Hermes Agent Ships Self-Evolution While Cloudflare Hands Everyone a Free Web Crawler

March 11, 2026 · 22 sources

NousResearch's Hermes agent dropped a weekend of updates including DSPy-powered self-evolution and autonomous Pokemon playing, while Cloudflare surprised everyone by launching a /crawl endpoint that reduces web scraping to a single API call. Meanwhile, the local AI movement gained momentum with Microsoft's BitNet running 100B models on CPUs and NVIDIA's Nemotron 3 Nano targeting low-end hardware.

Daily Wrap-Up

The AI world keeps compressing timelines in ways that feel genuinely disorienting. Today's feed was dominated by two stories that, taken together, paint a clear picture of where things are headed: agents are learning to improve themselves, and the infrastructure to feed them data just got radically simpler. NousResearch shipped a self-evolution system for their Hermes agent that uses DSPy to rewrite its own prompts and skills based on failures, essentially closing the loop on autonomous improvement. On the other side of the stack, Cloudflare dropped a /crawl endpoint that turns entire websites into structured data with a single API call, which is darkly funny coming from the company that built its business on stopping exactly this kind of thing.

The local inference story also had a strong day. Microsoft's BitNet framework, which runs 100B parameter models on CPUs using ternary weights, keeps making rounds. NVIDIA released Nemotron 3 Nano targeting machines with just 24GB of RAM. And people are running Qwen 3.5 9B through autonomous agents on RTX 3060s they bought to play Warzone. The gap between what you need cloud infrastructure for and what your laptop can handle keeps narrowing, and the agent ecosystem is racing to exploit that gap. Kimi K2.5 also emerged as a serious daily driver, with DHH and others praising its speed at 200 tokens per second through Fireworks AI.

The most practical takeaway for developers: if you're building agents, the "less tools, more discipline" principle that @morganlinton highlighted deserves serious attention. The rush to stuff agents with every available tool is producing worse results than carefully curating a minimal, well-optimized toolkit. Pair that with the shared-folder architecture pattern from @nityeshaga's Claude Code project manager writeup, where sub-agents write reports to a temp directory instead of flooding the orchestrator's context window, and you have two concrete patterns worth implementing this week.

Quick Hits

@pvncher celebrated the 1-year anniversary of RepoPrompt 1.0 with a demo combining agentic context building with ChatGPT's GPT-5.4 Pro for deep code reviews.

@EntireHQ launched their CLI for GitHub Copilot CLI users, recording "flight paths" of Copilot coding sessions.

@steipete shared that qmd 2.0 is out with a stable library interface for easier integration.

@badlogicgames retweeted updates about continuous /btw threads for Pi development.

@CorelythRun announced an AI assistant focused on task automation and tool syncing.

@thejayden recommended a Polymarket weather trading bot tutorial as a genuinely useful learning exercise amid prediction-market hype.

@PKodmad was impressed by an open-source skill for generating spec-conforming App Store screenshots, replacing paid SaaS tools.

Agents & Self-Improvement

The agent conversation took a meaningful turn this week with NousResearch shipping what might be the most ambitious open-source agent update in months. Their Hermes agent now includes a self-evolution system built on DSPy and GEPA that maintains populations of solutions, applies LLM-driven mutations targeting specific failure cases, and selects based on fitness. As @EXM7777 put it: "it's the best agent i've ever touched, not even close. It uses DSPy to rewrite its own skills and prompts based on failures... it rewrites its own code to get better over time." The system draws inspiration from Imbue's Darwinian Evolver research that achieved 95.1% on ARC-AGI-2, and the fact that it's open-source and runnable with a ChatGPT Pro subscription makes it accessible in a way that proprietary agent systems aren't.

But raw capability isn't everything. @morganlinton flagged what might be the most underappreciated insight in agent development right now: "the less tools you give them, the better they perform." This runs counter to the instinct most teams have when building agents, which is to connect everything and hope the model figures it out. The companies getting real value from agents are being "incredibly detailed, and disciplined about tool selection and optimization." Meanwhile, @neural_avb shared learnings on building eval harnesses for agentic systems, applying traditional ML experiment-loop principles to agent development, a reminder that rigorous evaluation infrastructure matters as much as the agents themselves.

The tension between agent capability and agent reliability is the defining challenge of this phase. Self-improving agents sound exciting until you realize that without proper evaluation frameworks, you're just automating the production of harder-to-debug failures.

Cloudflare's Crawling Pivot

Cloudflare did something that had the entire tech timeline doing double-takes. The company that sells anti-bot protection shipped a /crawl endpoint that lets anyone scrape entire websites with a single API call, returning clean HTML, Markdown, or JSON. @TukiFromKL captured the irony perfectly: "For years, Cloudflare sold anti-bot protection. Companies paid them to STOP crawlers. Now those same companies are watching Cloudflare hand everyone a free crawler that bypasses... other people's anti-bot protection. They didn't switch sides. They're playing both sides. And getting paid twice."

@Anubhavhing spelled out the implications for the scraping ecosystem: "Every web scraping startup that raised millions to solve this problem just became a single endpoint. Every freelancer charging $500 to 'extract website data' just lost their entire business model to a /crawl command." The hyperbole aside, there's a real structural shift here. Web scraping has been a cottage industry of Playwright scripts, proxy rotation, and rate-limiting logic. Reducing it to a platform feature changes who can access web data and how quickly they can do it. For agent builders in particular, this removes one of the more annoying infrastructure hurdles in building research and data-gathering agents.

Local AI Keeps Gaining Ground

The push to run serious models on consumer hardware had its strongest day in a while. Microsoft's BitNet framework, which uses ternary weights (just -1, 0, and +1) to run 100B parameter models on CPUs, continues to gain attention. @heygurisingh broke down the numbers: "2.37x to 6.17x faster than llama.cpp on x86, 82% lower energy consumption, memory drops by 16-32x vs full-precision models." With 27.4K GitHub stars and MIT licensing, BitNet represents a serious bet that the future of inference isn't just about bigger GPUs.

On the smaller end, NVIDIA's Nemotron 3 Nano targets machines with 24GB of RAM or VRAM, offering what @0x0SojalSec described as "sharp reasoning, amazing coding, and zero refusals via SOTA PRISM pipeline." And @sudoingX demonstrated the practical reality of budget local AI by running Qwen 3.5 9B through Hermes Agent on a single RTX 3060: "5.3 gigabytes of model on a card most people bought to play Warzone... plugged it into a full autonomous agent with 29 tools, terminal access, file operations, browser automation, persistent memory across sessions." The 3060's 12GB of VRAM, more than the 3070's 8GB, makes it what they called "the most underrated budget AI card on the market."

Claude Code for Non-Technical Work

@nityeshaga shared a detailed case study that deserves attention beyond the usual Claude Code discourse. They built "Claudie," an AI project manager that pulls from Gmail, Drive, Calendar, and meeting notes into real-time client dashboards, work that previously consumed a full person's time. The architectural journey is instructive: slash commands failed because MCP tool calls consumed too much context. Orchestrator-plus-sub-agents failed because sub-agent reports overwhelmed the main agent's context window.

The solution was elegantly simple: "We made every sub-agent output their final report into a temp folder and tell the orchestrator where to find it." This shared-folder pattern solved both context management and inter-agent communication. They also evolved from eleven fragmented skills to a single "handbook" skill organized into chapters, mirroring how you'd actually onboard a human employee. The framing of treating agent setup like hiring a real person, starting with a job description, then building an onboarding handbook, is a design pattern worth stealing.

Retrieval Reimagined for Agents

@AkariAsai introduced AgentIR, a research direction that challenges a fundamental assumption in how we build retrieval for AI agents. Most Deep Research agents still use search engines and embedding models designed for humans, retrieving based on a query plus maybe an instruction. AgentIR instead uses the agent's reasoning tokens during retrieval, training embedding models to leverage the rich context that agents generate while thinking through problems.

The results on BrowseComp-Plus tell the story: BM25 achieved 35%, Qwen3-Embed hit 50%, and AgentIR reached 67%. As Asai noted, "reasoning traces provide a much richer retrieval context than simply paraphrasing or concatenating queries. They capture evolving intent, decomposition, and search rationale in ways standard query-only retrieval misses." This feels like one of those ideas that seems obvious in retrospect but could meaningfully improve agent performance across the board.

The Kimi K2.5 Moment

Kimi K2.5 is quietly becoming the default model for a growing segment of developers who need speed and cost efficiency over peak intelligence. @0xSero reported running "5x 8 hour missions overnight on my Kimi $40 subscription" with all of them still running smoothly, each with 20+ sessions. The model is "often indistinguishable from Claude, about 5x cheaper, open weight." DHH called it his "daily driver for all the basic stuff," running at 200 tokens per second through Fireworks AI. The emerging pattern is clear: frontier models for hard problems, fast open-weight models for everything else, and the "everything else" category keeps expanding.

Google's Multimodal Embeddings

@googleaidevs announced Gemini Embedding 2, their first fully multimodal embedding model built on the Gemini architecture, now available in preview via the Gemini API and Vertex AI. While the announcement was light on benchmarks, the significance is in the "fully multimodal" qualifier. Embedding models that natively handle text, images, and other modalities without separate pipelines could simplify RAG architectures considerably. Combined with AgentIR's work on reasoning-aware retrieval, the embedding layer of the AI stack is getting substantially more capable.

Sources

Nityesh @nityeshaga · Jan 30

How We Built an AI Project Manager Using Claude Code Claude Code for non-technical work is going to sweep the world by storm in 2026. This is how we built Claudie, our internal project manager for the consulting business. This process provides a great peek into my role as an applied AI engineer. My Role I'm an applied AI engineer at @every. My job is to take everything we learn about AI — from client work, from the industry, from internal experiments — and turn it into systems that scale. Curriculum, automations, frameworks. I turn the insights clients give us on discovery calls to curriculum that designers can polish into final client-ready materials. When there's a repetitive task across sales, planning, or delivery, I build the automation, document it, and train the internal team to use. The highest-value internal automation I've built so far is the one I'm about to tell you about. What We Needed to Automate Every Consulting runs on Google Sheets. Every client gets a detailed dashboard — up to 12 tables per sheet — tracking people, teams, sessions, deliverables, feedback, and open items. Keeping these sheets accurate and up-to-date is genuinely a full person's job. @NataliaZarina, our consulting lead, was doing that job on top of 20 other things. She's managing client relationships, running sales, making final decisions on scope and delivery — and also manually updating dashboards, cross-referencing emails and calendar events, and keeping everything current. It was the work of two people, and she was doing both. So I automated the second person. Step 1: Write a Job Description The first thing I did was ask Natalia to write a job description. Not for an AI agent — for a human. I asked her to imagine she's hiring a project manager: what would she want this person to do, what qualities would they have, what would be an indicator of them succeeding in their role, and everything else you'd put in a real job description. See screenshot 1. Once I had this job description, I started thinking about how to turn it into an agent flow. That framing — treating it like hiring a real person — ended up guiding every architectural decision we made. More on that later. Step 0: Build the Tools Before any of the agent work could happen, we needed Claude Code to be able to access our Google Workspace. That's where the consulting business lives — Gmail, Calendar, Drive, Sheets. Google does not have an official MCP server for their Workspace tools. But here's something most people don't know: MCP is simply a wrapper on top of an API. If you have an API for something, you basically have an MCP for it. I used Claude Code's MCP Builder skill — I gave it the Google Workspace API and asked it to build me an MCP server, and it did. Once it was confirmed that Claude Code could work with Google Sheets, that was the biggest unknown resolved, and we knew it would be able to do the work we needed. Version 1: Slash Commands Now it was time for context engineering. The first thing we tried was to create a bunch of slash commands — simple instructions that tell Claude what to do for each piece of work. This treated slash commands as text expanders, which is what they are, but it didn't work. It failed for one critical reason: using MCP tools to read our data sources and populate our sheets was very expensive in terms of context. By the time the agent was able to read our data sources and understand what was needed, it would be out of context window. We all know what that does to quality — it just drops drastically. So that didn't work. Version 2: Orchestrator and Sub-Agents This is also exactly when Anthropic released the new Tasks feature. We decided the new architecture would work by having our main Claude be the orchestrator of sub-agents, creating tasks that each get worked on by one sub-agent. But this ran into another unexpected problem. The main Claude would have its context window overwhelmed when it started 10 or more sub-agents in parallel. Each sub-agent would return a detailed report of what they did, and having so many reports sent to the orchestrator at the same time would overwhelm its context window. For example, our very first tasks launch data investigation agents which look at our raw data sources and create a detailed report about what has happened with a client over a specific period of time, based on a particular source like Gmail or Calendar. The output of these sub-agents needs to be read by all the sub-agents down the line — up to 35 of them. There would definitely be a loss in signal if it was the job of the main orchestrator to pass all required information between sub-agents. The Fix: A Shared Folder So we made one little change. We made every sub-agent output their final report into a temp folder and tell the orchestrator where to find it. Now the main Claude reads reports as it sees fit, and every downstream sub-agent can read the reports from earlier phases directly. This totally solved the problem. And it also improved communication between sub-agents, because they could read each other's full output without the orchestrator having to summarize or relay anything. See screenshot 2. Version 3: From Skills to a Handbook With the orchestration working, I initially created separate skills for each specific piece of work — gather-gmail, gather-calendar, check-accuracy, check-formatting, and so on. Eleven skills in total. Each sub-agent would read the skill it needed and get all the context for its task. This worked, but it was ugly. These were very specific, narrow skills, and it created all sorts of fragility in the system. Not to mention that it was difficult for even the humans to read and maintain. That's when the job description framing came back around. We started by treating this like hiring a real person. We wrote them a job description. So what do you do once you've actually hired someone? You give them an onboarding handbook — a document that covers how you approach things on your team and tells them to use it to get the job done, all aspects of their job. So that's what we built. One single project management skill that contains our entire handbook, organized into chapters: • Foundation — who we are, the team, our tools and data sources, when to escalate, data accuracy standards • Daily Operations — how to gather data from all our sources • Client Dashboards — how the dashboards are structured, what the master dashboard tracks, how to run quality checks • New Clients — how to onboard a new client and set up their dashboard from scratch Now when a sub-agent spins up, it reads the foundation chapters first (just like a new hire would), then reads the chapters relevant to its specific task. The handbook replaced eleven fragmented skills with one coherent source of truth. Here's what the final architecture looks like: See screenshot 4. What This Felt Like This was the most exhilarating work I've done in two weeks, and it was all of the things at once. Working with @NataliaZarina was the most important part. We were on calls for hours, running Claude Code sessions on each of our computers and trading inputs. She has the taste — she knows what the dashboards should look like, what the data should contain, what quality means for our clients. I have the AI engineering. Working together on this was genuinely exciting. Then there's the speed. We went through three major architectural generations in a span of two weeks. Everything was changing so fast. And what was actually the most exciting was how hard we were driving Claude Code. I've been using Claude Code for programming for months, but I was not driving it this hard before. This last couple weeks, I was consistently running out of my usage limits. In fact, both Natalia and I were running out of our combined usage limits on the ultimate max plans on multiple days. When you're consuming that much AI inference, you can imagine how fast things are moving. And that was just exciting as fuck. This was also a completely novel problem. Applied AI engineering as a discipline is still new, and this was the first real big shift in how I think about it. Why Now, and Why 2026 Here's why I opened with the claim that Claude Code for non-technical work will sweep the world in 2026. We realized that if you give Claude Code access to the tools you use as a non-technical person and do the work to build a workflow that covers how you actually use those tools, that is all you need. That's how non-technical work works. The reason this hasn't been done until now is that we were running Claude Code at its limits. This would not have been possible with a previous version of the AI or a previous version of Claude Code. We're literally using the latest features and the latest model. It requires reasoning through and understanding of the underlying tools and how to operate them, along with planning capabilities and context management capabilities that did not exist even six months ago. But now they do. And we're only in January. Every piece of the stack that made this possible is brand new: • MCP Builder skill — I built our own Google Workspace MCP server by asking Claude Code to use the Google Workspace API. That was not possible before Anthropic released MCP Builder on Oct 16, 2025 • Opus 4.5 — Its reasoning and planning capabilities made the entire orchestration possible. The agent needs to understand complex sheet structures, figure out what data goes where, and coordinate across dozens of sub-agents. Released Nov 24, 2025. • The Tasks feature — Sub-agent orchestration through Tasks made Version 2 and 3 possible at all. This was released Jan 23, 2026. That's why I'm saying Claude Code for non-technical work will sweep 2026. The building blocks just arrived.

N NataliaZarina @NataliaZarina

We just brought on a new project manager for @every, an agent called Claudie. The job description is a https://t.co/HdKYJTtSdj file. It pulls from Gmail, Drive, Calendar & meeting notes into a real-time client dashboard. Custom skills & tasks enforce quality. @nityeshaga & I built it over the past 2 weeks because managing clients with hundreds of employees across multiple teams was drowning us in manual work. Setup for our latest client took Claudie 30 min & would've taken us 5 hrs. Welcome to the team!

CorelythRun @CorelythRun · Mar 7

An AI assistant built for real work. Automate tasks, sync your tools, and turn commands into actions.

Ihtesham Ali @ihtesham2005 · Mar 10

🚨 Alibaba just open sourced a GUI agent that lives inside your webpage and controls it with natural language. It's called Page Agent and it's not a browser extension. It's pure JavaScript no Python, no Puppeteer, no headless browser, no screenshots. Just one script tag and your web app understands natural language. Here's what it actually does: → Embed it with a single <script> tag or npm install → Control any web interface with plain English commands → Text-based DOM manipulation no OCR, no vision models needed → Bring your own LLM (GPT, Claude, Qwen, anything) → Ships a built-in UI with human-in-the-loop support → Turn 20-click ERP/CRM workflows into one sentence → Optional Chrome extension for multi-tab agent tasks → Works on any web app SaaS, admin panels, internal tools Companies are charging $30/month for AI copilots built on this exact idea. This is 3 lines of code. Your users. Your interface. The AI copilot layer for every web app just got open sourced. 1.6K stars. 100% Open Source. (Link in the comments)

PK 🐢 👩🏻‍💻 @PKodmad · Mar 10

Completely humbled. Did not know you could make a skill do that. People packaging this as Sass and paid apps. Here’s an open source free to use freaking skill and it made beautiful spec conforming screenshots for me.

P ParthJadhav8 @ParthJadhav8

Making app store screenshots sucks. made a SKILL that does it END TO END. It handles content, design, apple specific sizing. It even generates page to visualize screenshots. https://t.co/qM4DZOd3CJ Example below👇 https://t.co/j5jl0sStAJ

Akari Asai @AkariAsai · Mar 10

Many Deep Research agents still rely on search engines and embedding models built for humans, not agents. They retrieve from a query (+ maybe an instruction), but ignore the much richer context agents generate while reasoning. Retrieval should be redesigned for agents - introducing AgentIR. 💡Key Idea: Use the DR agent's reasoning tokens during retrieval and train embedding models to leverage these reasoning traces more effectively, rather than relying only on the final query. 📈Results: State-of-the-art performance on BrowseComp-Plus, outperforming powerful reasoning-aware embedding models (ReasonIR) or query re-writing methods. Interestingly, we find that reasoning traces provide a much richer retrieval context than simply paraphrasing or concatenating queries. They capture evolving intent, decomposition, and search rationale in ways standard query-only retrieval misses. Looking ahead, we believe future embedding (or retrieval systems in general) should be trained natively for Deep Research agents, so they can fully leverage these reasoning chains.

Z zijian42chen @zijian42chen

🚀 Introducing AgentIR, a retriever that reads your agent’s mind (literally!) 🧠 Unlike humans, agents explicitly expose thoughts in reasoning tokens. Put them to use! 📈 Simple, substantial gains for agents on BrowseComp-Plus, 35% (BM25) ➡️ 50% (Qwen3-Embed) ➡️ 67% (AgentIR) 🧵

Morgan @morganlinton · Mar 10

I can't stop thinking about this article Rhys wrote yesterday, I think he's really onto something. While there are a lot of really good points he makes, the one that really stands out to me, that I think is a real (growing) problem today is: "the less tools you give them, the better they perform" As more and more companies, rush to adopt AI Agents, I think one of the key mistakes they'll make/are making, is to stuff those agents full of tools, and then wonder why they're not performing as well. The companies that end up really getting value from agents, are going to be those that are incredibly detailed, and disciplined about tool selection and optimization.

R RhysSullivan @RhysSullivan

The Execution Layer

Entire @EntireHQ · Mar 10

Copilot flies the plane. Entire records the flight path. The Entire CLI is now available for @github Copilot CLI users. Who reaches the first Checkpoint? 🤖https://t.co/TH1DsEfAEi https://t.co/Vo7UEb9XQE

Machina @EXM7777 · Mar 10

do you understand what this means for agents? i've been locked in on Hermes for a few days now... and it's the best agent i've ever touched, not even close >it uses DSPy to rewrite its own skills and prompts based on failures >it plays Pokemon Red autonomously from your terminal >it rewrites its own code to get better over time >it can now uncensor any open-weight model in one command >it does everything OpenClaw does and 100x more plus you can use your ChatGPT Pro subscription (basically unlimited usage) BUILD MORE

N NousResearch @NousResearch

The last few days have been wild. Here's what we've shipped over the weekend. But first, we're giving away free Nous Portal subscriptions to the first 250 people who claim code AGENTHERMES01 at https://t.co/uomohWX1dw - and there's a lot of exciting new stuff to use it on: -> Pokemon Player 🎮 Hermes can now play Pokemon Red/FireRed autonomously via headless emulation. The new pokemon-agent package (https://t.co/SefIfSFMBP) and built-in skill provides a REST API game server, and Hermes drives it through its native tools - reading game state from RAM, making strategic battle decisions, navigating the overworld, and saving progress to memory across sessions. It just plays Pokemon. From your terminal. No display server needed. -> Self-Evolution 🧬 We shipped hermes-agent-self-evolution (https://t.co/9xlrFXTgPQ) and an optional skill - an evolutionary self-improvement system that uses DSPy + GEPA to optimize Hermes's own skills, prompts, and code. It maintains populations of solutions, applies LLM-driven mutations targeted at specific failure cases, and selects based on fitness. Inspired by Imbue's Darwinian Evolver research that achieved 95.1% on ARC-AGI-2. -> OBLITERATUS 🔓 The abliteration skill got a major update. Hermes can now uncensor any open-weight LLM (Llama, Qwen, Mistral, etc.) by surgically removing refusal directions from model weights - 9 CLI methods, 116 model presets, tournament evaluation. Just say "abliterate this model" and it handles the rest. -> Signal, iMessage + 7-Platform Gateway 📱 Hermes now runs on iMessage and Signal alongside Telegram, Discord, WhatsApp, Slack, and CLI. Full feature parity: voice messages, image handling, DM pairing. Your agent is reachable everywhere. -> Automatic Provider Failover 🔄 When your primary model goes down (rate limits, outages), Hermes now automatically switches to a configured fallback model. Supports all providers including Codex OAuth and Nous Portal. One line of config, zero downtime. -> Secret Redaction Everywhere 🔒 All tool outputs now redact API keys, tokens, and passwords before they reach the LLM. 22+ patterns covering AWS, Stripe, HuggingFace, GitHub, SSH private keys, database connection strings, and more. Your secrets never leak into context.

Google AI Developers @googleaidevs · Mar 10

Start building with Gemini Embedding 2, our most capable and first fully multimodal embedding model built on the Gemini architecture. Now available in preview via the Gemini API and in Vertex AI. https://t.co/jPE8KpN7Rf

Md Ismail Šojal 🕷️ @0x0SojalSec · Mar 10

Local LLM that runs on a low-end GPU or even fully on CPU and RAM.🫠 Yes, NVIDIA Nemotron 3 Nano! Speed demon with sharp reasoning, amazing coding, and zero refusals via SOTA PRISM pipeline. 24GB RAM/VRAM, and you can now fine-tune it locally. - https://t.co/MkshkuRd6f https://t.co/Pkdy61ouBN

AVB @neural_avb · Mar 10

Wrote down my learnings about setting up eval harnesses for agentic systems. I come from ML, so the build <-> experiment loop is kinda second nature to me. I have basically tried to apply those same principles here.

N neural_avb @neural_avb

How to design Experiments to Evaluate your Agentic Harness

eric provencher @pvncher · Mar 10

Today is the 1 year anniversary of @RepoPrompt 1.0 To highlight the occasion, here's a video combining agentic context building, with old school copy pasting, to enable you to use ChatGPT's GPT-5.4 Pro to get some of the deepest code reviews around. https://t.co/NqFtbnMnnx

Numman Ali @nummanali · Mar 10

Claude Code teams with tmux is really cool When you run with team mode enabled in tmux, it automatically opens the additional terminal in pane I don't really get my main agent to orchestrate, I chat to them myself CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=true claude https://t.co/GRBKlA4XRd

Peter Steinberger 🦞 @steipete · Mar 11

RT @tobi: qmd 2.0 is out. You can now use it in as a library with stable interface. Much easier to integrate. And all the other interfaces…

Tuki @TukiFromKL · Mar 11

🚨 Stop scrolling. This is the biggest betrayal in tech this year. The company that built its entire reputation on BLOCKING scrapers just shipped the most powerful scraping tool ever made. > Cloudflare just dropped a /crawl endpoint. One API call and you get an entire website back. Clean HTML, Markdown, or JSON. That's it. That's the whole thing. Let me break it down. For years, Cloudflare sold anti-bot protection. Companies paid them to STOP crawlers. Now those same companies are watching Cloudflare hand everyone a free crawler that bypasses… other people's anti-bot protection. They didn't switch sides. They're playing both sides. And getting paid twice.

C CloudflareDev @CloudflareDev

Introducing the new /crawl endpoint - one API call and an entire site crawled. No scripts. No browser management. Just the content in HTML, Markdown, or JSON. https://t.co/DCJMek8cZa

Andrej Karpathy @karpathy · Mar 11

@nummanali tmux grids are awesome, but i feel a need to have a proper "agent command center" IDE for teams of them, which I could maximize per monitor. E.g. I want to see/hide toggle them, see if any are idle, pop open related tools (e.g. terminal), stats (usage), etc.

Guri Singh @heygurisingh · Mar 11

Holy shit... Microsoft open sourced an inference framework that runs a 100B parameter LLM on a single CPU. It's called BitNet. And it does what was supposed to be impossible. No GPU. No cloud. No $10K hardware setup. Just your laptop running a 100-billion parameter model at human reading speed. Here's how it works: Every other LLM stores weights in 32-bit or 16-bit floats. BitNet uses 1.58 bits. Weights are ternary just -1, 0, or +1. That's it. No floats. No expensive matrix math. Pure integer operations your CPU was already built for. The result: - 100B model runs on a single CPU at 5-7 tokens/second - 2.37x to 6.17x faster than llama.cpp on x86 - 82% lower energy consumption on x86 CPUs - 1.37x to 5.07x speedup on ARM (your MacBook) - Memory drops by 16-32x vs full-precision models The wildest part: Accuracy barely moves. BitNet b1.58 2B4T their flagship model was trained on 4 trillion tokens and benchmarks competitively against full-precision models of the same size. The quantization isn't destroying quality. It's just removing the bloat. What this actually means: - Run AI completely offline. Your data never leaves your machine - Deploy LLMs on phones, IoT devices, edge hardware - No more cloud API bills for inference - AI in regions with no reliable internet The model supports ARM and x86. Works on your MacBook, your Linux box, your Windows machine. 27.4K GitHub stars. 2.2K forks. Built by Microsoft Research. 100% Open Source. MIT License.

Mario Zechner @badlogicgames · Mar 11

RT @noahsaso: /btw for pi 🐇 worth it https://t.co/H0cuYbzBpM https://t.co/X361s6tq5q full continuous /btw threads with the ability to /bt…

Anubhav @Anubhavhing · Mar 11

Crawling an entire website used to take: A Python script. Playwright or Selenium. Proxy rotation. Rate limiting logic. Error handling. 3 hours of debugging why page 47 returned a 403. Now it's one API call. Every web scraping startup that raised millions to solve this problem just became a single endpoint. Every freelancer charging $500 to "extract website data" just lost their entire business model to a /crawl command. HTML. Markdown. JSON. Pick your format. No scripts. No browser. No headache. The entire web scraping industry just got reduced to one line of code. Someone is going to use this to clone every competitor's website by Friday. 💀

C CloudflareDev @CloudflareDev

Introducing the new /crawl endpoint - one API call and an entire site crawled. No scripts. No browser management. Just the content in HTML, Markdown, or JSON. https://t.co/DCJMek8cZa

0xSero @0xSero · Mar 11

I ran 5x 8 hour missions overnight on my Kimi 40$ subscription. I woke up to all of them running smoothly still, each has 20+ sessions. This model is often indistinguishable from Claude, about 5x cheaper, open weight, and their Kimi code has to be incentivized I never run out

D dhh @dhh

Kimi K2.5 continues to be my daily driver for all the basic stuff where I don't need PhD-level intelligence. I just need it done quickly. Running it at 200 tps through @FireworksAI_HQ within @opencode is just such a delight. https://t.co/ukQIG7DmfG

Sudo su @sudoingX · Mar 11

single RTX 3060. 12GB VRAM. AMD EPYC 7282. Qwen 3.5 9B Q4_K_M running through Hermes Agent. 5.3 gigabytes of model on a card most people bought to play Warzone. compiled llama.cpp from source. loaded a 9 billion parameter model that outscores models 13x its size on reasoning benchmarks. plugged it into a full autonomous agent with 29 tools, terminal access, file operations, browser automation, persistent memory across sessions. will run the same octopus invaders prompt i've been testing across every config. Qwen 3.5 27B dense on a 3090. the 35B MoE. the 80B coder. hermes 4.3 36B. now the 9B on a 3060. same test. same standard. different floor. the 3060 has more VRAM than the 3070. 12GB vs 8GB. the most underrated budget AI card on the market and most people don't know it.

S sudoingX @sudoingX

qwen just dropped 4 new models. 0.8B runs on a phone. 9B runs on 6GB RAM. same Qwen3.5 family, same 256K context, all the way down. i just finished benchmarking the 35B MoE across 15 GPUs and published the full breakdown. now the entire family is here. 0.8B, 2B, 4B, 9B. community has been asking for smaller GPU benchmarks since the first thread. running the full lineup. every size, same benchmarks. scaling nodes as needed. small to big comparison data coming.

Jayden ⛩️ @thejayden · Mar 11

Between all the unlimited prediction-market slop posts, this article is actually quite good. You can copy-paste it into GPT and ask it to explain it like you’re five, then ask it to build a single-file script. It won’t make you rich, but you’ll learn new things.

A AlterEgo_eth @AlterEgo_eth

Building a Polymarket Weather Trading Bot in Python (Step-by-Step Guide)