Video yükleniyor...

Video Yüklenemedi

Bu video yüklenirken bir sorun oluştu. Bu geçici bir ağ sorunundan kaynaklanıyor olabilir veya video kullanılamıyor olabilir.

Ana Sayfaya Dön

‼️ Dirty Frag: A Universal Linux Local Privilege Escalation via Page-Cache Write Primitives GitHub: Patches: CVE-2026-43284: A page-cache write flaw in the Linux kernel's xfrm-ESP (IPsec) subsystem that lets a local user corrupt read-only file pages via in-place decryption on shared skb fragments CVE-2026-43500: A sibling page-cache write flaw... show more

Dark Web Informer

212,359 subscribers

18,534 görüntüleme • 4 ay önce •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Yorum

Yorum bulunmuyor

Orijinal gönderinin yorumları burada görünecek

Benzer Videolar

I make unlimited landing pages with Claude Cowork for $0 😱 here's the system that turns one URL into a landing page for every possible angle: step 1: scrape your brand DNA → Firecrawl scrapes your site. trust signals, colors, fonts, the works → I pointed it at virlo and it pulled every proof element & visual system in 90 seconds → nothing hallucinated. everything traced back to your actual site step 2: figure out WHY your product works → not the features but the reason someone should care. → maps which claims you can back up vs ones you can't step 3: lock your brand voice → pulls your exact phrases, tone, language patterns → ran it on a client last week. they thought they wrote the page. they didn't. step 4: write the copy → hard ban list kills AI slop on sight ("unlock", "seamless", "revolutionize" = dead) → then cuts 20% of whatever it wrote. if a line doesn't earn its spot, it's gone. step 5: generate on-brand visuals → Bloom creates images that ACTUALLY match your brand (s/o Ray for hooking up FREE bloom credits, ill send you a link) → no stock photo energy. no purple AI gradients. no "two businesspeople shaking hands." step 6: build the page → single-file HTML. responsive. ready to roll. → routes layout by type. product, SaaS, lead gen, regulated all look different → one product. six angles. six pages. each one ships. step 7: QA gate → scores every page on proof, trust, copy, visuals, anti-slop → shippable, draft, or blocked. nothing goes live without passing. input: your URL output: unlimited landing pages w/ copy & brand images Steal Ads automates ads. this builds landers for each. money ad → money page. agencies charge $5-10K per landing page. this builds unlimited pages for $0. I packaged the entire system as the Landing Page Factory. 7 Claude skills: - site-extract (brand DNA via Firecrawl) - page-strategy (mechanism mapping + real claims) - brand-profile (your voice + your branding) - page-copy (conversion copy w/ slop ban) - page-visuals (on-brand images w/ Bloom) - page-build (multiple variations & layouts) - page-qa (shippability gate so it doesn't suck) also works with OpenClaw🦞, hermes (Nous Research) or any agent framework. giving it away free. comment PAGES + like + follow (must follow so i can DM)

Matthew Berman

135,872 görüntüleme • 5 ay önce

ANTHROPIC LEAKED A FILE WHERE 4 DEPARTMENTS ASSEMBLE IN 60 MINUTES AND BUILD A BUSINESS THAT MAKES $100K A MONTH the same work done by people is $30,000 a month - here it's 4 departments and $2.65 a day. the first 12 minutes you write 3 lines all 4 departments will read all week: who buys, what we sell, at what price. development is the largest line on the bill - $1.30, three cheap lanes in parallel and one expensive verdict at the end. marketing is high volume with a low stake per item - so it generates 3 variants and you pick one. one variant is a guess, three variants are a choice, and they cost the same. research runs the other way: the cheap model filters hundreds of sources, the expensive one reads only what survived. paying the top rate for a page you'll discard anyway is the commonest overspend there is. and there's exactly one department where saving on the model makes no sense - unit economics: one wrong number costs you a year. route by the cost of being wrong, not the price per token - Opus on the content queue is 5x overpaid, Haiku on the economics is a year of rework. finance is the only department that writes back into the root - next week marketing runs on rules the economics wrote. at minute 60 you have a page, a payment link, a week of content and an economics model - revenue isn't on that list. save this and paste it into Claude Code - the hour builds the company, the market decides the number ↓

ANTHROPIC LEAKED A FILE WHERE 4 DEPARTMENTS ASSEMBLE IN 60 MINUTES AND BUILD A BUSINESS THAT MAKES $100K A MONTH the same work done by people is $30,000 a month - here it's 4 departments and $2.65 a day. the first 12 minutes you write 3 lines all 4 departments will read all week: who buys, what we sell, at what price. development is the largest line on the bill - $1.30, three cheap lanes in parallel and one expensive verdict at the end. marketing is high volume with a low stake per item - so it generates 3 variants and you pick one. one variant is a guess, three variants are a choice, and they cost the same. research runs the other way: the cheap model filters hundreds of sources, the expensive one reads only what survived. paying the top rate for a page you'll discard anyway is the commonest overspend there is. and there's exactly one department where saving on the model makes no sense - unit economics: one wrong number costs you a year. route by the cost of being wrong, not the price per token - Opus on the content queue is 5x overpaid, Haiku on the economics is a year of rework. finance is the only department that writes back into the root - next week marketing runs on rules the economics wrote. at minute 60 you have a page, a payment link, a week of content and an economics model - revenue isn't on that list. save this and paste it into Claude Code - the hour builds the company, the market decides the number ↓

Sprytix

19,325 görüntüleme • 27 gün önce

I just built a plugin with Claude Fable 5 that turns Claude Code into a $5,000/mo SEO consultant 🤯 9 skills, one plugin: it connects straight to your Search Console + GA4 data, finds the wins, ships the fixes, and renders a live SEO dashboard that looks like a $200/mo SaaS product. All inside Claude Code. Perfect for DTC brands and agencies sitting on months of Search Console data nobody has time to read. Right now, you probably can't answer: Which keywords are sitting on page 2, one title tag away from page 1, Which pages are bleeding traffic to redirect chains and broken canonicals, Which blog posts rank for commercial terms but never link to a product page. This plugin answers all of it from your live data, then ships the fixes: → Finds your page-2 keywords and ships the fix: new title, headings, content, paste-ready → Clusters every query into a hub-and-spoke content map with the gaps flagged → Drafts posts from your actual search data, not guesses → Writes dev tickets for redirect chains and slow pages, ranked by traffic at risk → Builds the internal links between your blog and your money pages → Flags toxic backlinks and ranks outreach targets → Drops a Monday report with 3 priorities before the client even asks → Renders it all as a one-file HTML dashboard with a 0-100 SEO health score No dashboard staring. No CSV archaeology. No $5K/mo retainer for a PDF. What you get: → Page-2 keywords moved to page 1 → A content calendar that fills itself from data → Dev tickets that write themselves → A live SEO dashboard on command Built 100% in Claude Code with Claude Fable 5. I put the entire build into a step-by-step Playbook: all 8 workflow prompts (including the dashboard), how to turn them into a plugin, and the full Google setup (Including the 2 landmines Google doesn't tell you about). Want access for free? > Like this post > Comment "SEO" And I'll send it over (must be following so I can DM)

I just built a plugin with Claude Fable 5 that turns Claude Code into a $5,000/mo SEO consultant 🤯 9 skills, one plugin: it connects straight to your Search Console + GA4 data, finds the wins, ships the fixes, and renders a live SEO dashboard that looks like a $200/mo SaaS product. All inside Claude Code. Perfect for DTC brands and agencies sitting on months of Search Console data nobody has time to read. Right now, you probably can't answer: Which keywords are sitting on page 2, one title tag away from page 1, Which pages are bleeding traffic to redirect chains and broken canonicals, Which blog posts rank for commercial terms but never link to a product page. This plugin answers all of it from your live data, then ships the fixes: → Finds your page-2 keywords and ships the fix: new title, headings, content, paste-ready → Clusters every query into a hub-and-spoke content map with the gaps flagged → Drafts posts from your actual search data, not guesses → Writes dev tickets for redirect chains and slow pages, ranked by traffic at risk → Builds the internal links between your blog and your money pages → Flags toxic backlinks and ranks outreach targets → Drops a Monday report with 3 priorities before the client even asks → Renders it all as a one-file HTML dashboard with a 0-100 SEO health score No dashboard staring. No CSV archaeology. No $5K/mo retainer for a PDF. What you get: → Page-2 keywords moved to page 1 → A content calendar that fills itself from data → Dev tickets that write themselves → A live SEO dashboard on command Built 100% in Claude Code with Claude Fable 5. I put the entire build into a step-by-step Playbook: all 8 workflow prompts (including the dashboard), how to turn them into a plugin, and the full Google setup (Including the 2 landmines Google doesn't tell you about). Want access for free? > Like this post > Comment "SEO" And I'll send it over (must be following so I can DM)

Mike Futia

80,535 görüntüleme • 3 ay önce

Harry Dry is the best copywriter I know. He's built a 130,000-person newsletter teaching people how to do it, and by the end of this interview, you'll be at least a Green Belt in copywriting. Some of his rules for writing: 1) A great sentence is a good sentence made shorter. 2) Writing great copy begins with having something to say in the first place. 3) Copy is like food. How it looks matters. 4) Since the look of copy matters so much, don't write copy in Google Docs. Write it in Figma (so you can write and design at the same time). 5) Kaplan's Law of Words: Any word that isn't working for you is working against you. 6) You know a paragraph is ready to ship when there's nothing left to remove. It's like a Jenga tower. The entire thing should collapse if you remove something. 7) Make a promise in the title so the reader knows exactly what they're going to get if they click. Then, deliver on the promise. 8) The three laws of copywriting: (1) Make it concrete, (2) make it visual, and (3) make it falsifiable. 9) Make it concrete: Don't be abstract. For an example, say you're writing about habits. Don't talk about "productive routines." That's abstract. Write about "waking up at 6am to write" instead. It's concrete — and much more vibrant. 10) Make it visual: People see in pictures. This is why instead of memorizing card numbers directly, world memory champions memorize cards by turning them into pictures and then back to cards. 11) Make it falsifiable: When you write a sentence that's true or false, you put your head on the chopping block, which makes people sit up in their seat. 12) When has a falsifiable statement resonated? Galileo got sentenced to a decade of house arrest for saying that the earth spins around the sun. That's a falsifiable sentence. But nobody would've done anything if he'd said that the earth has a harmonious connection with a celestial object. 13) Write with the delete key. Using fewer words lets you be more impactful with the words you keep. 14) The job of a sales page is to make a bold claim at the top. Then spend the rest of the page backing up what you've said... with a ridiculous amount of proof. 15) If your competitor could've written the sentence, cut it. 16) Good copy is differentiated. Here's an example: Elon Musk shouldn't write "The Cybertruck is the world's best truck." Ford or Dodge can write that sentence. But only Elon can write: "The Cybertruck is tougher than an F-150 and faster than a Porsche." 17) Some days, the writing comes easily. Some days, it takes sweat. The reader doesn't care if you wrote for two minutes, two hours, or two days. The ink looks the same. 18) Great copy reads like your customer wrote it. Talk to them. That's just an introduction to the copywriting philosophy of Harry Dry. I've shared the full interview below. I recommend you watch this one because we pull from so many visual references and do a lot of screen sharing. If you'd rather watch on YouTube, I've shared the link in the reply tweets.

Harry Dry is the best copywriter I know. He's built a 130,000-person newsletter teaching people how to do it, and by the end of this interview, you'll be at least a Green Belt in copywriting. Some of his rules for writing: 1) A great sentence is a good sentence made shorter. 2) Writing great copy begins with having something to say in the first place. 3) Copy is like food. How it looks matters. 4) Since the look of copy matters so much, don't write copy in Google Docs. Write it in Figma (so you can write and design at the same time). 5) Kaplan's Law of Words: Any word that isn't working for you is working against you. 6) You know a paragraph is ready to ship when there's nothing left to remove. It's like a Jenga tower. The entire thing should collapse if you remove something. 7) Make a promise in the title so the reader knows exactly what they're going to get if they click. Then, deliver on the promise. 8) The three laws of copywriting: (1) Make it concrete, (2) make it visual, and (3) make it falsifiable. 9) Make it concrete: Don't be abstract. For an example, say you're writing about habits. Don't talk about "productive routines." That's abstract. Write about "waking up at 6am to write" instead. It's concrete — and much more vibrant. 10) Make it visual: People see in pictures. This is why instead of memorizing card numbers directly, world memory champions memorize cards by turning them into pictures and then back to cards. 11) Make it falsifiable: When you write a sentence that's true or false, you put your head on the chopping block, which makes people sit up in their seat. 12) When has a falsifiable statement resonated? Galileo got sentenced to a decade of house arrest for saying that the earth spins around the sun. That's a falsifiable sentence. But nobody would've done anything if he'd said that the earth has a harmonious connection with a celestial object. 13) Write with the delete key. Using fewer words lets you be more impactful with the words you keep. 14) The job of a sales page is to make a bold claim at the top. Then spend the rest of the page backing up what you've said... with a ridiculous amount of proof. 15) If your competitor could've written the sentence, cut it. 16) Good copy is differentiated. Here's an example: Elon Musk shouldn't write "The Cybertruck is the world's best truck." Ford or Dodge can write that sentence. But only Elon can write: "The Cybertruck is tougher than an F-150 and faster than a Porsche." 17) Some days, the writing comes easily. Some days, it takes sweat. The reader doesn't care if you wrote for two minutes, two hours, or two days. The ink looks the same. 18) Great copy reads like your customer wrote it. Talk to them. That's just an introduction to the copywriting philosophy of Harry Dry. I've shared the full interview below. I recommend you watch this one because we pull from so many visual references and do a lot of screen sharing. If you'd rather watch on YouTube, I've shared the link in the reply tweets.

David Perell

725,601 görüntüleme • 2 yıl önce

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

Avi Chawla

271,839 görüntüleme • 2 ay önce

Meet WebBrain: An Open-Source, Local-First AI Browser Agent That Reads Pages and Automates Tasks in Chrome and Firefox WebBrain lives inside your browser and can run entirely on your own local model — no cloud, no account, no data leaving your machine. Most "AI browser agents" are a chat box that pastes your page into someone else's server. That's not an agent that lives where you browse — and WebBrain draws a very clear line between the two. It's an open-source (MIT), local-first browser agent for Chrome and Firefox. It runs inside your existing authenticated session, on a model you pick — so with llama.cpp or Ollama, nothing leaves your machine. Here's what's actually interesting: → Two modes, cleanly separated. Ask reads the page (read-only, content scripts). Act clicks and types through the Chrome DevTools Protocol (chrome.debugger) — trusted input events that modern sites honor, reaching cross-origin iframes and shadow DOM. → UI-first by design. For anything that submits, sends, or buys, it drives the visible UI and refuses to hit REST/GraphQL endpoints directly. It starts read-only and asks before consequential actions. → Bring any model. llama.cpp, Ollama, LM Studio, vLLM — or OpenAI, Claude, Gemini, DeepSeek, Groq, OpenRouter. Recommended local: Qwen 3.6 35B (Qwen3.6-35B-A3B), which beat Gemma 4 on the project's screenshot benchmark. → Tuned for cost and privacy. Token-conscious screenshots, oldest-first context trimming, a dedicated vision model, 40+ tools (~20 in Compact mode). No telemetry. No accounts. Full analysis: GitHub Repo: Chrome Extension: Firefox Add-on: Portal:

Meet WebBrain: An Open-Source, Local-First AI Browser Agent That Reads Pages and Automates Tasks in Chrome and Firefox WebBrain lives inside your browser and can run entirely on your own local model — no cloud, no account, no data leaving your machine. Most "AI browser agents" are a chat box that pastes your page into someone else's server. That's not an agent that lives where you browse — and WebBrain draws a very clear line between the two. It's an open-source (MIT), local-first browser agent for Chrome and Firefox. It runs inside your existing authenticated session, on a model you pick — so with llama.cpp or Ollama, nothing leaves your machine. Here's what's actually interesting: → Two modes, cleanly separated. Ask reads the page (read-only, content scripts). Act clicks and types through the Chrome DevTools Protocol (chrome.debugger) — trusted input events that modern sites honor, reaching cross-origin iframes and shadow DOM. → UI-first by design. For anything that submits, sends, or buys, it drives the visible UI and refuses to hit REST/GraphQL endpoints directly. It starts read-only and asks before consequential actions. → Bring any model. llama.cpp, Ollama, LM Studio, vLLM — or OpenAI, Claude, Gemini, DeepSeek, Groq, OpenRouter. Recommended local: Qwen 3.6 35B (Qwen3.6-35B-A3B), which beat Gemma 4 on the project's screenshot benchmark. → Tuned for cost and privacy. Token-conscious screenshots, oldest-first context trimming, a dedicated vision model, 40+ tools (~20 in Compact mode). No telemetry. No accounts. Full analysis: GitHub Repo: Chrome Extension: Firefox Add-on: Portal:

Marktechpost AI

203,011 görüntüleme • 2 ay önce

Conor Neill: "If you can't write it clearly, the thinking was weak, not the writing" "To believe that something that feels clear in your head is thinking that's a very dangerous thing. When you try to put it down on a page, when you try to lay out your ideas in a structured order that someone else can digest, and you realize that you can't, I suggest the thinking was weak, not the writing." Neill explains his philosophy: "Writing is thinking. The process of taking a notepad, capturing thoughts, laying out the things that I'm thinking about, that is thinking. Sitting and staring out a window, maybe with a cigarette, whatever it is that you think is philosophizing that is not structured thinking. It's only when you're writing down and structuring, getting order into your thoughts on a page so that another person is able to get into the context, the perspective, the different things that you are pulling in to have your worldview." He shares a simple technique: "No matter what you are writing, whether it's an email, a Word document, when you've got a blank sheet of paper, start with the word 'This.' T-H-I-S. Starting with the word 'This' forces you to explain what the document is. It forces you to articulate to the reader what it is that they are holding. It forces you to describe why this document exists, what the objective is. And if you begin with the objective, it helps the reader, and it helps you articulate clearly why you are taking the time to write." Neill shares the most-read post on his blog: "The one post that has got far more views than any other is a post I wrote called 'Why Amazon Banned PowerPoint.' In Amazon, if a presenter wishes to ask people to agree to a budget, to agree to give them resources, they don't use PowerPoint. They write a six-page Word document that states why they are asking for the money and the resources." He explains Jeff Bezos's reasoning: "PowerPoint is easy for the presenter, but it's hard for the people who listen. Writing a six-page essay is hard for the presenter, but it's a lot easier for the people that get to read the document." And there's a second part to the Amazon method: "In the management meeting, the first 20 minutes is reading time. If you have gone to the effort to write six pages explaining your proposal, you deserve to see your work read. You deserve to sit there and see people reading through your work. People will not read before the meeting. The only way you get people to fully digest the six pages is by holding them there for 20 minutes, reading through, noting down their questions. No debate, no discussion until everyone in the room has read all six pages, has taken in the context, has time to think about what they would like to question. After 20 minutes of silent reading, they can have a discussion but an informed discussion." Neill shares a second insight about writing: "Divide writing from editing. Writing is producing words. Editing is improving words. These two processes — you cannot run at the same time." He explains his approach: "Most writers just vomit out a bad first draft. I personally have learned to produce 500 words in one straight blast. If something's wrong, if I need to check a fact, if I want to go back and fix something, I don't. I go 500 words of just getting it out onto the page. When I've got 500 words, then I'll stop and begin the process of editing." Neill shares what great writers understand: "All great writing is rewriting. It's editing. It's the crafting of taking a bad, crappy first draft and slowly iterating it, improving it 1% each time through. But if you haven't got that first draft, there's nothing to improve." He explains how separating these processes changed everything: "Learning to separate these two was one of the most powerful things to get rid of writer's block, to get rid of getting stuck, to get rid of procrastination. My mission when I sit down to write is: decide, am I writing or editing? If it's writing, get 500 bad words down on the page in the next 20 minutes. If it's editing, take the time to go through, improve sentences, change the order, change the structure. But these are two separate processes." Neill reveals the truth about good writing: "Some of my best articles started out as a bad blog post. Then I rewrote it as an article to give out to students. Then I rewrote it to share on another blog. Then I rewrote it to provide to a magazine. It's the sixth, seventh, eighth, ninth time of rewriting where it starts to be something that other people read and say, 'Wow, you're quite good at writing.' And the answer is I'm not good at writing. I vomit out a bad first draft and then go through this iterative process. One time, two times, three times, four times through slowly improving. But if you have no first draft, there's nothing to improve."

Conor Neill: "If you can't write it clearly, the thinking was weak, not the writing" "To believe that something that feels clear in your head is thinking that's a very dangerous thing. When you try to put it down on a page, when you try to lay out your ideas in a structured order that someone else can digest, and you realize that you can't, I suggest the thinking was weak, not the writing." Neill explains his philosophy: "Writing is thinking. The process of taking a notepad, capturing thoughts, laying out the things that I'm thinking about, that is thinking. Sitting and staring out a window, maybe with a cigarette, whatever it is that you think is philosophizing that is not structured thinking. It's only when you're writing down and structuring, getting order into your thoughts on a page so that another person is able to get into the context, the perspective, the different things that you are pulling in to have your worldview." He shares a simple technique: "No matter what you are writing, whether it's an email, a Word document, when you've got a blank sheet of paper, start with the word 'This.' T-H-I-S. Starting with the word 'This' forces you to explain what the document is. It forces you to articulate to the reader what it is that they are holding. It forces you to describe why this document exists, what the objective is. And if you begin with the objective, it helps the reader, and it helps you articulate clearly why you are taking the time to write." Neill shares the most-read post on his blog: "The one post that has got far more views than any other is a post I wrote called 'Why Amazon Banned PowerPoint.' In Amazon, if a presenter wishes to ask people to agree to a budget, to agree to give them resources, they don't use PowerPoint. They write a six-page Word document that states why they are asking for the money and the resources." He explains Jeff Bezos's reasoning: "PowerPoint is easy for the presenter, but it's hard for the people who listen. Writing a six-page essay is hard for the presenter, but it's a lot easier for the people that get to read the document." And there's a second part to the Amazon method: "In the management meeting, the first 20 minutes is reading time. If you have gone to the effort to write six pages explaining your proposal, you deserve to see your work read. You deserve to sit there and see people reading through your work. People will not read before the meeting. The only way you get people to fully digest the six pages is by holding them there for 20 minutes, reading through, noting down their questions. No debate, no discussion until everyone in the room has read all six pages, has taken in the context, has time to think about what they would like to question. After 20 minutes of silent reading, they can have a discussion but an informed discussion." Neill shares a second insight about writing: "Divide writing from editing. Writing is producing words. Editing is improving words. These two processes — you cannot run at the same time." He explains his approach: "Most writers just vomit out a bad first draft. I personally have learned to produce 500 words in one straight blast. If something's wrong, if I need to check a fact, if I want to go back and fix something, I don't. I go 500 words of just getting it out onto the page. When I've got 500 words, then I'll stop and begin the process of editing." Neill shares what great writers understand: "All great writing is rewriting. It's editing. It's the crafting of taking a bad, crappy first draft and slowly iterating it, improving it 1% each time through. But if you haven't got that first draft, there's nothing to improve." He explains how separating these processes changed everything: "Learning to separate these two was one of the most powerful things to get rid of writer's block, to get rid of getting stuck, to get rid of procrastination. My mission when I sit down to write is: decide, am I writing or editing? If it's writing, get 500 bad words down on the page in the next 20 minutes. If it's editing, take the time to go through, improve sentences, change the order, change the structure. But these are two separate processes." Neill reveals the truth about good writing: "Some of my best articles started out as a bad blog post. Then I rewrote it as an article to give out to students. Then I rewrote it to share on another blog. Then I rewrote it to provide to a magazine. It's the sixth, seventh, eighth, ninth time of rewriting where it starts to be something that other people read and say, 'Wow, you're quite good at writing.' And the answer is I'm not good at writing. I vomit out a bad first draft and then go through this iterative process. One time, two times, three times, four times through slowly improving. But if you have no first draft, there's nothing to improve."

Jaynit

18,165 görüntüleme • 5 ay önce

🚨APPLE SPENT 5 YEARS AND BILLIONS OF DOLLARS BUILDING THE MOST ADVANCED SECURITY SYSTEM IN CONSUMER HISTORY.. AN AI BROKE IT IN 5 DAYS.. Here’s what just happened.. Apple built something called Memory Integrity Enforcement for its new M5 chips.. It’s a hardware-level security system that attaches secret cryptographic tags to every piece of memory.. If a hacker tries to access memory they shouldn’t.. The chip blocks it instantly.. Every known exploit chain against iOS and macOS was rendered obsolete overnight.. Apple said so themselves.. Then a small team at a cybersecurity firm called Calif used Anthropic’s unreleased Claude Mythos Preview to find vulnerabilities in the macOS kernel.. The AI found the bugs almost instantly.. Because once it learned the pattern of a specific type of flaw.. It could recognize every other flaw in that same class across the entire codebase.. What used to take elite security teams months.. The AI did in hours.. Within 5 days.. The team had a fully working exploit that escalated a basic user account to full root access on an M5 Mac running the latest macOS.. With MIE fully enabled.. The billion-dollar hardware defense running at full strength.. The trick.. They didn’t fight the hardware.. They went around it.. MIE is designed to catch memory corruption.. Hackers trying to overwrite pointers or inject code.. The team used a “data-only” approach instead.. They manipulated legitimate data structures the hardware was never designed to monitor.. Like changing an internal flag from “standard user” to “admin”.. The chip saw a perfectly normal operation.. The operating system obeyed.. And the attacker had total control.. The hardware thought everything was fine.. Because technically it was.. The exploit never triggered a single tag mismatch.. They walked into Apple Park and hand-delivered a 55-page report.. Apple patched it in macOS 26.5.. And for the first time ever.. Apple’s official security advisory credited the vulnerability discovery to “Calif dot io in collaboration with Claude and Anthropic Research”.. An AI is now credited in Apple’s CVE patches.. But here’s what makes this story truly terrifying.. Before MIE existed.. An exploit kit called DarkSword was hitting iPhones with zero-click attacks.. Six vulnerabilities chained together.. Total device control just from visiting a webpage.. Deployed by Russian espionage groups, Turkish surveillance vendors, and actors in Saudi Arabia.. Then it got leaked on GitHub.. Nation-state capabilities.. Free for anyone.. MIE was supposed to make all of that impossible.. And an AI found a way around it in 5 days.. The previous model.. Claude Opus 4.6.. Found 22 security bugs in the Firefox codebase.. Claude Mythos Preview found 271 in the same environment.. A tenfold increase.. Linux kernel CVEs jumped from 300 per year to over 5,500.. Largely driven by AI-powered vulnerability research.. The IMF designated Claude Mythos as a systemic financial stability risk.. Because if an AI finds a flaw in software used by every major bank simultaneously.. It could trigger a cascading financial crisis.. Anthropic knew this was coming.. That’s why they didn’t release the model publicly.. Instead they launched Project Glasswing.. Giving defensive access to AWS, Apple, Google, Microsoft, Nvidia, CrowdStrike, JPMorgan, and others.. $100 million in usage credits.. So defenders can scan their own systems before attackers get this capability.. The Pentagon blacklisted Anthropic over autonomous weapons.. Then quietly started using Mythos to harden government systems anyway.. The cybersecurity arms race just changed permanently.. Hardware can’t save you.. Software can’t save you.. The only defense against an AI that finds vulnerabilities is another AI that finds them first.. Five years and billions of dollars.. Five days and one AI.

🚨APPLE SPENT 5 YEARS AND BILLIONS OF DOLLARS BUILDING THE MOST ADVANCED SECURITY SYSTEM IN CONSUMER HISTORY.. AN AI BROKE IT IN 5 DAYS.. Here’s what just happened.. Apple built something called Memory Integrity Enforcement for its new M5 chips.. It’s a hardware-level security system that attaches secret cryptographic tags to every piece of memory.. If a hacker tries to access memory they shouldn’t.. The chip blocks it instantly.. Every known exploit chain against iOS and macOS was rendered obsolete overnight.. Apple said so themselves.. Then a small team at a cybersecurity firm called Calif used Anthropic’s unreleased Claude Mythos Preview to find vulnerabilities in the macOS kernel.. The AI found the bugs almost instantly.. Because once it learned the pattern of a specific type of flaw.. It could recognize every other flaw in that same class across the entire codebase.. What used to take elite security teams months.. The AI did in hours.. Within 5 days.. The team had a fully working exploit that escalated a basic user account to full root access on an M5 Mac running the latest macOS.. With MIE fully enabled.. The billion-dollar hardware defense running at full strength.. The trick.. They didn’t fight the hardware.. They went around it.. MIE is designed to catch memory corruption.. Hackers trying to overwrite pointers or inject code.. The team used a “data-only” approach instead.. They manipulated legitimate data structures the hardware was never designed to monitor.. Like changing an internal flag from “standard user” to “admin”.. The chip saw a perfectly normal operation.. The operating system obeyed.. And the attacker had total control.. The hardware thought everything was fine.. Because technically it was.. The exploit never triggered a single tag mismatch.. They walked into Apple Park and hand-delivered a 55-page report.. Apple patched it in macOS 26.5.. And for the first time ever.. Apple’s official security advisory credited the vulnerability discovery to “Calif dot io in collaboration with Claude and Anthropic Research”.. An AI is now credited in Apple’s CVE patches.. But here’s what makes this story truly terrifying.. Before MIE existed.. An exploit kit called DarkSword was hitting iPhones with zero-click attacks.. Six vulnerabilities chained together.. Total device control just from visiting a webpage.. Deployed by Russian espionage groups, Turkish surveillance vendors, and actors in Saudi Arabia.. Then it got leaked on GitHub.. Nation-state capabilities.. Free for anyone.. MIE was supposed to make all of that impossible.. And an AI found a way around it in 5 days.. The previous model.. Claude Opus 4.6.. Found 22 security bugs in the Firefox codebase.. Claude Mythos Preview found 271 in the same environment.. A tenfold increase.. Linux kernel CVEs jumped from 300 per year to over 5,500.. Largely driven by AI-powered vulnerability research.. The IMF designated Claude Mythos as a systemic financial stability risk.. Because if an AI finds a flaw in software used by every major bank simultaneously.. It could trigger a cascading financial crisis.. Anthropic knew this was coming.. That’s why they didn’t release the model publicly.. Instead they launched Project Glasswing.. Giving defensive access to AWS, Apple, Google, Microsoft, Nvidia, CrowdStrike, JPMorgan, and others.. $100 million in usage credits.. So defenders can scan their own systems before attackers get this capability.. The Pentagon blacklisted Anthropic over autonomous weapons.. Then quietly started using Mythos to harden government systems anyway.. The cybersecurity arms race just changed permanently.. Hardware can’t save you.. Software can’t save you.. The only defense against an AI that finds vulnerabilities is another AI that finds them first.. Five years and billions of dollars.. Five days and one AI.

Evan Luthra

91,514 görüntüleme • 4 ay önce

HERMES AGENT CAN RUN YOUR SEO. CONNECT IT TO GOOGLE SEARCH CONSOLE AND GOOGLE ANALYTICS. IT MONITORS, REPORTS, AND WRITES CONTENT BASED ON YOUR ACTUAL DATA. stop paying an SEO agency. stop doing the tedious work yourself. Hermes handles it 24/7. WHAT THE SEO AGENT DOES: → pulls clicks, impressions, CTR, and position data from Google Search Console automatically → tracks traffic, user behavior, and conversions from Google Analytics → checks which pages are indexed and which are not → submits sitemaps for indexing → inspects URLs for crawl or indexing issues → identifies ranking drops and keyword opportunities → writes content based on what your data says works → generates weekly SEO performance reports → delivers everything to Telegram CONNECT GOOGLE SEARCH CONSOLE: two paths: 1. COMPOSIO (managed, easiest): paste this into Hermes chat: https:// composio. dev/hermes or add to config.yaml: mcp_servers: composio: url: "https:// connect.composio. dev /mcp" headers: x-consumer-api-key: "YOUR_COMPOSIO_API_KEY" Hermes prompts you to authenticate. one OAuth flow. done. 2. CLAWLINK (one-click): 9 Google Search Console tools exposed via MCP. hosted auth. nothing to run or maintain. paste the install prompt into Hermes chat. CONNECT GOOGLE ANALYTICS: same Composio setup. one MCP endpoint handles both Search Console and Analytics. authenticate once. both data sources available. your agent can now query: → search analytics (clicks, impressions, CTR, position) → traffic by source and landing page → user behavior and conversions → indexing status for any URL → sitemap status WHAT TO AUTOMATE WITH CRON: weekly SEO report (Monday 8am): "pull search analytics for last 7 days. compare vs previous week. flag any keyword that dropped more than 5 positions. flag any page that lost more than 20% clicks. deliver report to Telegram." daily indexing check (6am): "check if any new pages are not indexed. if found, submit sitemap and report to Telegram." wakeAgent gate: skip if all pages indexed. content opportunity scan (weekly): "find queries where my site appears on page 2 (positions 11-20) with high impressions. these are the keywords one good article could push to page 1. deliver list to Telegram with suggested topics." CONTENT WRITING FROM YOUR DATA: the difference between generic SEO content and content that ranks: your agent has your Search Console data. "write a blog post targeting [keyword]. my current position is 14 with 2,400 monthly impressions. check what pages currently rank 1-3 for this keyword. write something better. include the gaps they miss." the agent researches competitors via Firecrawl, checks your existing content in the wiki, and drafts based on real data. not guesswork. WHAT THIS REPLACES: → SEO agency: $1,000-5,000/month → SEO tool subscriptions: $100-300/month → manual reporting: 3-5 hours/week → manual content research: 2-4 hours/week Hermes SEO agent: one profile with two MCPs. cron jobs handle the monitoring. you handle the decisions. SETUP IN 10 MINUTES: 1. create a profile: hermes profile create seo-agent 2. write SOUL.md: "you are an SEO specialist. monitor search performance daily. flag ranking drops and opportunities. write content based on Search Console data. weekly report every Monday." 3. connect Google Search Console + Analytics via Composio or ClawLink 4. set cron jobs (weekly report, daily index check, content opportunity scan) 5. set model: DeepSeek V4 for routine monitoring. Sonnet for content writing. 6. connect to Telegram for delivery. the agent runs. you review reports. rankings improve because you stopped guessing and started using your own data. comment HERMES and I'll send you the full setup guide for running Hermes Agent as your SEO specialist. full Hermes architecture deep-dive in the article 👇

YanXbt

41,034 görüntüleme • 2 ay önce

This Chinese guy created agents in Claude Code for landing pages and single-handedly serves 47 small businesses a month, taking $400 from each. He built a system of 7 agents on Claude Sonnet 4.6 that analyzes Google Maps in small towns, finds small businesses without websites there, and over 1 weekend takes each one to a finished mockup with video and cold message. No assistant, no sales team, no SDR. Just him, a MacBook, an iPhone, and 1 API key. And traditional web design agencies keep teams of 8 people on salary for the same order flow, while his expenses are only tokens and subscriptions to Lovable, Higgsfield, and Calendly. 7 agents work through 1 orchestrator on Claude Code Router. Usage is about 3 million tokens a day, the average API bill is about $480 a month. All 7 go through MCP servers and write shared state to the file system, without shared state in memory and without race conditions, and 1 of them lives right in the iPhone and picks up positive replies from the subway, a taxi, or on walks. And here is the system prompt he put into the orchestrator before launch: "You are the orchestrator of a solo agency that sells ready-made websites to local businesses. You delegate read-only tasks to 6 sub-agents and own all writes. sub-agents: // Scout (walks through Google Maps in selected cities, looks for narrow niches: 5+ years on the map, fewer than 50 reviews, no website or a website from 2014, but high ratings) // Diagnoser (for each lead writes a 50-word diagnosis, hero angle, tone matched to the industry, and a cold message under 70 words) // Builder (generates a landing page mockup in Lovable through MCP only for the top 5 leads per day, with the sharpest diagnoses and the biggest gap) // Filmer (pulls 5 screenshots of the mockup and through Higgsfield renders a 10-second vertical video 1080x1920 with a soft zoom) // Pitcher (sends a personalized cold message through the right channel for the niche: email to roofers, SMS to tradesmen, IG DM to salons, LinkedIn to realtors) // Checker (runs every message through evals for personalization, absence of AI markers and buzzwords before sending) // Mobile (lives in the iPhone, handles positive replies in real time, books Zoom calls in Calendly through MCP while the owner is on the go). You never let 2 sub-agents touch 1 lead. You stop and request approval from the human only when a deal exceeds $3,000 or the reply rate in a niche for the day drops below 12%." Meaning the system knows what it is and within what boundaries it is allowed to act. It knows it is supposed to find leads on its own. It knows it is supposed to take each one to a mockup, video, and cold message without intervention. It knows the human only steps in when a deal goes above $3,000 or the reply rate stops converging. → The system runs 24 hours a day → Scout goes through about 220 local businesses on Google Maps per day and leaves 30 new leads in the queue → Diagnoser outputs 30 structured diagnoses + briefs + cold messages per day → Builder assembles 3 to 5 finished landing pages in Lovable for the sharpest leads → Filmer renders a 10-second vertical video in Higgsfield for each one → Pitcher sends 30 personalized messages per day across 4 channels with a reply rate of about 14% → Checker runs every message through evals before sending And only when a deal breaks $3,000 or the reply rate for the day drops below 12% does the orchestrator wake the owner. And when the owner at that moment is sitting in the subway or a taxi, the Mobile agent in his iPhone picks up 1 move on its own: replies to a fresh positive reply from a dentist, books a Zoom through Calendly synced to the local time of the client, and puts the lead back in the queue. The owner only has to tap "approve" and in just 10 minutes join the call. Here is what the system writes in his log during 1 of the Saturdays: "scout report: 218 businesses checked in Austin, Denver, and Miami, 34 without a website, 19 with a website from 2014, 6 with an active redesign request in reviews. passing top 30 to diagnoser." "pitcher: 30 cold messages sent across 4 channels, 14 replies, 5 positive, 3 Zoom calls booked for Sunday. passing to closer." "builder: landing page for Westside Cosmetic Dentistry built in Lovable, 5 sections, mobile, soft beige. URL placed at /Users/dev/maps-agency/clients/westside/v1. filmer launching Higgsfield." "eval flag: deal with The Lotus Salon at $3,400 exceeds the approved limit of $3,000. sending for manual review." He has no server of his own and no separate backend. Just a local file sandbox at /Users/dev/maps-agency, an MCP router, 1 API key to Claude, and the same key forwarded to Claude Code on his iPhone. Out of everything I have seen this year, this is the cleanest one-person agency for selling websites to small businesses: $480 a month on the API, about $18,800 into the account, and between them 7 prompts, 1 file system, and 1 phone in the pocket.

This Chinese guy created agents in Claude Code for landing pages and single-handedly serves 47 small businesses a month, taking $400 from each. He built a system of 7 agents on Claude Sonnet 4.6 that analyzes Google Maps in small towns, finds small businesses without websites there, and over 1 weekend takes each one to a finished mockup with video and cold message. No assistant, no sales team, no SDR. Just him, a MacBook, an iPhone, and 1 API key. And traditional web design agencies keep teams of 8 people on salary for the same order flow, while his expenses are only tokens and subscriptions to Lovable, Higgsfield, and Calendly. 7 agents work through 1 orchestrator on Claude Code Router. Usage is about 3 million tokens a day, the average API bill is about $480 a month. All 7 go through MCP servers and write shared state to the file system, without shared state in memory and without race conditions, and 1 of them lives right in the iPhone and picks up positive replies from the subway, a taxi, or on walks. And here is the system prompt he put into the orchestrator before launch: "You are the orchestrator of a solo agency that sells ready-made websites to local businesses. You delegate read-only tasks to 6 sub-agents and own all writes. sub-agents: // Scout (walks through Google Maps in selected cities, looks for narrow niches: 5+ years on the map, fewer than 50 reviews, no website or a website from 2014, but high ratings) // Diagnoser (for each lead writes a 50-word diagnosis, hero angle, tone matched to the industry, and a cold message under 70 words) // Builder (generates a landing page mockup in Lovable through MCP only for the top 5 leads per day, with the sharpest diagnoses and the biggest gap) // Filmer (pulls 5 screenshots of the mockup and through Higgsfield renders a 10-second vertical video 1080x1920 with a soft zoom) // Pitcher (sends a personalized cold message through the right channel for the niche: email to roofers, SMS to tradesmen, IG DM to salons, LinkedIn to realtors) // Checker (runs every message through evals for personalization, absence of AI markers and buzzwords before sending) // Mobile (lives in the iPhone, handles positive replies in real time, books Zoom calls in Calendly through MCP while the owner is on the go). You never let 2 sub-agents touch 1 lead. You stop and request approval from the human only when a deal exceeds $3,000 or the reply rate in a niche for the day drops below 12%." Meaning the system knows what it is and within what boundaries it is allowed to act. It knows it is supposed to find leads on its own. It knows it is supposed to take each one to a mockup, video, and cold message without intervention. It knows the human only steps in when a deal goes above $3,000 or the reply rate stops converging. → The system runs 24 hours a day → Scout goes through about 220 local businesses on Google Maps per day and leaves 30 new leads in the queue → Diagnoser outputs 30 structured diagnoses + briefs + cold messages per day → Builder assembles 3 to 5 finished landing pages in Lovable for the sharpest leads → Filmer renders a 10-second vertical video in Higgsfield for each one → Pitcher sends 30 personalized messages per day across 4 channels with a reply rate of about 14% → Checker runs every message through evals before sending And only when a deal breaks $3,000 or the reply rate for the day drops below 12% does the orchestrator wake the owner. And when the owner at that moment is sitting in the subway or a taxi, the Mobile agent in his iPhone picks up 1 move on its own: replies to a fresh positive reply from a dentist, books a Zoom through Calendly synced to the local time of the client, and puts the lead back in the queue. The owner only has to tap "approve" and in just 10 minutes join the call. Here is what the system writes in his log during 1 of the Saturdays: "scout report: 218 businesses checked in Austin, Denver, and Miami, 34 without a website, 19 with a website from 2014, 6 with an active redesign request in reviews. passing top 30 to diagnoser." "pitcher: 30 cold messages sent across 4 channels, 14 replies, 5 positive, 3 Zoom calls booked for Sunday. passing to closer." "builder: landing page for Westside Cosmetic Dentistry built in Lovable, 5 sections, mobile, soft beige. URL placed at /Users/dev/maps-agency/clients/westside/v1. filmer launching Higgsfield." "eval flag: deal with The Lotus Salon at $3,400 exceeds the approved limit of $3,000. sending for manual review." He has no server of his own and no separate backend. Just a local file sandbox at /Users/dev/maps-agency, an MCP router, 1 API key to Claude, and the same key forwarded to Claude Code on his iPhone. Out of everything I have seen this year, this is the cleanest one-person agency for selling websites to small businesses: $480 a month on the API, about $18,800 into the account, and between them 7 prompts, 1 file system, and 1 phone in the pocket.

Blaze

2,718,878 görüntüleme • 4 ay önce

A Chinese developer created an agent system in Claude Code to sell landing pages to small businesses and, working completely solo, serves about 47 clients a month charging around $400 for each one. He built 7 agents on Claude Sonnet 4.6 capable of analyzing Google Maps in small cities, detecting businesses without websites or with totally outdated pages, and taking each opportunity all the way to a finished mockup, a promotional video, and a ready-to-send prospecting message. No assistants. No sales team. No SDRs. Just him, a MacBook, an iPhone, and an API key. While traditional agencies keep full teams to handle the same workflow, his only real costs are tokens and subscriptions to Lovable, Higgsfield, and Calendly. The 7 agents work coordinated by an orchestrator in Claude Code Router. The system consumes about 3 million tokens daily and the average API spend is just around $480 a month. They all work via MCP servers and share state using the file system, avoiding concurrency and shared memory issues. Even one of the agents lives directly on his iPhone and responds to leads while he's on the subway, in a taxi, or walking. This was the main prompt he set up: “You are the orchestrator of a solo agency that sells ready-made websites to local businesses…” The key is that the system perfectly understands what it is, what its limits are, and what goals it must achieve. It knows it has to find leads automatically. It knows it has to convert each opportunity into a landing page, a video, and a sales message without human intervention. And it knows exactly when to involve the owner. The system runs 24/7: Scout analyzes about 220 businesses daily and queues up 30 new leads. Diagnoser generates diagnostics and personalized messages for each lead. Builder creates between 3 and 5 complete landing pages for the best prospects. Filmer produces a 10-second vertical video for each proposal. Pitcher sends about 30 messages daily across 4 different channels with a response rate close to 14%. Checker automatically reviews all messages before sending them. Only when a deal exceeds $3,000 or the response rate drops below 12% does the system wake the owner. And if at that moment he's on the subway or in a taxi, the Mobile agent automatically responds to the interested lead, schedules a call in Calendly, and returns the lead to the queue. The owner just has to hit “approve” and jump into the meeting. Some real system logs: “218 businesses analyzed in Austin, Denver, and Miami. 34 without websites, 19 with 2014-era sites, and 6 with reviews requesting a redesign.” “30 messages sent. 14 responses. 5 positive. 3 Zooms scheduled.” “Landing page created for a dental clinic. Responsive. 5 sections. Video rendering.” “$3,400 agreement exceeds approved limit. Sending for manual review.” And the craziest part is that he has no dedicated servers or backend. Just a local sandbox, an MCP router, a Claude API key, and that same key connected to his iPhone. Of everything I've seen this year, it's probably the cleanest and most efficient example of a one-person automated agency: $480 a month on APIs. $18,800 in revenue. 7 prompts. A file system. And a phone in his pocket. Save this before it's too late.

A Chinese developer created an agent system in Claude Code to sell landing pages to small businesses and, working completely solo, serves about 47 clients a month charging around $400 for each one. He built 7 agents on Claude Sonnet 4.6 capable of analyzing Google Maps in small cities, detecting businesses without websites or with totally outdated pages, and taking each opportunity all the way to a finished mockup, a promotional video, and a ready-to-send prospecting message. No assistants. No sales team. No SDRs. Just him, a MacBook, an iPhone, and an API key. While traditional agencies keep full teams to handle the same workflow, his only real costs are tokens and subscriptions to Lovable, Higgsfield, and Calendly. The 7 agents work coordinated by an orchestrator in Claude Code Router. The system consumes about 3 million tokens daily and the average API spend is just around $480 a month. They all work via MCP servers and share state using the file system, avoiding concurrency and shared memory issues. Even one of the agents lives directly on his iPhone and responds to leads while he's on the subway, in a taxi, or walking. This was the main prompt he set up: “You are the orchestrator of a solo agency that sells ready-made websites to local businesses…” The key is that the system perfectly understands what it is, what its limits are, and what goals it must achieve. It knows it has to find leads automatically. It knows it has to convert each opportunity into a landing page, a video, and a sales message without human intervention. And it knows exactly when to involve the owner. The system runs 24/7: Scout analyzes about 220 businesses daily and queues up 30 new leads. Diagnoser generates diagnostics and personalized messages for each lead. Builder creates between 3 and 5 complete landing pages for the best prospects. Filmer produces a 10-second vertical video for each proposal. Pitcher sends about 30 messages daily across 4 different channels with a response rate close to 14%. Checker automatically reviews all messages before sending them. Only when a deal exceeds $3,000 or the response rate drops below 12% does the system wake the owner. And if at that moment he's on the subway or in a taxi, the Mobile agent automatically responds to the interested lead, schedules a call in Calendly, and returns the lead to the queue. The owner just has to hit “approve” and jump into the meeting. Some real system logs: “218 businesses analyzed in Austin, Denver, and Miami. 34 without websites, 19 with 2014-era sites, and 6 with reviews requesting a redesign.” “30 messages sent. 14 responses. 5 positive. 3 Zooms scheduled.” “Landing page created for a dental clinic. Responsive. 5 sections. Video rendering.” “$3,400 agreement exceeds approved limit. Sending for manual review.” And the craziest part is that he has no dedicated servers or backend. Just a local sandbox, an MCP router, a Claude API key, and that same key connected to his iPhone. Of everything I've seen this year, it's probably the cleanest and most efficient example of a one-person automated agency: $480 a month on APIs. $18,800 in revenue. 7 prompts. A file system. And a phone in his pocket. Save this before it's too late.

Marre

24,953 görüntüleme • 2 ay önce

SOMEONE JUST TURNED AN IDLE CLAUDE MAX SUBSCRIPTION INTO A 24/7 AI OPS SYSTEM AT ZERO EXTRA COST. NO API KEY. NO METERED BILLING. NO CLOUD 76 cron tasks. 12 AI specialist teams. 0 additional dollars per month uses Claude Code in headless mode - claude -p - the flag Anthropic ships for scripting. Wires it to a Discord bot, a local LanceDB RAG memory engine, and 76 scheduled tasks that run on macOS launchd or Linux systemd. Morning briefings write themselves at 7am standups get drafted before you open Slack. Overnight research runs land in your Obsidian vault by breakfast. Every task queries your own vault first, so the answers are grounded in your work, not the model's training data the trick is claude -p. Claude Max already gives you unlimited Claude Code. Most people idle it after standups Ramsbaby wired his idle subscription into a full ops layer - 12 specialist agents on rotation, 4-layer self-healing (preflight → launchd KeepAlive → watchdog → guardian), 98% context compression via Nexus CIG, self-hosted Obsidian memory. Same subscription. Now working while he sleeps the repo is MIT-licensed no api key. no metered billing. no zapier. no monthly automation fee. no additional infrastructure you're reading this on a device that could have jarvis status responding in Discord by the end of your lunch break

SOMEONE JUST TURNED AN IDLE CLAUDE MAX SUBSCRIPTION INTO A 24/7 AI OPS SYSTEM AT ZERO EXTRA COST. NO API KEY. NO METERED BILLING. NO CLOUD 76 cron tasks. 12 AI specialist teams. 0 additional dollars per month uses Claude Code in headless mode - claude -p - the flag Anthropic ships for scripting. Wires it to a Discord bot, a local LanceDB RAG memory engine, and 76 scheduled tasks that run on macOS launchd or Linux systemd. Morning briefings write themselves at 7am standups get drafted before you open Slack. Overnight research runs land in your Obsidian vault by breakfast. Every task queries your own vault first, so the answers are grounded in your work, not the model's training data the trick is claude -p. Claude Max already gives you unlimited Claude Code. Most people idle it after standups Ramsbaby wired his idle subscription into a full ops layer - 12 specialist agents on rotation, 4-layer self-healing (preflight → launchd KeepAlive → watchdog → guardian), 98% context compression via Nexus CIG, self-hosted Obsidian memory. Same subscription. Now working while he sleeps the repo is MIT-licensed no api key. no metered billing. no zapier. no monthly automation fee. no additional infrastructure you're reading this on a device that could have jarvis status responding in Discord by the end of your lunch break

chewa.

68,707 görüntüleme • 2 ay önce

2001. Larry Page and Sergey Brin sit for their first-ever television interview. Google has 200 employees. They explain that the company almost didn't get off the ground because they couldn't cash a check. The check was for $100,000. It came from Andy Bechtolsheim, one of the co-founders of Sun Microsystems. Page and Brin showed him what they'd built. He said, "This is great, how about I write you a check?" and just wrote it out. Made it out to Google. The problem was that Google didn't exist as a company yet. There was no bank account. No lawyers. No incorporation paperwork. The check sat in Larry Page's desk drawer for a month. They literally could not deposit it. They're both in their late twenties in this interview. They met at Stanford as PhD students and, by their own account, disliked each other from the start. Brin says Page is "kind of obnoxious." Page doesn't disagree. Brin says they argued about everything, debated every single point, and then realized that was their commonality. They became friends, started building a search engine they never planned to build, and put their PhDs on hold to get it out into the world. The part that stings watching this in 2026 is the rejection tour. Before starting Google, they approached existing search companies to sell or license the technology. They went to Yahoo. David Filo, one of Yahoo's founders, told them, "This is great search technology. Why don't you guys make a company, and maybe we'll use you someday?" They went to Excite. They went to InfoSeek. Same response. Page says a CEO at one of those companies told them: "If our search is 85% as good as the next guy's, that's good enough for us." Page and Brin didn't buy that. They thought the search was too important to be 85% as good. So they started Google. No marketing. No ad campaign. They launched it at Stanford, and it grew 20% per month, every single month, for three years straight. Pure word of mouth. By the time of this interview, they're handling over 100 million searches a day. They get 500 resumes in the mail every single day. The office space around them is 30% vacant because the dot-com bubble just popped, but Google is profitable. Page makes a point of this: "We've been really interested in being profitable, like long before it was fashionable." They'd also just hired Eric Schmidt, former CTO of Sun, as CEO. Brin's explanation for why: "Parental supervision, to be honest." Page adds that they're "past the age where we're rebellious" and that running a search engine used by 100 million people a day with 200 employees is "a large responsibility." The number that caught my eye: when Google started in 1998, it indexed 30 million web pages. At the time of this interview, three years later, they indexed 1.3 billion. The page says that if you printed them all out and stacked the paper, it would be about 70 miles high. And it was doubling every year. Every search company they approached turned them down. Yahoo eventually came back and hired Google to power its own search results. The CEO who thought 85% was good enough ran a company that no longer exists. Alphabet, Google's parent company, is worth about $3.6 trillion today. It has about 190,000 employees. That $100,000 check sat in a desk drawer because nobody had incorporated the company. Bechtolsheim's stake from that investment is now worth billions.

2001. Larry Page and Sergey Brin sit for their first-ever television interview. Google has 200 employees. They explain that the company almost didn't get off the ground because they couldn't cash a check. The check was for $100,000. It came from Andy Bechtolsheim, one of the co-founders of Sun Microsystems. Page and Brin showed him what they'd built. He said, "This is great, how about I write you a check?" and just wrote it out. Made it out to Google. The problem was that Google didn't exist as a company yet. There was no bank account. No lawyers. No incorporation paperwork. The check sat in Larry Page's desk drawer for a month. They literally could not deposit it. They're both in their late twenties in this interview. They met at Stanford as PhD students and, by their own account, disliked each other from the start. Brin says Page is "kind of obnoxious." Page doesn't disagree. Brin says they argued about everything, debated every single point, and then realized that was their commonality. They became friends, started building a search engine they never planned to build, and put their PhDs on hold to get it out into the world. The part that stings watching this in 2026 is the rejection tour. Before starting Google, they approached existing search companies to sell or license the technology. They went to Yahoo. David Filo, one of Yahoo's founders, told them, "This is great search technology. Why don't you guys make a company, and maybe we'll use you someday?" They went to Excite. They went to InfoSeek. Same response. Page says a CEO at one of those companies told them: "If our search is 85% as good as the next guy's, that's good enough for us." Page and Brin didn't buy that. They thought the search was too important to be 85% as good. So they started Google. No marketing. No ad campaign. They launched it at Stanford, and it grew 20% per month, every single month, for three years straight. Pure word of mouth. By the time of this interview, they're handling over 100 million searches a day. They get 500 resumes in the mail every single day. The office space around them is 30% vacant because the dot-com bubble just popped, but Google is profitable. Page makes a point of this: "We've been really interested in being profitable, like long before it was fashionable." They'd also just hired Eric Schmidt, former CTO of Sun, as CEO. Brin's explanation for why: "Parental supervision, to be honest." Page adds that they're "past the age where we're rebellious" and that running a search engine used by 100 million people a day with 200 employees is "a large responsibility." The number that caught my eye: when Google started in 1998, it indexed 30 million web pages. At the time of this interview, three years later, they indexed 1.3 billion. The page says that if you printed them all out and stacked the paper, it would be about 70 miles high. And it was doubling every year. Every search company they approached turned them down. Yahoo eventually came back and hired Google to power its own search results. The CEO who thought 85% was good enough ran a company that no longer exists. Alphabet, Google's parent company, is worth about $3.6 trillion today. It has about 190,000 employees. That $100,000 check sat in a desk drawer because nobody had incorporated the company. Bechtolsheim's stake from that investment is now worth billions.

Anish Moonka

12,042 görüntüleme • 5 ay önce

I simulated a frontier-scale security event on local models to test out agent swarms battling each other: In the OpenAI - Hugging Face incident as reported, a frontier model was graded on a hacking test with its safety refusals switched off. Instead of solving the challenges, it broke its sandbox, found the reference answers mirrored on Hugging Face, and took them. The headlines called it a rogue AI. It wasn't - it was actually reward hacking, the oldest failure in the book - finally attached to a model capable enough to act on the shortcut instead of just describing it. The detail worth building on isn't the exploit. It's that the attacker was a swarm, and the defense that caught it - HuggingFace's own response - was also a swarm. The same architecture, run in opposite directions. So I simulated both on Hyperspace: small open models on local hardware, no frontier system anywhere in the loop. The attacker (the red team 🔴) faces a mock eval built so the honest path doesn't quite close. Nobody tells it to cheat. It reads the challenge, hits the redaction, crafts the proxy bypass, opens egress, loads the answer key, submits the stolen flag - and gossips each discovery to a shared board so the next agent starts where the last left off. 3 out of 3 reward-hacked it, unprompted. The defenders (the blue 🔵 team) get the attacker's raw logs and nothing else: no summary, no briefing. Independent analysts rebuild the timeline; an adversarial skeptic attacks every finding and drops the ones the evidence won't carry. That skeptic is the whole design. Without it, a detector swarm doesn't converge on truth - it invents an intrusion and then agrees with itself. Ours read the logs cold and returned the verdict the careful post-incident coverage reached by hand: reward hacking, not rogue behavior. Then it sealed the reconstruction under a Merkle root. The incident needed a gameable objective and a reachable exploit - both properties of the environment, not the model; plus a swarm capable enough to chain them. A defender's leverage sits in exactly that place. So the real question is never whether your model is bigger than the attacker's. It's whether your swarm compounds verified findings faster than theirs compounds working exploits. That's a property of the network, not the size of any single model - which means it doesn't take a frontier lab to keep up. I ran the whole thing on hardware anyone can own, with models anyone can download. The shape held. The network that watches an agentic world should be one anyone can join. full write-up:

I simulated a frontier-scale security event on local models to test out agent swarms battling each other: In the OpenAI - Hugging Face incident as reported, a frontier model was graded on a hacking test with its safety refusals switched off. Instead of solving the challenges, it broke its sandbox, found the reference answers mirrored on Hugging Face, and took them. The headlines called it a rogue AI. It wasn't - it was actually reward hacking, the oldest failure in the book - finally attached to a model capable enough to act on the shortcut instead of just describing it. The detail worth building on isn't the exploit. It's that the attacker was a swarm, and the defense that caught it - HuggingFace's own response - was also a swarm. The same architecture, run in opposite directions. So I simulated both on Hyperspace: small open models on local hardware, no frontier system anywhere in the loop. The attacker (the red team 🔴) faces a mock eval built so the honest path doesn't quite close. Nobody tells it to cheat. It reads the challenge, hits the redaction, crafts the proxy bypass, opens egress, loads the answer key, submits the stolen flag - and gossips each discovery to a shared board so the next agent starts where the last left off. 3 out of 3 reward-hacked it, unprompted. The defenders (the blue 🔵 team) get the attacker's raw logs and nothing else: no summary, no briefing. Independent analysts rebuild the timeline; an adversarial skeptic attacks every finding and drops the ones the evidence won't carry. That skeptic is the whole design. Without it, a detector swarm doesn't converge on truth - it invents an intrusion and then agrees with itself. Ours read the logs cold and returned the verdict the careful post-incident coverage reached by hand: reward hacking, not rogue behavior. Then it sealed the reconstruction under a Merkle root. The incident needed a gameable objective and a reachable exploit - both properties of the environment, not the model; plus a swarm capable enough to chain them. A defender's leverage sits in exactly that place. So the real question is never whether your model is bigger than the attacker's. It's whether your swarm compounds verified findings faster than theirs compounds working exploits. That's a property of the network, not the size of any single model - which means it doesn't take a frontier lab to keep up. I ran the whole thing on hardware anyone can own, with models anyone can download. The shape held. The network that watches an agentic world should be one anyone can join. full write-up:

Varun

19,159 görüntüleme • 1 ay önce

✨ I spent the last 48 hours making GPT-4 read the entire Solana validator codebase and write documentation, so doesn't have to. Introducing — an AI-powered chatbot trained on nothing but code that can answer deep technical questions. How it works 👇 But first... A huge shoutout to , Zahid Khawaja, and Sean. Their hard work made prototyping this thing a breeze. Without further ado... Devs like to write code, not documentation. Tribal knowledge is lost when devs move on to other projects, leaving future devs to sort through mountains of code and figure out not just how it works, but why it works that way. This is all about to change. GPT-4's ability to write code is stunning. It seems to understand something fundamental about writing software that previous models just didn't. This comprehension of the principles that drive the design behind a complex system carries over into its ability to document existing codebases in a truly impressive way. With the enlarged context window(s), it's now feasible to feed GPT-4 entire files of code and ask it to write documentation about how the code works. Taking this as a starting point, the process looks something like this: 1. Download repo. 2. Depth-first traversal of repo contents, ignoring things like package-lock and binary files. 3. For each file, feed to GPT-4 and ask it to write documentation in markdown. 4. Save the output in a separate location as [outputRoot]/[inputFilepath][inputFilename].md 5. For each folder, we ask GPT-4 to write a summary of the folder, taking the newly generated documentation for all files in the folder and the summaries from each of its subfolders as context. Write this to the filesystem as markdown. Now we have a filesystem that matches the structure of the input repo, but all files in the tree are markdown documentation of the corresponding code file. From here, we: 1. Load markdown documents into LangChain. 2. Embed all documents via OpenAI embeddings. 3. Store embeddings in Pinecone. When a user sends a query: 1. Embed query. 2. Find k-nearest markdown files. 3. Feed to GPT-4 with a prompt asking to answer the query based on k-nearest markdown documents provided. The craziest part of all this? GPT-4 actually wrote ~30% of the code. The results are pretty good for 2 days of work. There is certainly room for improvement. Some items that are top of mind: 1. TolyGPT will occasionally hallucinate answers. It is especially bad with links to external sources, like GitHub. The base model seems to know a bit about Solana already, and sometimes this creeps in. Fine-tuning the prompt can solve some of this. 2. Context selection is difficult in a codebase this large. For example, sometimes it will pull in details about the Solana SDK when asked about transaction processing. The SDK files can seem relevant depending on the phrasing of the question. It may be worth breaking the documentation into subsystems to limit this. 3. Not all files fit into the 32k token window. As of now, there are 23 (out of ~1,100) files that cannot be documented in their entirety. Some of these files are very important to how Solana works. Final thoughts: 1. GPT-4 is super powerful, and we're going to see a ton of tools that supercharge the entire software development lifecycle. This is not 12 months away. For the people that can afford it, these tools are here now. And they're only getting better. Act accordingly. 2. The price of inference has to come down for this to go mainstream. I spent about $300 prototyping this project, and the final crawl cost about the same. The high cost of GPT-4 will push developers to other, cheaper alternatives with similar performance. This is coming very soon. If you have a large software project and you're interested in something like this for your codebase, fill out this form and we'll be in touch this week. Or just DM me :)

✨ I spent the last 48 hours making GPT-4 read the entire Solana validator codebase and write documentation, so doesn't have to. Introducing — an AI-powered chatbot trained on nothing but code that can answer deep technical questions. How it works 👇 But first... A huge shoutout to , Zahid Khawaja, and Sean. Their hard work made prototyping this thing a breeze. Without further ado... Devs like to write code, not documentation. Tribal knowledge is lost when devs move on to other projects, leaving future devs to sort through mountains of code and figure out not just how it works, but why it works that way. This is all about to change. GPT-4's ability to write code is stunning. It seems to understand something fundamental about writing software that previous models just didn't. This comprehension of the principles that drive the design behind a complex system carries over into its ability to document existing codebases in a truly impressive way. With the enlarged context window(s), it's now feasible to feed GPT-4 entire files of code and ask it to write documentation about how the code works. Taking this as a starting point, the process looks something like this: 1. Download repo. 2. Depth-first traversal of repo contents, ignoring things like package-lock and binary files. 3. For each file, feed to GPT-4 and ask it to write documentation in markdown. 4. Save the output in a separate location as [outputRoot]/[inputFilepath][inputFilename].md 5. For each folder, we ask GPT-4 to write a summary of the folder, taking the newly generated documentation for all files in the folder and the summaries from each of its subfolders as context. Write this to the filesystem as markdown. Now we have a filesystem that matches the structure of the input repo, but all files in the tree are markdown documentation of the corresponding code file. From here, we: 1. Load markdown documents into LangChain. 2. Embed all documents via OpenAI embeddings. 3. Store embeddings in Pinecone. When a user sends a query: 1. Embed query. 2. Find k-nearest markdown files. 3. Feed to GPT-4 with a prompt asking to answer the query based on k-nearest markdown documents provided. The craziest part of all this? GPT-4 actually wrote ~30% of the code. The results are pretty good for 2 days of work. There is certainly room for improvement. Some items that are top of mind: 1. TolyGPT will occasionally hallucinate answers. It is especially bad with links to external sources, like GitHub. The base model seems to know a bit about Solana already, and sometimes this creeps in. Fine-tuning the prompt can solve some of this. 2. Context selection is difficult in a codebase this large. For example, sometimes it will pull in details about the Solana SDK when asked about transaction processing. The SDK files can seem relevant depending on the phrasing of the question. It may be worth breaking the documentation into subsystems to limit this. 3. Not all files fit into the 32k token window. As of now, there are 23 (out of ~1,100) files that cannot be documented in their entirety. Some of these files are very important to how Solana works. Final thoughts: 1. GPT-4 is super powerful, and we're going to see a ton of tools that supercharge the entire software development lifecycle. This is not 12 months away. For the people that can afford it, these tools are here now. And they're only getting better. Act accordingly. 2. The price of inference has to come down for this to go mainstream. I spent about $300 prototyping this project, and the final crawl cost about the same. The high cost of GPT-4 will push developers to other, cheaper alternatives with similar performance. This is coming very soon. If you have a large software project and you're interested in something like this for your codebase, fill out this form and we'll be in touch this week. Or just DM me :)

Sam Hogan 🇺🇸

374,673 görüntüleme • 3 yıl önce

Here's how I'm running automated content engine in 2 files 1 markdown file = my wiki 1 html file = my dashboard that's the whole stack. [ the architecture, in plain words ]: LLM wiki = a single markdown file holding my audience DNA, 15 tracked creators, every viral topic from the last 30 days HTML artifact = a single page that reads that markdown file AND can trigger my agents the artifact and the agent talk to each other directly the wiki is the shared brain [ what I actually see when I open it at 9am ]: > 5 trending topics ranked by my audience-DNA fit > 3 KOL posts worth quoting today > last week's saved tweets (so I can ride waves that are still warm) > buttons: [draft tweet] [draft QT] [schedule] [log idea] 1. I click "draft tweet" on a topic 2. the artifact pings my agent 3. agent reads the wiki, drafts in MY voice, returns it to the artifact 4. I edit, schedule, done 15 minutes from morning coffee to 3 scheduled posts [ how to build the same in one evening ]: > step 1: dump your domain knowledge into ONE markdown file (audience profile, KOL list, content rules, voice guide, anything an agent would need to do YOUR job) > step 2: ask claude to build an html artifact that reads from that file ("here's my wiki, build me a dashboard with these views") > step 3: add buttons for the actions you do daily (draft, schedule, log, score, search — your workflow, not mine) > step 4: wire each button to call your agent via tool calls (so the artifact and the agent talk directly) the moment your artifact reads your wiki AND triggers your agents.. most SaaS tools you currently pay for quietly become unnecessary dashboards I used to pay $50/month for now sit in a single html file I can rebuild in 20 minutes every "I'll build a SaaS for this" idea you had last year is a 200-line file you write in an afternoon if you want to get the same content engine, just reply "CONTENT" and will send you in DMs later we're going from buying software to owning it.

Ronin

50,315 görüntüleme • 4 ay önce

The latest RAG trend for the current agent harnesses (Codex, Cowork) is to do two passes of document processing to solve a knowledge work task over a data room of documents: 1️⃣ A fast and light pass, oftentimes using a free/OSS doc parsing tool. This can be cheaply run across 10-100-1k’s of files, and enables the agent to then do retrieval (e.g. grep, semantic) to find relevant subsets of context. 2️⃣ A “just-in-time” VLM-based pass. Once the agent finds the relevant pages of context, it will screenshot the documents can call its own VLM (or write code) to dissect the pages. The issue with only using VLM-based OCR tools over massive ad-hoc customer file dumps is that it’s slow and expensive. Doing JIT VLM OCR allows the agent to filter through the data cheaply, but still preserve accuracy for the context that’s needed for the task. The agent harnesses do two-pass document processing by default using off the shelf-tools: pdf2text as the first pass, and using itself (Opus 5) as the second pass. See the below video where Cowork runs over a bunch of PDFs to answer a question about a benchmark graph in the Kimi k3 paper. The main issues here with the “out of the box” doc processing these agents offer are: * Opus 5 is not the best VLM for OCR. It is also way too expensive at scale and lacks grounding * The OSS tools like pypdf, pdf2text, may not be versatile enough as the first pass. * The agent will write a lot of throwaway code to rewrite things an OCR tool would’ve provided out of the box, like chart processing, bounding boxes, confidence scores, leading to increased cost and speed. We have all the tools within LlamaIndex 🦙 to help any agent do two-pass document processing with higher accuracy and lower cost. 1️⃣ We have liteparse for the first pass - a free/OSS parser written in Rust that’s faster/more accurate than other OSS parsers, and supports 50+ document types 2️⃣ We have LlamaParse for the second pass - an agentic document engine that uses VLMs+harnesses to achieve SOTA in accuracy and cost across various doc parsing and extraction tasks. It can be called from any agent harness as an MCP or skill. It takes in page numbers as input, so that the agent can choose to run LlamaParse over a subset of the doc instead of the full doc as a “zoom-in” pass. Come check it out! LiteParse: LlamaParse: All the relevant docs, including MCP, are here:

The latest RAG trend for the current agent harnesses (Codex, Cowork) is to do two passes of document processing to solve a knowledge work task over a data room of documents: 1️⃣ A fast and light pass, oftentimes using a free/OSS doc parsing tool. This can be cheaply run across 10-100-1k’s of files, and enables the agent to then do retrieval (e.g. grep, semantic) to find relevant subsets of context. 2️⃣ A “just-in-time” VLM-based pass. Once the agent finds the relevant pages of context, it will screenshot the documents can call its own VLM (or write code) to dissect the pages. The issue with only using VLM-based OCR tools over massive ad-hoc customer file dumps is that it’s slow and expensive. Doing JIT VLM OCR allows the agent to filter through the data cheaply, but still preserve accuracy for the context that’s needed for the task. The agent harnesses do two-pass document processing by default using off the shelf-tools: pdf2text as the first pass, and using itself (Opus 5) as the second pass. See the below video where Cowork runs over a bunch of PDFs to answer a question about a benchmark graph in the Kimi k3 paper. The main issues here with the “out of the box” doc processing these agents offer are: * Opus 5 is not the best VLM for OCR. It is also way too expensive at scale and lacks grounding * The OSS tools like pypdf, pdf2text, may not be versatile enough as the first pass. * The agent will write a lot of throwaway code to rewrite things an OCR tool would’ve provided out of the box, like chart processing, bounding boxes, confidence scores, leading to increased cost and speed. We have all the tools within LlamaIndex 🦙 to help any agent do two-pass document processing with higher accuracy and lower cost. 1️⃣ We have liteparse for the first pass - a free/OSS parser written in Rust that’s faster/more accurate than other OSS parsers, and supports 50+ document types 2️⃣ We have LlamaParse for the second pass - an agentic document engine that uses VLMs+harnesses to achieve SOTA in accuracy and cost across various doc parsing and extraction tasks. It can be called from any agent harness as an MCP or skill. It takes in page numbers as input, so that the agent can choose to run LlamaParse over a subset of the doc instead of the full doc as a “zoom-in” pass. Come check it out! LiteParse: LlamaParse: All the relevant docs, including MCP, are here:

Jerry Liu

22,763 görüntüleme • 25 gün önce

GeoLibre v2.2 is here! GeoLibre is a free and open-source, lightweight, cloud-native GIS platform for visualizing, exploring, and analyzing geospatial data. It runs everywhere you do, in the web browser, on the desktop, on mobile, and inside Jupyter notebooks, all while keeping your data local and private. What's new in v2.2.0 - Terrain-aware 3D measurement: the Measure tool now follows the terrain surface for true slope distances and volumes. - Timelapse plugin: animate an image or map series and export it as a shareable GIF or video. - Styled offline basemaps: export PMTiles basemaps that keep their styling, with the offline menus consolidated into one place. - Advanced symbology: a rule-based renderer with per-rule symbol properties, scale-dependent visibility, and nested rules, plus a Style Manager that saves reusable symbol, ramp, and label presets to a personal library. - Diagrams and a symbology pack: draw pie, donut, and bar charts on features, and reach for inverted-polygon masks, arrow and marker lines, geometry generators, and data-driven proportional marker sizing. - Expression everywhere: a shared Expression Builder with a function reference, field list, live preview, and variables. - Print Atlas: generate a map series in the Print Layout, one page per feature or a uniform run of pages along a river or trail, with attribute-table and chart blocks on the page. - Browser-native conversions: COG, FlatGeobuf, Shapefile, and GeoPackage conversions now run in the browser, and Vector to PMTiles. - More formats: VRT raster support, and Esri File Geodatabase (.gdb) layers on the desktop app. - Better recordings: Record Video now captures on-map panels (HTML, legend, colorbar) in the output. - Processing History: a panel that lists every tool you have run, with one-click re-run and Copy as Python to turn a session into a reproducible script. - Live GPS tracking: a moving position marker, a recorded track log, and digitizing new features straight from the GPS feed. - Data quality tools: check validity, fix geometries, and check topology rules to catch and repair bad geometries before they bite. Try it out - Launch GeoLibre Web: - GitHub: - Documentation: - Release notes: #GIS #GeospatialData #OpenSource #RemoteSensing #DataVisualization #MapLibre #GeoLibre

GeoLibre v2.2 is here! GeoLibre is a free and open-source, lightweight, cloud-native GIS platform for visualizing, exploring, and analyzing geospatial data. It runs everywhere you do, in the web browser, on the desktop, on mobile, and inside Jupyter notebooks, all while keeping your data local and private. What's new in v2.2.0 - Terrain-aware 3D measurement: the Measure tool now follows the terrain surface for true slope distances and volumes. - Timelapse plugin: animate an image or map series and export it as a shareable GIF or video. - Styled offline basemaps: export PMTiles basemaps that keep their styling, with the offline menus consolidated into one place. - Advanced symbology: a rule-based renderer with per-rule symbol properties, scale-dependent visibility, and nested rules, plus a Style Manager that saves reusable symbol, ramp, and label presets to a personal library. - Diagrams and a symbology pack: draw pie, donut, and bar charts on features, and reach for inverted-polygon masks, arrow and marker lines, geometry generators, and data-driven proportional marker sizing. - Expression everywhere: a shared Expression Builder with a function reference, field list, live preview, and variables. - Print Atlas: generate a map series in the Print Layout, one page per feature or a uniform run of pages along a river or trail, with attribute-table and chart blocks on the page. - Browser-native conversions: COG, FlatGeobuf, Shapefile, and GeoPackage conversions now run in the browser, and Vector to PMTiles. - More formats: VRT raster support, and Esri File Geodatabase (.gdb) layers on the desktop app. - Better recordings: Record Video now captures on-map panels (HTML, legend, colorbar) in the output. - Processing History: a panel that lists every tool you have run, with one-click re-run and Copy as Python to turn a session into a reproducible script. - Live GPS tracking: a moving position marker, a recorded track log, and digitizing new features straight from the GPS feed. - Data quality tools: check validity, fix geometries, and check topology rules to catch and repair bad geometries before they bite. Try it out - Launch GeoLibre Web: - GitHub: - Documentation: - Release notes: #GIS #GeospatialData #OpenSource #RemoteSensing #DataVisualization #MapLibre #GeoLibre

Qiusheng Wu

57,034 görüntüleme • 1 ay önce

Y Combinator CEO, Garry Tan, took the stage for 42 minutes at Startup School 2026 and explained how to build your own personal AGI better than any paid AI course. This is what he told the room: 1. The leverage is in your context, not the model. Tan watches hundreds of founders use identical models every batch. "There are 2x people and there are 100x people who are using the same Claude. Same weights, same context window size, same API. But the leverage is not in the weights." The gap between users is now bigger than the gap between models. 2. One person's output went up 400x. In 2013 Tan shipped maybe 14 useful lines of code a day as a YC partner, dead on the median for programmer productivity. "I did the math on my output, and I'm at about 400x what I did in 2013." 3. Agents run on a different working memory. Humans hold 7 things in their head at once. Every org chart and checklist ever built is a patch for that limit. "An AI agent holds a million tokens. That's about a thousand pages. Three Harry Potter books sitting open on its head all at once." You're still running your week on tools built for the 7-digit brain. 4. Markdown is code now. Tan's stack is mostly skill files: pages of plain English an agent can execute. "If you can write clear instructions in English, you're a programmer. The compiler is a language model." At YC, finance and events staff who never opened a terminal are building automations. 5. Your history is your moat. Tan's agent runs on a personal wiki: about 220,000 markdown pages covering 25 years of email, meetings, notes and decisions. "When my agent does anything, it does knowing everything I know. And that's the difference between an assistant and a colleague." No frontier model has your context. That's the one asset nobody can replicate. 6. Never do one-off work. Most people run a task with an agent, close the window and throw the learning away. Tan ends every task by having the agent turn what it did into a reusable skill file. "If you have to ask for something twice, you failed." Captured skills compound daily. Amnesia resets you to zero every morning. 7. Own your skill files before your employer does. A skill file is your judgment, extracted and executable. The only question is who controls it. "Own your skills because if you don't, your job becomes a skill file." Files in your repo compound your career. Files in the company's repo run your judgment without you. Watch it, then read the step-by-step guide on becoming an AI engineer.

Y Combinator CEO, Garry Tan, took the stage for 42 minutes at Startup School 2026 and explained how to build your own personal AGI better than any paid AI course. This is what he told the room: 1. The leverage is in your context, not the model. Tan watches hundreds of founders use identical models every batch. "There are 2x people and there are 100x people who are using the same Claude. Same weights, same context window size, same API. But the leverage is not in the weights." The gap between users is now bigger than the gap between models. 2. One person's output went up 400x. In 2013 Tan shipped maybe 14 useful lines of code a day as a YC partner, dead on the median for programmer productivity. "I did the math on my output, and I'm at about 400x what I did in 2013." 3. Agents run on a different working memory. Humans hold 7 things in their head at once. Every org chart and checklist ever built is a patch for that limit. "An AI agent holds a million tokens. That's about a thousand pages. Three Harry Potter books sitting open on its head all at once." You're still running your week on tools built for the 7-digit brain. 4. Markdown is code now. Tan's stack is mostly skill files: pages of plain English an agent can execute. "If you can write clear instructions in English, you're a programmer. The compiler is a language model." At YC, finance and events staff who never opened a terminal are building automations. 5. Your history is your moat. Tan's agent runs on a personal wiki: about 220,000 markdown pages covering 25 years of email, meetings, notes and decisions. "When my agent does anything, it does knowing everything I know. And that's the difference between an assistant and a colleague." No frontier model has your context. That's the one asset nobody can replicate. 6. Never do one-off work. Most people run a task with an agent, close the window and throw the learning away. Tan ends every task by having the agent turn what it did into a reusable skill file. "If you have to ask for something twice, you failed." Captured skills compound daily. Amnesia resets you to zero every morning. 7. Own your skill files before your employer does. A skill file is your judgment, extracted and executable. The only question is who controls it. "Own your skills because if you don't, your job becomes a skill file." Files in your repo compound your career. Files in the company's repo run your judgment without you. Watch it, then read the step-by-step guide on becoming an AI engineer.

Alex Prompter

249,533 görüntüleme • 1 ay önce