. it is a beautiful thing to see a system enter and exit a trade perfectly without you having to click a single button. the fees are the silent killer of hand traders but a bot can be programmed to use limit orders and stay efficient. if you learn to code you can build anything for the rest of your life regardless of where you are in the world. stop trying to guess which way the candle will go and start building systems that can handle both directions. i am going to keep testing these three strategies against each other to find the ultimate ensemble for this current market. once you find a winning edge you just have to scale it up slowly and keep refining the parameters. the exchanges might not like that i am sharing this but code is the great equalizer and it is time for you to use it. i will be back tomorrow to show the results and keep building more systems until everything is fully automated

Moon Dev

11,948 次观看 • 5 个月前

🚨 The legendary Joe McMoneagle, the CIA’s Remote Viewer No. 1, played a pivotal role in the highly classified Stargate program, where he used remote viewing to penetrate the world’s deepest intelligence secrets. His sessions pinpointed undetectable nuclear submarines, exposed covert Soviet weapons programs, and provided life-saving intelligence during high-stakes missions in Vietnam and Europe, earning him the Legion of Merit. McMoneagle also describes out-of-body experiences and remote viewing sessions that revealed unusual structures on Mars—specifically pyramidal formations and fortress-like ruins. He claims to possess JPL negatives that match his descriptions and believes further analysis is needed. These findings come amid growing evidence that Mars once had rivers, a magnetosphere, and even fossilized bacteria—an idea that gained traction after NASA’s ALH 84001 meteorite discovery, President Clinton’s 1996 announcement, and analyses from plasma physicist John Brandenburg suggesting Mars may have suffered a nuclear catastrophe. From double-blind experiments demonstrating psi perception as a quantifiable skill to the precise methods behind remote viewing accuracy, McMoneagle challenges the limits of human perception. This episode explores the mechanics of consciousness, the future of intelligence gathering, and the controversial possibility that Mars may hold evidence of past civilization. Key Revelations: 1. Military-Grade Remote Viewing: Over 200 verified remote viewing sessions provided crucial intelligence from the late 1970s through the early 1980s. His work in identifying hidden Soviet submarine construction sites and military facilities led to major breakthroughs in U.S. defense strategy. In 1979, his session revealed an unknown Soviet sub with slanted missile tubes, later confirmed by satellite imagery. This intelligence helped shape U.S. naval countermeasures during the Cold War. 2. Unusual Structures on Mars: McMoneagle was tasked by the Department of Defense to remote view Mars at “1 million B.C.” He described massive pyramidal structures, collapsed fortifications, and signs of a lost civilization. He later obtained JPL negatives that he claims validate some of his descriptions. Recently, satellite imagery of a square anomaly on Mars—discussed by Elon Musk and mainstream scientists—has reignited interest in artificial-looking structures. If McMoneagle’s JPL negatives can be formally analyzed, they could contribute to the growing case that Mars’ past is more complex than officially acknowledged. 3. The Case for Life on Mars: McMoneagle’s claims align with a growing body of scientific evidence suggesting Mars may have supported life: • The 1996 ALH 84001 meteorite contained possible fossilized bacteria, prompting President Clinton’s public statement about potential Martian life. • Mars’ atmosphere contains Xenon-129 and Argon-40, which plasma physicist John Brandenburg argues are signatures of nuclear cataclysm—potential evidence of a large-scale ancient war. • Haim Eshed, former Israeli space security chief, has stated that human contact with extraterrestrials is ongoing and that the U.S. has knowledge of non-human presence on Mars. 4. Evolutionary Roots of Psychic Insight: McMoneagle believes remote viewing is an evolutionary ability that was once crucial for human survival. Before spoken language, early humans relied on telepathic perception to track predators, coordinate hunts, and sense threats—just as animals do today. Over time, this innate skill atrophied, but it remains latent within the human mind. 5. Refined Stargate Protocols: From 1972 to 1979, McMoneagle helped refine strict double-blind protocols for remote viewing. These included blind targeting, left-brain monitoring techniques, and strict data analysis procedures to separate real perceptions from imagination. These methods proved essential for achieving military-grade accuracy in intelligence gathering. 6. Double-Blind Validation & Telepathy: One of the most startling studies involved a Soviet experiment testing remote influence on biological targets. • Mice were divided into control and test groups through a randomized selection process. • Remote viewers were instructed to increase anxiety in the target mice. • Post-mortem chemical analysis confirmed heightened stress markers exclusively in the targeted group. McMoneagle also explores the mechanics of psi perception, emphasizing that knowing is a “sense” unto itself. This aligns with findings from The Telepathy Tapes, a hit podcast exploring cases of nonverbal communication in highly intuitive individuals, including autistic children with extraordinary clairvoyance. 7. Integration with Modern Technology: McMoneagle’s open-source remote viewing methodologies are now being integrated with artificial intelligence and pattern recognition software. Since the early 2000s, tests have explored whether AI-enhanced psi data analysis could detect non-human signatures hidden in vast data sets. 8. High-Stakes Intelligence Operations: Remote viewing played a role in critical military intelligence operations beyond Soviet submarine detection. • His sessions identified hidden crash sites at classified locations such as Fort Meade and Dugway Proving Grounds. • He tracked covert Soviet weapons facilities in real-time, allowing for rapid intelligence action when traditional surveillance failed. • The MX Missile Program was scrapped after McMoneagle and his team proved they could remotely locate mobile nuclear warheads—saving the U.S. over $100 billion. 9. The Power of Intent: Remote viewing accuracy hinges on disciplined intent and mental precision. • McMoneagle trains viewers to suppress ego and detach from analytical overlay, a process he refined through thousands of sessions. • In military applications, those who followed these strict focus protocols demonstrated significantly higher accuracy rates. 10. Future Contact with Aliens: McMoneagle believes remote viewing could be humanity’s first tool for non-physical contact with extraterrestrial intelligence. • He suggests that advanced civilizations may already be observing human development and that psi-based intelligence collection could be used to detect and interact with them. • While skeptics demand a “smoking gun,” he maintains that his Mars sessions are one more compelling data point in a much larger puzzle. The ultimate question: Can remote viewing be the key to unlocking interstellar diplomacy?
0:47

🚨 The legendary Joe McMoneagle, the CIA’s Remote Viewer No. 1, played a pivotal role in the highly classified Stargate program, where he used remote viewing to penetrate the world’s deepest intelligence secrets. His sessions pinpointed undetectable nuclear submarines, exposed covert Soviet weapons programs, and provided life-saving intelligence during high-stakes missions in Vietnam and Europe, earning him the Legion of Merit. McMoneagle also describes out-of-body experiences and remote viewing sessions that revealed unusual structures on Mars—specifically pyramidal formations and fortress-like ruins. He claims to possess JPL negatives that match his descriptions and believes further analysis is needed. These findings come amid growing evidence that Mars once had rivers, a magnetosphere, and even fossilized bacteria—an idea that gained traction after NASA’s ALH 84001 meteorite discovery, President Clinton’s 1996 announcement, and analyses from plasma physicist John Brandenburg suggesting Mars may have suffered a nuclear catastrophe. From double-blind experiments demonstrating psi perception as a quantifiable skill to the precise methods behind remote viewing accuracy, McMoneagle challenges the limits of human perception. This episode explores the mechanics of consciousness, the future of intelligence gathering, and the controversial possibility that Mars may hold evidence of past civilization. Key Revelations: 1. Military-Grade Remote Viewing: Over 200 verified remote viewing sessions provided crucial intelligence from the late 1970s through the early 1980s. His work in identifying hidden Soviet submarine construction sites and military facilities led to major breakthroughs in U.S. defense strategy. In 1979, his session revealed an unknown Soviet sub with slanted missile tubes, later confirmed by satellite imagery. This intelligence helped shape U.S. naval countermeasures during the Cold War. 2. Unusual Structures on Mars: McMoneagle was tasked by the Department of Defense to remote view Mars at “1 million B.C.” He described massive pyramidal structures, collapsed fortifications, and signs of a lost civilization. He later obtained JPL negatives that he claims validate some of his descriptions. Recently, satellite imagery of a square anomaly on Mars—discussed by Elon Musk and mainstream scientists—has reignited interest in artificial-looking structures. If McMoneagle’s JPL negatives can be formally analyzed, they could contribute to the growing case that Mars’ past is more complex than officially acknowledged. 3. The Case for Life on Mars: McMoneagle’s claims align with a growing body of scientific evidence suggesting Mars may have supported life: • The 1996 ALH 84001 meteorite contained possible fossilized bacteria, prompting President Clinton’s public statement about potential Martian life. • Mars’ atmosphere contains Xenon-129 and Argon-40, which plasma physicist John Brandenburg argues are signatures of nuclear cataclysm—potential evidence of a large-scale ancient war. • Haim Eshed, former Israeli space security chief, has stated that human contact with extraterrestrials is ongoing and that the U.S. has knowledge of non-human presence on Mars. 4. Evolutionary Roots of Psychic Insight: McMoneagle believes remote viewing is an evolutionary ability that was once crucial for human survival. Before spoken language, early humans relied on telepathic perception to track predators, coordinate hunts, and sense threats—just as animals do today. Over time, this innate skill atrophied, but it remains latent within the human mind. 5. Refined Stargate Protocols: From 1972 to 1979, McMoneagle helped refine strict double-blind protocols for remote viewing. These included blind targeting, left-brain monitoring techniques, and strict data analysis procedures to separate real perceptions from imagination. These methods proved essential for achieving military-grade accuracy in intelligence gathering. 6. Double-Blind Validation & Telepathy: One of the most startling studies involved a Soviet experiment testing remote influence on biological targets. • Mice were divided into control and test groups through a randomized selection process. • Remote viewers were instructed to increase anxiety in the target mice. • Post-mortem chemical analysis confirmed heightened stress markers exclusively in the targeted group. McMoneagle also explores the mechanics of psi perception, emphasizing that knowing is a “sense” unto itself. This aligns with findings from The Telepathy Tapes, a hit podcast exploring cases of nonverbal communication in highly intuitive individuals, including autistic children with extraordinary clairvoyance. 7. Integration with Modern Technology: McMoneagle’s open-source remote viewing methodologies are now being integrated with artificial intelligence and pattern recognition software. Since the early 2000s, tests have explored whether AI-enhanced psi data analysis could detect non-human signatures hidden in vast data sets. 8. High-Stakes Intelligence Operations: Remote viewing played a role in critical military intelligence operations beyond Soviet submarine detection. • His sessions identified hidden crash sites at classified locations such as Fort Meade and Dugway Proving Grounds. • He tracked covert Soviet weapons facilities in real-time, allowing for rapid intelligence action when traditional surveillance failed. • The MX Missile Program was scrapped after McMoneagle and his team proved they could remotely locate mobile nuclear warheads—saving the U.S. over $100 billion. 9. The Power of Intent: Remote viewing accuracy hinges on disciplined intent and mental precision. • McMoneagle trains viewers to suppress ego and detach from analytical overlay, a process he refined through thousands of sessions. • In military applications, those who followed these strict focus protocols demonstrated significantly higher accuracy rates. 10. Future Contact with Aliens: McMoneagle believes remote viewing could be humanity’s first tool for non-physical contact with extraterrestrial intelligence. • He suggests that advanced civilizations may already be observing human development and that psi-based intelligence collection could be used to detect and interact with them. • While skeptics demand a “smoking gun,” he maintains that his Mars sessions are one more compelling data point in a much larger puzzle. The ultimate question: Can remote viewing be the key to unlocking interstellar diplomacy?

Jesse Michels

263,672 次观看 • 1 年前

la dernière fois je vous ai parlé du réseau ferroviaire chinois qui roule du désert brûlant au grand nord gelé, aujourd'hui si vous prévoyez d'y aller (et je vous y encourage) , allez voir les ponts, vous allez voir ces ouvrages posés dans des gorges où vous vous demandez juste comment un putain d’humain a pu bâtir ça mdrrrrr le dernier c'est le pont du grand canyon de Huajiang dans le guizhou ouvert en 2025 et qui est le + haut pont du monde, 625 mètres entre le tablier et la rivière…. pour vous donner un aperçu du gigantisme sachez que vous pourriez glisser l’empire state à NYC en dessous et il resterait ENCORE de la marge mdrrr et pour info il enjambe une faille si profonde qu'on la surnomme la fissure de la terre et fait passer la traversée de 2 heures à 2 minutes maintenant de mon côté j'étais tellement obsédé par l'ouvrage que je suis allé éplucher ses spécifications techniques et ses notes de génie civil et voilà le vrai tour de force, au dessus d'un tel vide impossible de construire depuis le bas, alors ils ont fabriqué le tablier en usine, 93 sections faites de plus de 100 000 pièces d'acier acheminées par la route puis hissées par câbles depuis le centre vers les 2 rives jusqu'à se rejoindre au dessus du gouffre, y’a 22 000 tonnes d'acier soit + du double de la Tour Eiffel et pour dompter les vents violents ducanyon ils ont modélisé les rafales au lidar et en soufflerie puis ajouté visiblement ds déflecteurs et amortisseurs pour que rien ne bouge à + de 600 mètres mais le truc qui doit vraiment vous scotcher c'est le délai, de la première pierre en janvier 2022 à l'ouverture, y’a eu 3 ans et 8 mois pour le plus haut pont de la planète dans le karst le plus hostile qui soit et le guizhou une des provinces les plus pauvres de Chine concentre à lui seul près de la moitié des 100 ponts les + hauts du monde pour moi c’est clairement le genre de une civilisation qui traite l'impossible comme une simple question de calendrier et je trouve ça absolument incroyable MAIS ce qui me marque le plus dans tout ça c’est que la Chine ne laisse jamais une région crever dans son coin, elle la connecte & vous voyez ce ce pont dans une province pauvre c'est de l'économie qui débarque, du tourisme des flux logistiques, des data centers et des centaines de milliers d'emplois créés sur place, résultat un habitant qui cherche du travail le trouve près de chez lui!!! chez nous c'est l'inverse, on laisse mourir des territoires entiers, on ferme les lignes, les services et les usines et on demande aux gens de partir à des dizaines de kilomètres pour bosser, désenclaver un territoire c'est le premier acte d'un pays qui croit encore en tous ses habitants et juste pour ça la Chine mérite un grand respect!! voilà, voilà ;)
0:29

la dernière fois je vous ai parlé du réseau ferroviaire chinois qui roule du désert brûlant au grand nord gelé, aujourd'hui si vous prévoyez d'y aller (et je vous y encourage) , allez voir les ponts, vous allez voir ces ouvrages posés dans des gorges où vous vous demandez juste comment un putain d’humain a pu bâtir ça mdrrrrr le dernier c'est le pont du grand canyon de Huajiang dans le guizhou ouvert en 2025 et qui est le + haut pont du monde, 625 mètres entre le tablier et la rivière…. pour vous donner un aperçu du gigantisme sachez que vous pourriez glisser l’empire state à NYC en dessous et il resterait ENCORE de la marge mdrrr et pour info il enjambe une faille si profonde qu'on la surnomme la fissure de la terre et fait passer la traversée de 2 heures à 2 minutes maintenant de mon côté j'étais tellement obsédé par l'ouvrage que je suis allé éplucher ses spécifications techniques et ses notes de génie civil et voilà le vrai tour de force, au dessus d'un tel vide impossible de construire depuis le bas, alors ils ont fabriqué le tablier en usine, 93 sections faites de plus de 100 000 pièces d'acier acheminées par la route puis hissées par câbles depuis le centre vers les 2 rives jusqu'à se rejoindre au dessus du gouffre, y’a 22 000 tonnes d'acier soit + du double de la Tour Eiffel et pour dompter les vents violents ducanyon ils ont modélisé les rafales au lidar et en soufflerie puis ajouté visiblement ds déflecteurs et amortisseurs pour que rien ne bouge à + de 600 mètres mais le truc qui doit vraiment vous scotcher c'est le délai, de la première pierre en janvier 2022 à l'ouverture, y’a eu 3 ans et 8 mois pour le plus haut pont de la planète dans le karst le plus hostile qui soit et le guizhou une des provinces les plus pauvres de Chine concentre à lui seul près de la moitié des 100 ponts les + hauts du monde pour moi c’est clairement le genre de une civilisation qui traite l'impossible comme une simple question de calendrier et je trouve ça absolument incroyable MAIS ce qui me marque le plus dans tout ça c’est que la Chine ne laisse jamais une région crever dans son coin, elle la connecte & vous voyez ce ce pont dans une province pauvre c'est de l'économie qui débarque, du tourisme des flux logistiques, des data centers et des centaines de milliers d'emplois créés sur place, résultat un habitant qui cherche du travail le trouve près de chez lui!!! chez nous c'est l'inverse, on laisse mourir des territoires entiers, on ferme les lignes, les services et les usines et on demande aux gens de partir à des dizaines de kilomètres pour bosser, désenclaver un territoire c'est le premier acte d'un pays qui croit encore en tous ses habitants et juste pour ça la Chine mérite un grand respect!! voilà, voilà ;)

Mehdi (e/λ)

35,832 次观看 • 1 个月前

$NVDA $MU $SNDK $LITE PAPER OVERVIEW AND CORE CLAIMS The paper “KV Cache Transform Coding for Compact Storage in LLM Inference” introduces kvtc, a transform-coding pipeline that compresses transformer key-value (KV) caches primarily for storage and transfer in LLM serving, rather than for accelerating the per-token attention kernel during active decoding. The method combines 3 stages: (1) feature decorrelation via a PCA basis computed from a calibration dataset and reused across requests; (2) adaptive, variable-precision quantization with bit allocation solved via dynamic programming (DP), including groupwise scaling/shift overhead; and (3) lossless entropy coding (DEFLATE via nvCOMP in the reference implementation) to exploit residual redundancy after quantization. The central empirical claim is that KV tensors contain large, exploitable redundancy across heads and layers, enabling approximately 20× compression versus a 16-bit baseline with negligible degradation across a broad set of accuracy and long-context benchmarks, with materially higher compression (≥40×) available at modest quality cost in some regimes. The system claim is that such compression materially improves the economics of multi-turn, prefix-reuse serving by extending effective KV cache capacity in GPU HBM and host tiers (DRAM/NVMe) and by reducing inter-node and GPU↔host bandwidth demands, thereby improving cache hit rates and reducing time-to-first-token (TTFT) relative to recomputation when caches would otherwise be evicted. KV CACHE AS THE DOMINANT STATE VARIABLE IN INFERENCE ECONOMICS KV cache growth is linear in context length and is multiplicative in layers and attention heads, making it an increasingly dominant constraint as (a) context lengths expand, (b) models add layers and maintain large hidden dimensions, and (c) production workloads shift toward iterative and tool-augmented interactions that repeatedly reuse long prefixes. The paper uses the canonical 16-bit KV cache size formula (4·l·h·d_head·t) bytes and reports 16-bit KV cache sizes per 1K tokens of context that are already operationally large: 128MiB for Llama 3.1 8B, 160MiB for Mistral NeMo 12B, and 320MiB for Llama 3.3 70B Instruct. In binary units, these figures imply per-token KV footprints of 128KiB/token (Llama 3.1 8B), 160KiB/token (Mistral NeMo 12B), and 320KiB/token (Llama 3.3 70B Instruct) at 16-bit. For a 10K-token prompt (10×1K in the paper’s binary convention), the 16-bit KV cache sizes scale to approximately 1.25GiB (Llama 3.1 8B), 1.56GiB (Mistral NeMo 12B), and 3.13GiB (Llama 3.3 70B Instruct). These magnitudes explain why stale caches create a throughput–latency dilemma: retaining them in HBM maximizes responsiveness on future turns but crowds out concurrent sessions; evicting them forces quadratic-cost prefill recomputation and increases TTFT; offloading them to host or storage introduces large transfer overhead and consumes DRAM/NVMe capacity. A key operational nuance emphasized is that modern serving stacks increasingly treat KV caches as a database, leveraging block paging and shared-prefix reuse. In the common disaggregated serving design (separate prefill and decode nodes), KV cache transfer becomes a dominant category of cross-node traffic. Under that design, any reduction in KV cache size directly increases effective fabric capacity and reduces tail latency attributable to congestion, while also enabling longer cache lifetimes in “hot” (HBM) and “warm” (CPU DRAM) tiers that raise cache hit rates and reduce recomputation frequency. The paper’s quantitative example illustrates the economic stakes: a 1,000-line code file tokenized at ~10 tokens/line yields ~10K tokens; for Llama 3.3 70B, an 8-bit KV cache for that context is ~1.6GiB. Reuse across subsequent turns or parallel chats around the same file is valuable, but HBM scarcity makes retaining many such caches infeasible without compression. TECHNICAL MECHANISM: WHY KV CACHES ARE COMPRESSIBLE AND HOW KVTC EXPLOITS IT The technical rationale begins with an empirical observation: keys (and, to a lesser extent, values) across different attention heads can be aligned into a shared latent space using orthogonal transformations (Procrustes alignment). This supports the hypothesis that head-specific projections introduce rotations of a common subspace rather than completely distinct information, implying that concatenating across heads and layers should reveal low-rank structure suitable for linear decorrelation and dimensionality reduction. The method operationalizes this using a PCA/SVD basis learned from calibration data rather than recomputing a decomposition per prompt. This design choice targets production viability: per-prompt SVD is computationally expensive and scales poorly with long prompts and frequent cache updates. kvtc is explicitly structured as an offline-calibrated, online-applied codec: Calibration (performed 1 time per model and compression setting for DP allocation) A calibration dataset is forwarded through the model to collect KV caches. Token positions are pooled, and a subset of positions is sampled. Keys and values are processed separately. Several implementation choices are highlighted as decisive for stability: Rotary positional embeddings are effectively removed prior to compression (“undo positional rotations”), because positional rotations degrade the apparent low-rank structure of keys. “Attention sink” tokens (the earliest tokens in the sequence) and a sliding window of most recent tokens are excluded from compression because they disproportionately affect attention patterns and are empirically more sensitive to reconstruction error. Cross-layer concatenation is used: keys (or values) from multiple layers and heads at the same token position are concatenated along the feature axis to form a higher-dimensional feature vector. PCA is computed over these concatenated vectors, improving robustness relative to per-layer or per-head PCA. The PCA basis is computed via SVD of centered calibration data, using randomized SVD for scalability with a target rank cutoff. The paper reports calibration regimes of 160K tokens for several models with a 10K PCA dimension cutoff (8K for Qwen variants with fewer KV heads), selected to fit within a single 80GB H100 memory envelope and complete within minutes. A critical economic detail is that the same PCA basis can be reused across multiple compression ratios; only the DP-derived precision assignment changes per compression target. Compression (applied between inference phases) Compression operates on stored KV cache tensors, not on weights, and does not modify attention computation. The KV cache is projected into the PCA basis, quantized, packed, and then entropy-coded. Compression is positioned as a background or between-phase operation (after decoding, or between prefill and decode), executed on GPU or CPU depending on where the cache currently resides. The design intent is that compression should not sit on the critical per-token decoding path; it is a storage and transport optimization. Decompression (performed prior to reuse) Decompression reverses the entropy coding and quantization and applies the inverse PCA projection. A practical latency optimization is proposed: inverse projection can be performed layer-by-layer using submatrices of the PCA basis, allowing generation to begin before the full cache is reconstructed, reducing TTFT. Quantization and bit allocation are the core differentiators versus simpler PCA truncation. PCA provides ordered components by variance; kvtc uses DP to allocate a global bit budget across PCA coordinates (and across groups of coordinates) to minimize reconstruction error in the decorrelated domain. Groups of subsequent PCA coordinates share 16-bit shift and scale factors (a microscaling-inspired design), and the DP algorithm jointly selects group size and precision type under a bit budget, including the overhead of per-group metadata. DP commonly assigns 0 bits to many trailing PCA components, which both increases compression and provides a mechanism to trim the PCA basis to the subset of components that actually carry payload, reducing compute and storage overhead of the projection matrices in deployment. Lossless entropy coding then exploits the structure induced by quantization. DEFLATE is used in the reference implementation, and the paper emphasizes that the incremental gain from the lossless stage is content-dependent but meaningful, with an average uplift of ~1.23× on top of quantization in the reported regime. An ablation in the appendices indicates that GPU-friendly variants (GDeflate) can achieve nearly identical compression ratios (≤0.1 difference in measured cases), implying that throughput-optimized lossless codecs can likely be substituted without sacrificing meaningful compression. EMPIRICAL RESULTS: ACCURACY, COMPRESSION, AND LATENCY General-purpose 8B–12B dense models The paper evaluates Llama 3.1 8B, MN-Minitron 8B, and Mistral NeMo 12B across math/knowledge (GSM8K, MMLU) and long-context tasks (Qasper, Lost in the Middle, RULER Variable Tracking) under a simulated multi-turn regime where compression/decompression is applied periodically, with a sliding window of recent tokens excluded. A consistent pattern appears: kvtc maintains near-vanilla performance through 16× compression settings, and remains competitive at 32×, with degradation becoming task- and model-dependent at 64×, particularly on long-context retrieval metrics when compression is pushed aggressively. Selected quantitative anchor points from the paper’s standard-error table (all values are reported with the paper’s evaluation setup and token-window exclusions): Llama 3.1 8B Vanilla: GSM8K 56.8, MMLU 60.5, Qasper 40.4, LITM 99.4, RULER-VT 99.8 kvtc16×: GSM8K 56.9, MMLU 60.1, Qasper 40.7, LITM 99.3, RULER-VT 99.1 kvtc32×: GSM8K 57.8, MMLU 60.6, Qasper 39.4, LITM 99.1, RULER-VT 98.9 kvtc64×: GSM8K 57.2, MMLU 60.7, Qasper 37.8, LITM 90.2, RULER-VT 95.9 These results indicate that, for this model, long-context sensitivity emerges at 64× with meaningful drops in LITM and RULER-VT, while math/knowledge scores remain stable, implying a differential sensitivity consistent with key-vector precision being more critical for retrieval-style behavior. Mistral NeMo 12B Vanilla: GSM8K 61.9, MMLU 64.5, Qasper 38.4, LITM 99.5, RULER-VT 99.8 kvtc16×: GSM8K 62.0, MMLU 64.4, Qasper 37.6, LITM 99.8, RULER-VT 99.5 kvtc32×: GSM8K 62.2, MMLU 63.8, Qasper 37.5, LITM 99.6, RULER-VT 98.7 kvtc64×: GSM8K 61.9, MMLU 61.4, Qasper 38.0, LITM 95.3, RULER-VT 98.0 Here, degradation at 64× is visible but materially smaller than the Llama 3.1 8B LITM drop, suggesting model-architecture or training-data differences can change the tolerance envelope for aggressive KV cache distortion. MN-Minitron 8B Vanilla: GSM8K 59.1, MMLU 64.3, Qasper 38.2, LITM 99.8, RULER-VT 99.4 kvtc16×: GSM8K 60.3, MMLU 64.1, Qasper 38.6, LITM 99.3, RULER-VT 98.8 kvtc32×: GSM8K 59.1, MMLU 63.7, Qasper 37.7, LITM 86.9, RULER-VT 96.0 kvtc64×: GSM8K 57.8, MMLU 62.1, Qasper 38.1, LITM 59.5, RULER-VT 93.4 This model shows markedly higher sensitivity on LITM at 32× and 64×, despite stable short-context metrics, reinforcing that “compression safety” is not monotonic in parameter count and that pruning/distillation choices can alter KV cache redundancy or robustness. Comparisons to baselines The paper compares kvtc to quantization baselines (KIVI, GEAR, FP8) and eviction baselines (H2O, TOVA), plus an SVD-based prefill-optimization method (xKV). Across the reported tasks: Low-bit quantization methods at modest compression (2-bit KV schemes) show earlier degradation in long-context behavior than kvtc at substantially higher compression settings. Eviction methods perform poorly as generic compressors for long-context tasks, consistent with their objective function (selective pruning) being misaligned with “lossless-ish storage for reuse.” xKV shows competitive results on some tasks but a consistent underperformance on Qasper relative to kvtc and vanilla in the provided tables, consistent with method-specific distortions introduced by its decomposition regime. Reasoning models and high-variance tasks For DeepSeek-R1-distilled Qwen 2.5 reasoning models, the paper evaluates AIME 2024/2025 and LiveCodeBench coding. Results are averaged over 8 runs with large variance, but a key inference is that kvtc at ~9×–21× compression achieves broadly similar AIME scores within variance bands, while coding performance remains stable at ~9× and degrades more visibly at ~18×–21× on the 7B model. An important nuance is that smaller reasoning models already have smaller KV footprints (reported ~29KiB/token for Qwen R1 1.5B versus 131KiB/token for Llama 3.1 8B), so the economic value of aggressive KV cache compression is proportionally higher for large models and long contexts than for small models with short contexts, unless the serving system’s bottleneck is dominated by cache transfer rather than HBM capacity. Multi-GPU inference and pipeline parallel For Llama 3.3 70B Instruct run pipeline-parallel across 4 GPUs (20 layers per GPU), the paper compresses KV cache chunks independently per GPU. On MATH-500, the reported accuracy declines from 75.6 (vanilla) to 74.4 at 10× and 72.6 at 20×, with standard errors near ~1.9. NIAH and LITM remain at 100.0 for all tested ratios in that table. The paper notes that joint compression across chunks could improve accuracy for some offload scenarios but is not required for feasibility, highlighting an engineering trade-off between deployment simplicity in distributed settings and optimal global compression. Latency and TTFT economics A critical system result is the measured compression/decompression latency on an H100 for a non-fused implementation. For Mistral NeMo 12B in bfloat16: BS=8, CTX=8K: compression 379ms, decompression 267ms; vanilla recompute TTFT 3098ms; kvtc decompression TTFT 380ms BS=2, CTX=16K: compression 194ms, decompression 143ms; vanilla recompute TTFT 1780ms; kvtc decompression TTFT 208ms These measurements imply that, when a cache would otherwise be recomputed, decompressing a stored compressed cache can reduce TTFT by ~8×–9× in these scenarios, even without kernel fusion. The decomposition of runtime shows PCA projection and entropy coding as the largest contributors, implying that GPU-optimized kernels and faster GPU-native lossless codecs could reduce overhead further. The fundamental economic conclusion is that, in multi-turn settings with long prefixes, compression-induced overhead is likely dominated by the avoided prefill compute and avoided transfer overhead for uncompressed caches. KEY DEPLOYMENT-SENSITIVE DESIGN CHOICES AND FAILURE MODES Several design choices appear to be “hard requirements” rather than optional optimizations: Sink tokens and sliding window exclusions The paper’s ablations show that compressing early “sink” tokens can catastrophically degrade accuracy at high compression ratios (example: Llama 3.1 8B at 64× collapses on multiple tasks when sink tokens are compressed). Similarly, compressing the most recent tokens hurts performance, motivating a sliding window (default 128 tokens) that remains uncompressed. This introduces a predictable engineering constraint: kvtc is not a uniform compression of the full cache; it is a policy-driven, token-position-dependent codec. Production integration therefore requires correct handling of token positions, attention sinks, and window management, and these policies must be aligned with attention-kernel behavior and model-specific sink dynamics. RoPE handling Removing positional rotations prior to compression is described as important for preserving low-rank structure. In deployment, this implies that the codec must be position-aware and must invert and reapply RoPE correctly. This is an additional source of complexity relative to pure per-token quantization and is sensitive to model variants and RoPE parameterizations. Calibration set representativeness The method’s quality hinges on the PCA basis generalizing from calibration data to production data. The paper demonstrates relative stability with 160K–200K calibration tokens and explores domain shifts (general web text vs math traces vs code). Results suggest that moderate domain mismatch is tolerated at 16×–64×, while extreme compression (e.g., 256× in ablations) becomes materially more sensitive to calibration choice. In production, this implies that operators targeting the “negligible degradation” regime should be able to calibrate with broadly representative corpora, while operators targeting ultra-high compression for specialized workloads should expect tighter coupling between calibration domain and achieved quality. PCA matrix storage overhead and operational footprint A non-trivial hidden cost is the need to store PCA projection matrices per model. The paper reports that, prior to DP trimming, PCA matrices stored at 16-bit can amount to a meaningful fraction of model parameter count (examples reported: ~2.4% for Llama 3.3 70B, ~8.7% for Llama 3.1 8B). This overhead is amortized across all cached sessions for a model but competes with HBM/DRAM budgets in multi-model serving. DP-driven trimming can reduce this overhead at higher compression ratios by removing zero-bit components, but the directionality is not guaranteed at low compression ratios if many components remain active. In distributed inference (pipeline parallel), per-chunk PCA can reduce matrix sizes, but may reduce cross-layer decorrelation benefits if fewer layers are concatenated. SYSTEM-LEVEL IMPLICATIONS FOR GENERATIVE AI INFRASTRUCTURE GPU AND HBM The principal infrastructure implication is that KV cache compression at storage time targets the dominant memory allocator stressor in stateful serving: the accumulation of idle or warm conversation state. For workloads with long reusable prefixes (code assistants, enterprise agents with large system prompts, repeated RAG scaffolds, document chat), the limiting resource frequently becomes HBM reserved for KV caches rather than compute. By compressing stale caches by ~20× (or more), the same HBM budget can retain a materially larger working set of cached prefixes, increasing cache hit rates and reducing recomputation. This effect is multiplicative with cache-aware routing and prefix sharing: more prefixes can remain resident (hot or warm) and can be routed to nodes that already hold them, improving both throughput and tail latency. However, kvtc as described does not reduce the active KV cache footprint during the actual attention computation for a currently decoding sequence, because the model operates on decompressed KV caches during decoding. Therefore, the method does not directly reduce HBM bandwidth consumed by attention kernels during steady-state decode, and does not directly address the “memory traffic per generated token” bottleneck that motivates online KV quantization and eviction strategies. The primary HBM benefit is increased effective capacity for caches between turns and reduced HBM pressure from storing many idle sessions, not reduced per-token decode bandwidth. Compression and decompression themselves consume GPU compute and memory bandwidth. The measured decompression TTFT of ~208ms–380ms in the provided benchmarks indicates that the overhead is real but can be materially smaller than recomputation of long prefixes. In an HBM-constrained serving environment, this overhead can be interpreted as a trade between (a) maintaining more caches warm and paying decompression on reuse versus (b) evicting caches and paying full prefill recomputation. The decision boundary will depend on distribution of inter-turn idle times, probability of reuse, and SLA sensitivity to TTFT. kvtc expands the feasible region where keeping caches is economically rational, especially for long prompts. CPU AND DRAM The method implies a stronger role for CPU DRAM as a warm KV cache tier. A ~20× compression ratio changes the practical scale of “warm state” that can be stored per server. Using the paper’s reported KV cache sizes, a 10K-token 16-bit KV cache for Llama 3.3 70B is ~3.13GiB; compressing by ~20× would reduce this to ~160MiB. At that size, storing hundreds to thousands of warm conversation states in DRAM becomes materially more feasible, increasing cache hit rates and reducing NVMe dependence. This can shift system design from “HBM-only hot caches with aggressive eviction” toward “HBM hot + DRAM warm with long retention,” which is structurally analogous to CPU page cache hierarchies in classical systems design. CPU compute implications depend on where compression is executed. The paper explicitly allows compression on CPU if the cache is already in storage, but the strongest bandwidth savings are achieved when compression happens before moving KV caches off the GPU. If an operator chooses GPU-side compression prior to PCIe/NVLink transfer, CPU compute overhead is modest (orchestrating and DP calibration offline). If an operator instead transfers uncompressed caches to CPU for compression, bandwidth savings are forfeited and CPU memory bandwidth becomes a bottleneck. Therefore, the most economically coherent deployment path is GPU-native compression/decompression with CPU DRAM used as the warm storage reservoir.
12:09

$NVDA $MU $SNDK $LITE PAPER OVERVIEW AND CORE CLAIMS The paper “KV Cache Transform Coding for Compact Storage in LLM Inference” introduces kvtc, a transform-coding pipeline that compresses transformer key-value (KV) caches primarily for storage and transfer in LLM serving, rather than for accelerating the per-token attention kernel during active decoding. The method combines 3 stages: (1) feature decorrelation via a PCA basis computed from a calibration dataset and reused across requests; (2) adaptive, variable-precision quantization with bit allocation solved via dynamic programming (DP), including groupwise scaling/shift overhead; and (3) lossless entropy coding (DEFLATE via nvCOMP in the reference implementation) to exploit residual redundancy after quantization. The central empirical claim is that KV tensors contain large, exploitable redundancy across heads and layers, enabling approximately 20× compression versus a 16-bit baseline with negligible degradation across a broad set of accuracy and long-context benchmarks, with materially higher compression (≥40×) available at modest quality cost in some regimes. The system claim is that such compression materially improves the economics of multi-turn, prefix-reuse serving by extending effective KV cache capacity in GPU HBM and host tiers (DRAM/NVMe) and by reducing inter-node and GPU↔host bandwidth demands, thereby improving cache hit rates and reducing time-to-first-token (TTFT) relative to recomputation when caches would otherwise be evicted. KV CACHE AS THE DOMINANT STATE VARIABLE IN INFERENCE ECONOMICS KV cache growth is linear in context length and is multiplicative in layers and attention heads, making it an increasingly dominant constraint as (a) context lengths expand, (b) models add layers and maintain large hidden dimensions, and (c) production workloads shift toward iterative and tool-augmented interactions that repeatedly reuse long prefixes. The paper uses the canonical 16-bit KV cache size formula (4·l·h·d_head·t) bytes and reports 16-bit KV cache sizes per 1K tokens of context that are already operationally large: 128MiB for Llama 3.1 8B, 160MiB for Mistral NeMo 12B, and 320MiB for Llama 3.3 70B Instruct. In binary units, these figures imply per-token KV footprints of 128KiB/token (Llama 3.1 8B), 160KiB/token (Mistral NeMo 12B), and 320KiB/token (Llama 3.3 70B Instruct) at 16-bit. For a 10K-token prompt (10×1K in the paper’s binary convention), the 16-bit KV cache sizes scale to approximately 1.25GiB (Llama 3.1 8B), 1.56GiB (Mistral NeMo 12B), and 3.13GiB (Llama 3.3 70B Instruct). These magnitudes explain why stale caches create a throughput–latency dilemma: retaining them in HBM maximizes responsiveness on future turns but crowds out concurrent sessions; evicting them forces quadratic-cost prefill recomputation and increases TTFT; offloading them to host or storage introduces large transfer overhead and consumes DRAM/NVMe capacity. A key operational nuance emphasized is that modern serving stacks increasingly treat KV caches as a database, leveraging block paging and shared-prefix reuse. In the common disaggregated serving design (separate prefill and decode nodes), KV cache transfer becomes a dominant category of cross-node traffic. Under that design, any reduction in KV cache size directly increases effective fabric capacity and reduces tail latency attributable to congestion, while also enabling longer cache lifetimes in “hot” (HBM) and “warm” (CPU DRAM) tiers that raise cache hit rates and reduce recomputation frequency. The paper’s quantitative example illustrates the economic stakes: a 1,000-line code file tokenized at ~10 tokens/line yields ~10K tokens; for Llama 3.3 70B, an 8-bit KV cache for that context is ~1.6GiB. Reuse across subsequent turns or parallel chats around the same file is valuable, but HBM scarcity makes retaining many such caches infeasible without compression. TECHNICAL MECHANISM: WHY KV CACHES ARE COMPRESSIBLE AND HOW KVTC EXPLOITS IT The technical rationale begins with an empirical observation: keys (and, to a lesser extent, values) across different attention heads can be aligned into a shared latent space using orthogonal transformations (Procrustes alignment). This supports the hypothesis that head-specific projections introduce rotations of a common subspace rather than completely distinct information, implying that concatenating across heads and layers should reveal low-rank structure suitable for linear decorrelation and dimensionality reduction. The method operationalizes this using a PCA/SVD basis learned from calibration data rather than recomputing a decomposition per prompt. This design choice targets production viability: per-prompt SVD is computationally expensive and scales poorly with long prompts and frequent cache updates. kvtc is explicitly structured as an offline-calibrated, online-applied codec: Calibration (performed 1 time per model and compression setting for DP allocation) A calibration dataset is forwarded through the model to collect KV caches. Token positions are pooled, and a subset of positions is sampled. Keys and values are processed separately. Several implementation choices are highlighted as decisive for stability: Rotary positional embeddings are effectively removed prior to compression (“undo positional rotations”), because positional rotations degrade the apparent low-rank structure of keys. “Attention sink” tokens (the earliest tokens in the sequence) and a sliding window of most recent tokens are excluded from compression because they disproportionately affect attention patterns and are empirically more sensitive to reconstruction error. Cross-layer concatenation is used: keys (or values) from multiple layers and heads at the same token position are concatenated along the feature axis to form a higher-dimensional feature vector. PCA is computed over these concatenated vectors, improving robustness relative to per-layer or per-head PCA. The PCA basis is computed via SVD of centered calibration data, using randomized SVD for scalability with a target rank cutoff. The paper reports calibration regimes of 160K tokens for several models with a 10K PCA dimension cutoff (8K for Qwen variants with fewer KV heads), selected to fit within a single 80GB H100 memory envelope and complete within minutes. A critical economic detail is that the same PCA basis can be reused across multiple compression ratios; only the DP-derived precision assignment changes per compression target. Compression (applied between inference phases) Compression operates on stored KV cache tensors, not on weights, and does not modify attention computation. The KV cache is projected into the PCA basis, quantized, packed, and then entropy-coded. Compression is positioned as a background or between-phase operation (after decoding, or between prefill and decode), executed on GPU or CPU depending on where the cache currently resides. The design intent is that compression should not sit on the critical per-token decoding path; it is a storage and transport optimization. Decompression (performed prior to reuse) Decompression reverses the entropy coding and quantization and applies the inverse PCA projection. A practical latency optimization is proposed: inverse projection can be performed layer-by-layer using submatrices of the PCA basis, allowing generation to begin before the full cache is reconstructed, reducing TTFT. Quantization and bit allocation are the core differentiators versus simpler PCA truncation. PCA provides ordered components by variance; kvtc uses DP to allocate a global bit budget across PCA coordinates (and across groups of coordinates) to minimize reconstruction error in the decorrelated domain. Groups of subsequent PCA coordinates share 16-bit shift and scale factors (a microscaling-inspired design), and the DP algorithm jointly selects group size and precision type under a bit budget, including the overhead of per-group metadata. DP commonly assigns 0 bits to many trailing PCA components, which both increases compression and provides a mechanism to trim the PCA basis to the subset of components that actually carry payload, reducing compute and storage overhead of the projection matrices in deployment. Lossless entropy coding then exploits the structure induced by quantization. DEFLATE is used in the reference implementation, and the paper emphasizes that the incremental gain from the lossless stage is content-dependent but meaningful, with an average uplift of ~1.23× on top of quantization in the reported regime. An ablation in the appendices indicates that GPU-friendly variants (GDeflate) can achieve nearly identical compression ratios (≤0.1 difference in measured cases), implying that throughput-optimized lossless codecs can likely be substituted without sacrificing meaningful compression. EMPIRICAL RESULTS: ACCURACY, COMPRESSION, AND LATENCY General-purpose 8B–12B dense models The paper evaluates Llama 3.1 8B, MN-Minitron 8B, and Mistral NeMo 12B across math/knowledge (GSM8K, MMLU) and long-context tasks (Qasper, Lost in the Middle, RULER Variable Tracking) under a simulated multi-turn regime where compression/decompression is applied periodically, with a sliding window of recent tokens excluded. A consistent pattern appears: kvtc maintains near-vanilla performance through 16× compression settings, and remains competitive at 32×, with degradation becoming task- and model-dependent at 64×, particularly on long-context retrieval metrics when compression is pushed aggressively. Selected quantitative anchor points from the paper’s standard-error table (all values are reported with the paper’s evaluation setup and token-window exclusions): Llama 3.1 8B Vanilla: GSM8K 56.8, MMLU 60.5, Qasper 40.4, LITM 99.4, RULER-VT 99.8 kvtc16×: GSM8K 56.9, MMLU 60.1, Qasper 40.7, LITM 99.3, RULER-VT 99.1 kvtc32×: GSM8K 57.8, MMLU 60.6, Qasper 39.4, LITM 99.1, RULER-VT 98.9 kvtc64×: GSM8K 57.2, MMLU 60.7, Qasper 37.8, LITM 90.2, RULER-VT 95.9 These results indicate that, for this model, long-context sensitivity emerges at 64× with meaningful drops in LITM and RULER-VT, while math/knowledge scores remain stable, implying a differential sensitivity consistent with key-vector precision being more critical for retrieval-style behavior. Mistral NeMo 12B Vanilla: GSM8K 61.9, MMLU 64.5, Qasper 38.4, LITM 99.5, RULER-VT 99.8 kvtc16×: GSM8K 62.0, MMLU 64.4, Qasper 37.6, LITM 99.8, RULER-VT 99.5 kvtc32×: GSM8K 62.2, MMLU 63.8, Qasper 37.5, LITM 99.6, RULER-VT 98.7 kvtc64×: GSM8K 61.9, MMLU 61.4, Qasper 38.0, LITM 95.3, RULER-VT 98.0 Here, degradation at 64× is visible but materially smaller than the Llama 3.1 8B LITM drop, suggesting model-architecture or training-data differences can change the tolerance envelope for aggressive KV cache distortion. MN-Minitron 8B Vanilla: GSM8K 59.1, MMLU 64.3, Qasper 38.2, LITM 99.8, RULER-VT 99.4 kvtc16×: GSM8K 60.3, MMLU 64.1, Qasper 38.6, LITM 99.3, RULER-VT 98.8 kvtc32×: GSM8K 59.1, MMLU 63.7, Qasper 37.7, LITM 86.9, RULER-VT 96.0 kvtc64×: GSM8K 57.8, MMLU 62.1, Qasper 38.1, LITM 59.5, RULER-VT 93.4 This model shows markedly higher sensitivity on LITM at 32× and 64×, despite stable short-context metrics, reinforcing that “compression safety” is not monotonic in parameter count and that pruning/distillation choices can alter KV cache redundancy or robustness. Comparisons to baselines The paper compares kvtc to quantization baselines (KIVI, GEAR, FP8) and eviction baselines (H2O, TOVA), plus an SVD-based prefill-optimization method (xKV). Across the reported tasks: Low-bit quantization methods at modest compression (2-bit KV schemes) show earlier degradation in long-context behavior than kvtc at substantially higher compression settings. Eviction methods perform poorly as generic compressors for long-context tasks, consistent with their objective function (selective pruning) being misaligned with “lossless-ish storage for reuse.” xKV shows competitive results on some tasks but a consistent underperformance on Qasper relative to kvtc and vanilla in the provided tables, consistent with method-specific distortions introduced by its decomposition regime. Reasoning models and high-variance tasks For DeepSeek-R1-distilled Qwen 2.5 reasoning models, the paper evaluates AIME 2024/2025 and LiveCodeBench coding. Results are averaged over 8 runs with large variance, but a key inference is that kvtc at ~9×–21× compression achieves broadly similar AIME scores within variance bands, while coding performance remains stable at ~9× and degrades more visibly at ~18×–21× on the 7B model. An important nuance is that smaller reasoning models already have smaller KV footprints (reported ~29KiB/token for Qwen R1 1.5B versus 131KiB/token for Llama 3.1 8B), so the economic value of aggressive KV cache compression is proportionally higher for large models and long contexts than for small models with short contexts, unless the serving system’s bottleneck is dominated by cache transfer rather than HBM capacity. Multi-GPU inference and pipeline parallel For Llama 3.3 70B Instruct run pipeline-parallel across 4 GPUs (20 layers per GPU), the paper compresses KV cache chunks independently per GPU. On MATH-500, the reported accuracy declines from 75.6 (vanilla) to 74.4 at 10× and 72.6 at 20×, with standard errors near ~1.9. NIAH and LITM remain at 100.0 for all tested ratios in that table. The paper notes that joint compression across chunks could improve accuracy for some offload scenarios but is not required for feasibility, highlighting an engineering trade-off between deployment simplicity in distributed settings and optimal global compression. Latency and TTFT economics A critical system result is the measured compression/decompression latency on an H100 for a non-fused implementation. For Mistral NeMo 12B in bfloat16: BS=8, CTX=8K: compression 379ms, decompression 267ms; vanilla recompute TTFT 3098ms; kvtc decompression TTFT 380ms BS=2, CTX=16K: compression 194ms, decompression 143ms; vanilla recompute TTFT 1780ms; kvtc decompression TTFT 208ms These measurements imply that, when a cache would otherwise be recomputed, decompressing a stored compressed cache can reduce TTFT by ~8×–9× in these scenarios, even without kernel fusion. The decomposition of runtime shows PCA projection and entropy coding as the largest contributors, implying that GPU-optimized kernels and faster GPU-native lossless codecs could reduce overhead further. The fundamental economic conclusion is that, in multi-turn settings with long prefixes, compression-induced overhead is likely dominated by the avoided prefill compute and avoided transfer overhead for uncompressed caches. KEY DEPLOYMENT-SENSITIVE DESIGN CHOICES AND FAILURE MODES Several design choices appear to be “hard requirements” rather than optional optimizations: Sink tokens and sliding window exclusions The paper’s ablations show that compressing early “sink” tokens can catastrophically degrade accuracy at high compression ratios (example: Llama 3.1 8B at 64× collapses on multiple tasks when sink tokens are compressed). Similarly, compressing the most recent tokens hurts performance, motivating a sliding window (default 128 tokens) that remains uncompressed. This introduces a predictable engineering constraint: kvtc is not a uniform compression of the full cache; it is a policy-driven, token-position-dependent codec. Production integration therefore requires correct handling of token positions, attention sinks, and window management, and these policies must be aligned with attention-kernel behavior and model-specific sink dynamics. RoPE handling Removing positional rotations prior to compression is described as important for preserving low-rank structure. In deployment, this implies that the codec must be position-aware and must invert and reapply RoPE correctly. This is an additional source of complexity relative to pure per-token quantization and is sensitive to model variants and RoPE parameterizations. Calibration set representativeness The method’s quality hinges on the PCA basis generalizing from calibration data to production data. The paper demonstrates relative stability with 160K–200K calibration tokens and explores domain shifts (general web text vs math traces vs code). Results suggest that moderate domain mismatch is tolerated at 16×–64×, while extreme compression (e.g., 256× in ablations) becomes materially more sensitive to calibration choice. In production, this implies that operators targeting the “negligible degradation” regime should be able to calibrate with broadly representative corpora, while operators targeting ultra-high compression for specialized workloads should expect tighter coupling between calibration domain and achieved quality. PCA matrix storage overhead and operational footprint A non-trivial hidden cost is the need to store PCA projection matrices per model. The paper reports that, prior to DP trimming, PCA matrices stored at 16-bit can amount to a meaningful fraction of model parameter count (examples reported: ~2.4% for Llama 3.3 70B, ~8.7% for Llama 3.1 8B). This overhead is amortized across all cached sessions for a model but competes with HBM/DRAM budgets in multi-model serving. DP-driven trimming can reduce this overhead at higher compression ratios by removing zero-bit components, but the directionality is not guaranteed at low compression ratios if many components remain active. In distributed inference (pipeline parallel), per-chunk PCA can reduce matrix sizes, but may reduce cross-layer decorrelation benefits if fewer layers are concatenated. SYSTEM-LEVEL IMPLICATIONS FOR GENERATIVE AI INFRASTRUCTURE GPU AND HBM The principal infrastructure implication is that KV cache compression at storage time targets the dominant memory allocator stressor in stateful serving: the accumulation of idle or warm conversation state. For workloads with long reusable prefixes (code assistants, enterprise agents with large system prompts, repeated RAG scaffolds, document chat), the limiting resource frequently becomes HBM reserved for KV caches rather than compute. By compressing stale caches by ~20× (or more), the same HBM budget can retain a materially larger working set of cached prefixes, increasing cache hit rates and reducing recomputation. This effect is multiplicative with cache-aware routing and prefix sharing: more prefixes can remain resident (hot or warm) and can be routed to nodes that already hold them, improving both throughput and tail latency. However, kvtc as described does not reduce the active KV cache footprint during the actual attention computation for a currently decoding sequence, because the model operates on decompressed KV caches during decoding. Therefore, the method does not directly reduce HBM bandwidth consumed by attention kernels during steady-state decode, and does not directly address the “memory traffic per generated token” bottleneck that motivates online KV quantization and eviction strategies. The primary HBM benefit is increased effective capacity for caches between turns and reduced HBM pressure from storing many idle sessions, not reduced per-token decode bandwidth. Compression and decompression themselves consume GPU compute and memory bandwidth. The measured decompression TTFT of ~208ms–380ms in the provided benchmarks indicates that the overhead is real but can be materially smaller than recomputation of long prefixes. In an HBM-constrained serving environment, this overhead can be interpreted as a trade between (a) maintaining more caches warm and paying decompression on reuse versus (b) evicting caches and paying full prefill recomputation. The decision boundary will depend on distribution of inter-turn idle times, probability of reuse, and SLA sensitivity to TTFT. kvtc expands the feasible region where keeping caches is economically rational, especially for long prompts. CPU AND DRAM The method implies a stronger role for CPU DRAM as a warm KV cache tier. A ~20× compression ratio changes the practical scale of “warm state” that can be stored per server. Using the paper’s reported KV cache sizes, a 10K-token 16-bit KV cache for Llama 3.3 70B is ~3.13GiB; compressing by ~20× would reduce this to ~160MiB. At that size, storing hundreds to thousands of warm conversation states in DRAM becomes materially more feasible, increasing cache hit rates and reducing NVMe dependence. This can shift system design from “HBM-only hot caches with aggressive eviction” toward “HBM hot + DRAM warm with long retention,” which is structurally analogous to CPU page cache hierarchies in classical systems design. CPU compute implications depend on where compression is executed. The paper explicitly allows compression on CPU if the cache is already in storage, but the strongest bandwidth savings are achieved when compression happens before moving KV caches off the GPU. If an operator chooses GPU-side compression prior to PCIe/NVLink transfer, CPU compute overhead is modest (orchestrating and DP calibration offline). If an operator instead transfers uncompressed caches to CPU for compression, bandwidth savings are forfeited and CPU memory bandwidth becomes a bottleneck. Therefore, the most economically coherent deployment path is GPU-native compression/decompression with CPU DRAM used as the warm storage reservoir.

TheValueist

16,549 次观看 • 6 个月前