Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Search across a half million git repos at Algolia-level speeds ⚡ I'm never using GitHub code search ever again.

Steven Tey

58,273 subscribers

156,650 views • 3 years ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

We built the best codebase search to give any agent grounded information from GitHub repos. Introducing Sandbox Search. Point it at any repo, and we’ll spin up a secure coding agent in its own sandbox to do research for you. Use inside claude code, openclaw, cursor, and more.

We built the best codebase search to give any agent grounded information from GitHub repos. Introducing Sandbox Search. Point it at any repo, and we’ll spin up a secure coding agent in its own sandbox to do research for you. Use inside claude code, openclaw, cursor, and more.

Arlan

12,412 views • 4 months ago

we built an extremely good code search agent. it’s called tracer, and it can search across the entire github in less than 5 minutes and come back with a clean report. you can try it by running 'bunx nia-wizard'

we built an extremely good code search agent. it’s called tracer, and it can search across the entire github in less than 5 minutes and come back with a clean report. you can try it by running 'bunx nia-wizard'

Arlan

91,438 views • 6 months ago

We're a grassroots campaign. I'm so grateful to the nearly $5 million we've raised from over 200,000 people across the country who believe in our movement to impeach Donald Trump and pass laws to make sure nothing like this can ever happen again. #NY12

We're a grassroots campaign. I'm so grateful to the nearly $5 million we've raised from over 200,000 people across the country who believe in our movement to impeach Donald Trump and pass laws to make sure nothing like this can ever happen again. #NY12

George Conway ⚖️🇺🇸

78,955 views • 2 months ago

MARS ROVER CAPTURES STUNNING NEW FOOTAGE ON SOL 1688 AMID SEARCH FOR SIGNS OF LIFE NASA’s Perseverance rover is still rolling across the Martian surface, snapping high-def footage like a pro using its Right Mastcam-Z camera. This latest video was taken on Sol 1688, as the rover continues its mission to search for ancient microbial life in Jezero Crater. It’s not just about pretty rocks: every frame helps scientists understand Mars’ history, climate, and whether it was ever home to life. So yeah, it's dusty, it’s dry... but it’s also science history in the making. Source: Black Hole

MARS ROVER CAPTURES STUNNING NEW FOOTAGE ON SOL 1688 AMID SEARCH FOR SIGNS OF LIFE NASA’s Perseverance rover is still rolling across the Martian surface, snapping high-def footage like a pro using its Right Mastcam-Z camera. This latest video was taken on Sol 1688, as the rover continues its mission to search for ancient microbial life in Jezero Crater. It’s not just about pretty rocks: every frame helps scientists understand Mars’ history, climate, and whether it was ever home to life. So yeah, it's dusty, it’s dry... but it’s also science history in the making. Source: Black Hole

Mario Nawfal

243,195 views • 9 months ago

Hermes Agent vs OpenClaw using Qwen 35B Local Model We asked agents to scrape GitHub star history for both tools, find what caused the growth spikes, build a live dashboard in the browser. MacBook Pro M5 Max 64Gb OpenClaw: 203k tokens, 12m 01s - wrote a bash script Hermes: 257k tokens, 33m 01s - wrote a SKILL.md OpenClaw hit GitHub API, got truncated responses, paginated through contributors, pulled star-history JSON, found a security incident in OpenClaw's history, fetched SVGs, fixed broken HTML from trimming, rewrote it clean. Hermes parallel tool calls across GitHub API, web search, and browser. Hit Google rate limit, auto-switched to DuckDuckGo. Fetched article contents, mapped viral moments, then built the dashboard. Both shipped a live dashboard with star growth charts and spike annotations

Hermes Agent vs OpenClaw using Qwen 35B Local Model We asked agents to scrape GitHub star history for both tools, find what caused the growth spikes, build a live dashboard in the browser. MacBook Pro M5 Max 64Gb OpenClaw: 203k tokens, 12m 01s - wrote a bash script Hermes: 257k tokens, 33m 01s - wrote a SKILL.md OpenClaw hit GitHub API, got truncated responses, paginated through contributors, pulled star-history JSON, found a security incident in OpenClaw's history, fetched SVGs, fixed broken HTML from trimming, rewrote it clean. Hermes parallel tool calls across GitHub API, web search, and browser. Hit Google rate limit, auto-switched to DuckDuckGo. Fetched article contents, mapped viral moments, then built the dashboard. Both shipped a live dashboard with star growth charts and spike annotations

atomicbot.ai

605,729 views • 3 months ago

I just built a Google Ads Builder in Claude Code that turns one URL into a complete, launch-ready campaign 🤯 Point it at your website and it builds the whole Google Search campaign — keywords, ad groups, every headline, the negative list — structured the way a good paid-search manager would. All inside Claude Code. Perfect for DTC brands and agencies who need Search live but don't have days to build it right. If you're staring at a blank Google Ads account, researching keywords one browser tab at a time, writing 15 headlines per ad group to a 30-character limit, guessing at a negative list while budget quietly leaks on junk clicks... This builds the entire campaign from a single input — your homepage URL: → Reads your site and works out what you sell and who for → Groups keywords into tight, high-intent ad groups → Writes every Responsive Search Ad — 15 headlines, 4 descriptions, to Google's exact limits → Builds a negative-keyword list so you stop paying for junk clicks → Adds sitelinks, callouts, match types, bidding, and a budget split → Exports a Google Ads Editor CSV you import in one click No blank-account paralysis. No keyword rabbit holes. No character-counting 60 headlines by hand. What you get: → A launch-ready campaign from just your URL → Tight ad groups built for Quality Score, not a keyword dump → Every ad written to Google's limits and policy → A one-click CSV import plus a full campaign dashboard Built 100% in Claude Code. No API keys, no Google Ads login. And I'm not sending a playbook this time — I'm giving the whole skill away for free. The actual file. Install it and build your own campaign in 5 minutes. Want the skill? > Like this post > Comment "BUILD" And I'll send it over (must be following so I can DM)

I just built a Google Ads Builder in Claude Code that turns one URL into a complete, launch-ready campaign 🤯 Point it at your website and it builds the whole Google Search campaign — keywords, ad groups, every headline, the negative list — structured the way a good paid-search manager would. All inside Claude Code. Perfect for DTC brands and agencies who need Search live but don't have days to build it right. If you're staring at a blank Google Ads account, researching keywords one browser tab at a time, writing 15 headlines per ad group to a 30-character limit, guessing at a negative list while budget quietly leaks on junk clicks... This builds the entire campaign from a single input — your homepage URL: → Reads your site and works out what you sell and who for → Groups keywords into tight, high-intent ad groups → Writes every Responsive Search Ad — 15 headlines, 4 descriptions, to Google's exact limits → Builds a negative-keyword list so you stop paying for junk clicks → Adds sitelinks, callouts, match types, bidding, and a budget split → Exports a Google Ads Editor CSV you import in one click No blank-account paralysis. No keyword rabbit holes. No character-counting 60 headlines by hand. What you get: → A launch-ready campaign from just your URL → Tight ad groups built for Quality Score, not a keyword dump → Every ad written to Google's limits and policy → A one-click CSV import plus a full campaign dashboard Built 100% in Claude Code. No API keys, no Google Ads login. And I'm not sending a playbook this time — I'm giving the whole skill away for free. The actual file. Install it and build your own campaign in 5 minutes. Want the skill? > Like this post > Comment "BUILD" And I'll send it over (must be following so I can DM)

Mike Futia

60,662 views • 1 month ago

If you are using an if statement inside a for loop, did you know there's a better way of doing this is Swift? I'm talking about the where clause. Using the where clause in a for loop can provide several benefits over using an if statement inside a for loop: Readability: It explicitly states the condition that must be met for an element to be processed, making the purpose of the loop clearer to anyone reading the code. Efficiency: When using a where clause, the filtering of elements is done at the level of the sequence, which can be more efficient than filtering within the loop with an if statement. Brevity: Using a where clause often results in more concise code, as it eliminates the need for an if statement and associated braces. This can make the code more manageable and easier to maintain.

If you are using an if statement inside a for loop, did you know there's a better way of doing this is Swift? I'm talking about the where clause. Using the where clause in a for loop can provide several benefits over using an if statement inside a for loop: Readability: It explicitly states the condition that must be met for an element to be processed, making the purpose of the loop clearer to anyone reading the code. Efficiency: When using a where clause, the filtering of elements is done at the level of the sequence, which can be more efficient than filtering within the loop with an if statement. Brevity: Using a where clause often results in more concise code, as it eliminates the need for an if statement and associated braces. This can make the code more manageable and easier to maintain.

Everton Carneiro

23,416 views • 2 years ago

I still don't understand why everyone is not using this yet. Thanks to it, a year ago I increased my income to 17,000 dollars a month Andrey Karpathy, co-founder of OpenAI, published a simple idea that got 16 million views: stop using AI to write code, use it to build a second brain You point Claude Code to a folder, drop any source in there (an article, transcript, PDF) and Claude reads it, links it, and saves it into a living wiki of everything you know. It compounds like interest: the more you feed it, the smarter it gets Here is the gist: Install Obsidian, create a vault, open it in Claude Code Paste the file with Karpathy's wiki idea and tell Claude to build it Claude creates three folders: raw for sources, wiki for its pages, CLAUDE which runs everything Drop any source into raw and say "ingest this" Ask questions across everything, forever Five minutes to set up, and you never start with an empty chat again The full step-by-step guide is in the article. Save to bookmarks

I still don't understand why everyone is not using this yet. Thanks to it, a year ago I increased my income to 17,000 dollars a month Andrey Karpathy, co-founder of OpenAI, published a simple idea that got 16 million views: stop using AI to write code, use it to build a second brain You point Claude Code to a folder, drop any source in there (an article, transcript, PDF) and Claude reads it, links it, and saves it into a living wiki of everything you know. It compounds like interest: the more you feed it, the smarter it gets Here is the gist: Install Obsidian, create a vault, open it in Claude Code Paste the file with Karpathy's wiki idea and tell Claude to build it Claude creates three folders: raw for sources, wiki for its pages, CLAUDE which runs everything Drop any source into raw and say "ingest this" Ask questions across everything, forever Five minutes to set up, and you never start with an empty chat again The full step-by-step guide is in the article. Save to bookmarks

Bober_smart

2,787,702 views • 6 days ago

Remember the guy who lost 7,500 BTC because he threw away his old hard drive? That’s James Howells, and today those coins are worth over $500 million forever locked behind a seed phrase he can’t recover. We have heard stories like that. panic, loss, phone wipe One mistake, and it’s game over. Now imagine never having to write down a seed phrase again. That’s what Vultisig is building. A seedless, multi-chain vault where your keys are split across trusted devices using MPC/TSS tech meaning no single point of failure, ever. It’s crypto security built for the AI Agent era: → self-custodial automation → multi-chain access → multi-factor protection → zero seed phrases. ever. The next evolution of self-custody is here. And it’s powered by Vultisig.

Remember the guy who lost 7,500 BTC because he threw away his old hard drive? That’s James Howells, and today those coins are worth over $500 million forever locked behind a seed phrase he can’t recover. We have heard stories like that. panic, loss, phone wipe One mistake, and it’s game over. Now imagine never having to write down a seed phrase again. That’s what Vultisig is building. A seedless, multi-chain vault where your keys are split across trusted devices using MPC/TSS tech meaning no single point of failure, ever. It’s crypto security built for the AI Agent era: → self-custodial automation → multi-chain access → multi-factor protection → zero seed phrases. ever. The next evolution of self-custody is here. And it’s powered by Vultisig.

OHJAY ⭕️ || 🇬🇧

24,493 views • 10 months ago

Every serious Claude Code user is using this repo. if you're not, you're leaving 90% of Claude Code's power on the table. It's called claude-code-best-practice - 84 sourced tips, implementation examples for every major feature, workflow comparisons across 8 major repos, and the actual tips from Boris Cherny (creator of Claude Code) compiled in one place. Here's what's actually in it: → 84 tips organized by category -- prompting, planning, CLAUDE.md, agents, commands, skills, hooks, workflows, debugging, utilities, daily habits → best practice + implemented examples for every core concept: subagents, commands, skills, hooks, MCP servers, plugins, settings, memory, checkpointing, CLI flags → workflow comparison table -- Superpowers, BMAD-METHOD, Get Shit Done, OpenSpec, gstack, HumanLayer -- what makes each unique, how many agents/commands/skills each has → orchestration workflow -- Command → Agent → Skill pattern with a live demo → Boris Cherny tips compiled across 3 tweet threads (13 + 10 + 12 tips) and 5 podcast/video appearances → "billion dollar questions" section -- open questions about CLAUDE.md, agents vs commands vs skills, specs -- that nobody has definitively answered yet here's a few of the tips that actually change how you use it: → use subagents with "say use subagents" to throw more compute at a problem -- offload tasks to keep your main context clean → spin up a second Claude to review your plan as a staff engineer before executing → CLAUDE.md should target under 200 lines -- wrap domain-specific rules in ` ` tags so Claude doesn't ignore them as files grow → compress KV context at max 50%, not at the end -- avoid the "agent dumb zone" by doing manual /compact proactively → after a mediocre fix: "knowing everything you know now, scrap this and implement the elegant solution" was #1 trending on GitHub in March 2026. 19.7K GitHub stars. 1.7K forks. MIT license. 100% open source. (link in the comments)

Every serious Claude Code user is using this repo. if you're not, you're leaving 90% of Claude Code's power on the table. It's called claude-code-best-practice - 84 sourced tips, implementation examples for every major feature, workflow comparisons across 8 major repos, and the actual tips from Boris Cherny (creator of Claude Code) compiled in one place. Here's what's actually in it: → 84 tips organized by category -- prompting, planning, CLAUDE.md, agents, commands, skills, hooks, workflows, debugging, utilities, daily habits → best practice + implemented examples for every core concept: subagents, commands, skills, hooks, MCP servers, plugins, settings, memory, checkpointing, CLI flags → workflow comparison table -- Superpowers, BMAD-METHOD, Get Shit Done, OpenSpec, gstack, HumanLayer -- what makes each unique, how many agents/commands/skills each has → orchestration workflow -- Command → Agent → Skill pattern with a live demo → Boris Cherny tips compiled across 3 tweet threads (13 + 10 + 12 tips) and 5 podcast/video appearances → "billion dollar questions" section -- open questions about CLAUDE.md, agents vs commands vs skills, specs -- that nobody has definitively answered yet here's a few of the tips that actually change how you use it: → use subagents with "say use subagents" to throw more compute at a problem -- offload tasks to keep your main context clean → spin up a second Claude to review your plan as a staff engineer before executing → CLAUDE.md should target under 200 lines -- wrap domain-specific rules in ` ` tags so Claude doesn't ignore them as files grow → compress KV context at max 50%, not at the end -- avoid the "agent dumb zone" by doing manual /compact proactively → after a mediocre fix: "knowing everything you know now, scrap this and implement the elegant solution" was #1 trending on GitHub in March 2026. 19.7K GitHub stars. 1.7K forks. MIT license. 100% open source. (link in the comments)

Sukh Sroay

113,759 views • 4 months ago

PLUGGED CLAUDE CODE INTO OBSIDIAN AND MY DEAD VAULT STARTED FIRING LIKE A NEURAL NETWORK 13,000 notes written by hand over 4 years, and search worked fine that was the problem, because search only gives back what you already remember well enough to ask for The graph view never helped either, just a pretty picture nobody actually reads Here is what almost nobody uses: Obsidian keeps your entire vault as plain markdown in a folder on disk, with no database, no export, no API to beg for access So an agent doesn't need permission, it needs a path I gave Claude Code the vault directory and one rule don't answer me, read the links It stopped treating notes as files and started treating them as a net, where backlinks became edges and tags became weights and every note it opened dragged 20 more into context Then it wrote a file I never asked for One note welding two clusters 4 years apart an old client teardown and a half-dead idea buried in a daily note from last winter, a connection I had circled for months and never closed The agent closed it on the first pass, not because it was smarter, but because it had no memory of writing either one That's the part nobody says out loud: a second brain is not a memory problem, it's a distance problem, and you are standing too close to your own notes Everyone builds a second brain, nobody gives it a nervous system

PLUGGED CLAUDE CODE INTO OBSIDIAN AND MY DEAD VAULT STARTED FIRING LIKE A NEURAL NETWORK 13,000 notes written by hand over 4 years, and search worked fine that was the problem, because search only gives back what you already remember well enough to ask for The graph view never helped either, just a pretty picture nobody actually reads Here is what almost nobody uses: Obsidian keeps your entire vault as plain markdown in a folder on disk, with no database, no export, no API to beg for access So an agent doesn't need permission, it needs a path I gave Claude Code the vault directory and one rule don't answer me, read the links It stopped treating notes as files and started treating them as a net, where backlinks became edges and tags became weights and every note it opened dragged 20 more into context Then it wrote a file I never asked for One note welding two clusters 4 years apart an old client teardown and a half-dead idea buried in a daily note from last winter, a connection I had circled for months and never closed The agent closed it on the first pass, not because it was smarter, but because it had no memory of writing either one That's the part nobody says out loud: a second brain is not a memory problem, it's a distance problem, and you are standing too close to your own notes Everyone builds a second brain, nobody gives it a nervous system

Spike 1%

12,561 views • 10 days ago

My project has 39,205 lines of code, and Cursor can't answer questions about it. Cursor's context seems to be capped at around 10,000 tokens. Unfortunately, this is not enough for any decent-sized project. If you have a large codebase, check out Augment Code. This thing is faaaast! I'm currently using their Visual Studio Code plugin, but you can also use them on JetBrains, Neovim, and even Vim. (I'm a Neovim fan, but Copilot's implementation for Neovim is nowhere as good as Augment Code.) Augment Code was gracious enough to sponsor this post. After you install their extension and run it for the first time, it will index your entire codebase. This is why it can answer questions as fast as it does, regardless of the size of your codebase. Augment Code supports chat and completions like every other AI coding assistant, but its killer feature is "Next Edit." When you make a change, two things happen: 1. The model analyzes the change to determine the ripple effects across your *entire* codebase. 2. The model suggests everything you need to update to ensure everything works correctly. This is pretty wild!

My project has 39,205 lines of code, and Cursor can't answer questions about it. Cursor's context seems to be capped at around 10,000 tokens. Unfortunately, this is not enough for any decent-sized project. If you have a large codebase, check out Augment Code. This thing is faaaast! I'm currently using their Visual Studio Code plugin, but you can also use them on JetBrains, Neovim, and even Vim. (I'm a Neovim fan, but Copilot's implementation for Neovim is nowhere as good as Augment Code.) Augment Code was gracious enough to sponsor this post. After you install their extension and run it for the first time, it will index your entire codebase. This is why it can answer questions as fast as it does, regardless of the size of your codebase. Augment Code supports chat and completions like every other AI coding assistant, but its killer feature is "Next Edit." When you make a change, two things happen: 1. The model analyzes the change to determine the ripple effects across your entire codebase. 2. The model suggests everything you need to update to ensure everything works correctly. This is pretty wild!

Santiago

247,833 views • 1 year ago

Obsidian became a neural network with 184,320 links and now runs as the most powerful second brain ever built for $0. 12,480 notes in 1 vault, 0 folders, 0 tabs, and he hasn't opened Notion in 7 months. Every night at 2 AM Claude Code walks the vault, reads all 12,480 notes, embeds them, and hunts for the pairs that should be linked but never were. Last night it found 61 and he accepted all 61. That's the move nobody makes by hand, because 12,480 notes means 77 million possible pairs, and a human links maybe 4 a day while the agent links 61 while he sleeps. Capture takes 2 seconds, a voice memo lands as a note in 1.4, a clip becomes a linked node on its own friction is what kills a second brain so he deleted the friction. Recall latency sits at 38 milliseconds, which means the graph answers before he finishes doubting himself. 36 clusters formed without him, decay score holds at 0.03, and 97% of everything he ever wrote is still reachable, while most people's notes rot past 90% in a year and turn into a graveyard with good search. Yesterday it surfaced 9 forgotten notes, one of them 412 days old, and that one became a product he shipped 48 hours later. Plain .md files on a laptop, no cloud, no subscription, no company that can delete his brain on a Tuesday. The graph doesn't store what he knows. It stores what he'd never have connected alone.

Obsidian became a neural network with 184,320 links and now runs as the most powerful second brain ever built for $0. 12,480 notes in 1 vault, 0 folders, 0 tabs, and he hasn't opened Notion in 7 months. Every night at 2 AM Claude Code walks the vault, reads all 12,480 notes, embeds them, and hunts for the pairs that should be linked but never were. Last night it found 61 and he accepted all 61. That's the move nobody makes by hand, because 12,480 notes means 77 million possible pairs, and a human links maybe 4 a day while the agent links 61 while he sleeps. Capture takes 2 seconds, a voice memo lands as a note in 1.4, a clip becomes a linked node on its own friction is what kills a second brain so he deleted the friction. Recall latency sits at 38 milliseconds, which means the graph answers before he finishes doubting himself. 36 clusters formed without him, decay score holds at 0.03, and 97% of everything he ever wrote is still reachable, while most people's notes rot past 90% in a year and turn into a graveyard with good search. Yesterday it surfaced 9 forgotten notes, one of them 412 days old, and that one became a product he shipped 48 hours later. Plain .md files on a laptop, no cloud, no subscription, no company that can delete his brain on a Tuesday. The graph doesn't store what he knows. It stores what he'd never have connected alone.

Spike 1%

11,546 views • 8 days ago

K-Means is simple. Making it fast on GPU isn't. Flash-KMeans is an IO-aware implementation of exact k-means that rethinks the algorithm around modern GPU bottlenecks. By attacking the memory bottlenecks directly, Flash-KMeans achieves: - 30x speedup over cuML - 200x speedup over FAISS Using the same exact algorithm, just engineered for today’s hardware. At the million-scale, Flash-KMeans can complete a k-means iteration in milliseconds. Here's why this matters today: K-means has always been an offline primitive. Something you run once to preprocess data and move on. These speedups change that. ↳ Vector databases like FAISS use k-means to build search indices. Faster k-means means you can re-index dynamically as data changes, not batch it overnight. ↳ LLM quantization methods need k-means to find optimal weight codebooks, per layer, repeatedly. What takes hours could now take minutes. ↳ MoE models need fast token routing at inference time. Millisecond k-means makes it viable to run this inside the inference loop, not just in preprocessing. The 200x over FAISS is the number to internalize. FAISS is the industry standard. Most production vector search systems sit on top of it. Link to the paper and code in next tweet!

K-Means is simple. Making it fast on GPU isn't. Flash-KMeans is an IO-aware implementation of exact k-means that rethinks the algorithm around modern GPU bottlenecks. By attacking the memory bottlenecks directly, Flash-KMeans achieves: - 30x speedup over cuML - 200x speedup over FAISS Using the same exact algorithm, just engineered for today’s hardware. At the million-scale, Flash-KMeans can complete a k-means iteration in milliseconds. Here's why this matters today: K-means has always been an offline primitive. Something you run once to preprocess data and move on. These speedups change that. ↳ Vector databases like FAISS use k-means to build search indices. Faster k-means means you can re-index dynamically as data changes, not batch it overnight. ↳ LLM quantization methods need k-means to find optimal weight codebooks, per layer, repeatedly. What takes hours could now take minutes. ↳ MoE models need fast token routing at inference time. Millisecond k-means makes it viable to run this inside the inference loop, not just in preprocessing. The 200x over FAISS is the number to internalize. FAISS is the industry standard. Most production vector search systems sit on top of it. Link to the paper and code in next tweet!

Daily Dose of Data Science

23,748 views • 4 months ago

K-Means is simple. Making it fast on GPU isn't. Flash-KMeans is an IO-aware implementation of exact k-means that rethinks the algorithm around modern GPU bottlenecks. By attacking the memory bottlenecks directly, Flash-KMeans achieves: - 30x speedup over cuML - 200x speedup over FAISS Using the same exact algorithm, just engineered for today’s hardware. At the million-scale, Flash-KMeans can complete a k-means iteration in milliseconds. Here's why this matters today: K-means has always been an offline primitive. Something you run once to preprocess data and move on. These speedups change that. ↳ Vector databases like FAISS use k-means to build search indices. Faster k-means means you can re-index dynamically as data changes, not batch it overnight. ↳ LLM quantization methods need k-means to find optimal weight codebooks, per layer, repeatedly. What takes hours could now take minutes. ↳ MoE models need fast token routing at inference time. Millisecond k-means makes it viable to run this inside the inference loop, not just in preprocessing. The 200x over FAISS is the number to internalize. FAISS is the industry standard. Most production vector search systems sit on top of it. Link to the paper and code in next tweet!

K-Means is simple. Making it fast on GPU isn't. Flash-KMeans is an IO-aware implementation of exact k-means that rethinks the algorithm around modern GPU bottlenecks. By attacking the memory bottlenecks directly, Flash-KMeans achieves: - 30x speedup over cuML - 200x speedup over FAISS Using the same exact algorithm, just engineered for today’s hardware. At the million-scale, Flash-KMeans can complete a k-means iteration in milliseconds. Here's why this matters today: K-means has always been an offline primitive. Something you run once to preprocess data and move on. These speedups change that. ↳ Vector databases like FAISS use k-means to build search indices. Faster k-means means you can re-index dynamically as data changes, not batch it overnight. ↳ LLM quantization methods need k-means to find optimal weight codebooks, per layer, repeatedly. What takes hours could now take minutes. ↳ MoE models need fast token routing at inference time. Millisecond k-means makes it viable to run this inside the inference loop, not just in preprocessing. The 200x over FAISS is the number to internalize. FAISS is the industry standard. Most production vector search systems sit on top of it. Link to the paper and code in next tweet!

Akshay 🚀

36,317 views • 5 months ago

most AI chatbots break when you ask a question that requires info from multiple sources for example try asking: “which client contracts are finishing up this month?” you’ll get a half-answer — or none at all why? because traditional chatbots only look at small snippets of your docs - they don’t understand how things connect across clients, services, timelines that’s where knowledge graphs come in they let you turn messy contracts into a web of relationships — like: "Client → Contract Type → Service Provided → End Date" so instead of guessing from a few chunks of text, your chatbot can search across all your clients and contracts to give accurate answers I made a full walkthrough on how I built this: – how to organize your contracts so an AI can actually use them – how to define what matters (like who signed what, and when) – how to get the AI to figure out what info it needs and where to find it – and how to feed that back into your chatbot so it gives accurate answers reply “graph” and I’ll DM it to you (must be following)

most AI chatbots break when you ask a question that requires info from multiple sources for example try asking: “which client contracts are finishing up this month?” you’ll get a half-answer — or none at all why? because traditional chatbots only look at small snippets of your docs - they don’t understand how things connect across clients, services, timelines that’s where knowledge graphs come in they let you turn messy contracts into a web of relationships — like: "Client → Contract Type → Service Provided → End Date" so instead of guessing from a few chunks of text, your chatbot can search across all your clients and contracts to give accurate answers I made a full walkthrough on how I built this: – how to organize your contracts so an AI can actually use them – how to define what matters (like who signed what, and when) – how to get the AI to figure out what info it needs and where to find it – and how to feed that back into your chatbot so it gives accurate answers reply “graph” and I’ll DM it to you (must be following)

Tyler

24,994 views • 1 year ago

We release Diamond Maps💎 unlocking accurate and efficient guidance for diffusion models. Our experiments show that our methods scale incredibly well. Excited to see what people will build with this! Accurate guidance has been a notoriously hard problem, but in this work, we’re bringing TWO (!) solutions to the table. The recipe for success: 1️⃣ Speed: Use distilled models (flow maps, mean flows, consistency models). 2️⃣ Exploration: Inject stochasticity to properly explore your search space. Because this fundamentally improves anything using flow matching and diffusion, we see a lot of potential for applications across audio, robotics, molecules, and beyond. Paper: Code: Huge thanks to an amazing team: Douglas Chen, Luca Eyring, Ishin Shah, Giri Anantharaman, Yutong (Kelly) He, Zeynep Akata, Tommi Jaakkola, Nicholas Boffi, and Max Simchowitz. It was awesome bringing this to life together!

We release Diamond Maps💎 unlocking accurate and efficient guidance for diffusion models. Our experiments show that our methods scale incredibly well. Excited to see what people will build with this! Accurate guidance has been a notoriously hard problem, but in this work, we’re bringing TWO (!) solutions to the table. The recipe for success: 1️⃣ Speed: Use distilled models (flow maps, mean flows, consistency models). 2️⃣ Exploration: Inject stochasticity to properly explore your search space. Because this fundamentally improves anything using flow matching and diffusion, we see a lot of potential for applications across audio, robotics, molecules, and beyond. Paper: Code: Huge thanks to an amazing team: Douglas Chen, Luca Eyring, Ishin Shah, Giri Anantharaman, Yutong (Kelly) He, Zeynep Akata, Tommi Jaakkola, Nicholas Boffi, and Max Simchowitz. It was awesome bringing this to life together!

Peter Holderrieth

60,526 views • 4 months ago