This tool is blowing up on GitHub right now OpenClaude is Claude Code rebuilt to run on any provider. Same terminal, same tools, same subagents, except every agent can sit on a different model. Split by what the work actually needs, not by which model you like. > Code and terminal work: Claude or Codex. Both are built around long tool chains, and that is most of what an agent does. > Bulk file operations and context gathering: whatever is cheapest and fastest, including a local model on your own machine. Hundreds of reads, no judgment involved. > Long documents and huge context: Gemini. That is where the million-token window earns its keep. > Review: something from a different lab than the one that wrote the code. The point is a second opinion. > Anything you do at volume: an open model on your hardware. It costs nothing per run once it is set up. That is the whole trick: you stop paying flagship prices for work a cheap model could do. You can also send jobs to the background. Start one, close the terminal, check the log later, kill it if it goes wrong. And it maps your repo so agents stop burning turns working out where anything lives. Repo: Gitlawb/openclaude

THIS TOOL IS FREAKING AWESOME Zoetrope draws your Claude Code session as a live graph: > the main agent > every subagent it spawned > tools each one is running right now. Cards show status, current tool, call count and tokens spent. Edges animate while an agent works. You can drag the timeline backwards and the session rewinds. Agents un-finish, tool counts drop, the graph shrinks. Click any agent for the prompt that spawned it, its model, and every tool call with timings. Runs in your terminal, or in the browser with no install at all. Just drop a transcript on the page. Useful the first time you wonder what your agent has been doing for the last twenty minutes. Repo is furkankly/zoetrope

Everyone is talking about a second brain. Almost nobody says what it is actually for It is not a note app. It is what stops you researching the same thing twice. The part nobody stores, and you can start today: keep a rejected file. Every claim you checked and threw out, with the reason. A vendor number that appeared nowhere outside their own blog. A stat that turned out to be from an older release. A repo that looked good until you read the licence. It works the same outside research. Candidates you passed on and why. Tools your team trialled and dropped. Features you decided not to build. Suppliers you quoted and rejected. Everyone saves what they found. Almost nobody saves what they ruled out, so six months later somebody rules it out again from scratch. That reasoning lives in one person's head and nowhere else. I keep mine in HippocampOS. Facts with the source attached, and a separate brain per role so nothing bleeds between them. Asked one brain about a topic I never gave it, and it came back empty rather than inventing something. Graph below is a 15 minutes of loading. Add HippocampOS to your work at:

The Claude Code ecosystem you did not know existed 18 repositories, sorted by what they do: Official Anthropic (173k stars) - the official skill library. PDF, DOCX, XLSX, PPTX manipulation and art generation, straight from the source. (24k stars) - plugins aimed at knowledge workers rather than engineers, for Claude Cowork and Claude Code. (6k stars) - a GitHub Action that reviews every diff for vulnerabilities. (8k stars) - the official Python SDK for building on top of Claude Code. Agent Harness (280k stars) - a 7-phase workflow from brainstorm to review. It deletes implementation written before a failing test exists. In Anthropic's official marketplace since January. (245k stars) - 68 subagents, 286 skills, 94 commands, MIT. Built by an Anthropic hackathon winner. (130k stars) - Garry Tan's own setup, with opinionated agent roles for product, design, engineering, release and QA. (70k stars) - orchestration platform for running many agents against one codebase. Skills (46k stars) - 50 marketing skills. One writes down who your customer is, the other 49 read that file first. (16k stars) - grades anything Claude writes out of 50 across five dimensions and rewrites below 35. (123k stars) - 79 design styles, 192 palettes, and a list of what to avoid per industry. (18k stars) - 15 skills for keeping the context window small. Loads only what the task needs. (4.5k stars) - from the Remotion team. Describe a video, get React code that renders it. MCP (93k stars) - the catalogue everyone starts from when wiring Claude to something external. (6k stars) - 78 tools that put Claude Code inside your TradingView chart. It writes Pine Script and fixes its own compile errors. (2.5k stars) - brings MCP tools into ChatGPT, Grok, Perplexity and Gemini. Lists (53k stars) - the hand-curated list. Contributions get reviewed, bulk AI submissions get rejected. (1.3k stars) - the opposite approach. An automated index that has 1061 plugin repositories catalogued as of August 31.

Karpathy left an agent running overnight on his own training code, the code he had already spent a long time optimizing himself. It came back with about 20 improvements he had missed. The setup is almost insultingly small. One file the agent is allowed to edit, one it cannot touch, one number to beat. The agent changes something, trains for five minutes, checks the score. Better, it keeps the commit. Worse, it throws the change away and tries again. That is the entire loop. Two days, roughly 700 experiments. Time to GPT-2 quality went from 2.02 hours down to 1.80. It also caught a missing multiplier in his attention code that he had walked past. The part worth sitting with: you never edit Python here. You edit a markdown file that tells the agent how to do research. Karpathy says that file is the real benchmark now. Repo is karpathy/autoresearch.

This Stanford lecture is a must-watch for anyone who wants to understand how neural networks like ChatGPT are built and how they actually work 104 minutes from Yann Dubois, a Stanford researcher who helped build the tools the field uses to benchmark models like ChatGPT. He takes you from raw internet text all the way to a working ChatGPT-style assistant, and along the way the "magic" quietly falls apart. By the end you realize what's actually happening underneath is almost nothing like what most people assume. Full lecture below.

Palantir sells this to governments for millions. Someone put a version of it on GitHub for free. Palantir's job is to pull scattered company data into one model where every decision can be traced later. Semantica does that for AI agents. Each decision becomes an object in a graph, tied to the facts that produced it, each fact linked to its source. Six months later you can walk backwards from any answer. The model never touches the reasoning. It only reads messy text and pulls out entities. The rest runs on plain rules, so the same input gives the same explanation every time. Vector memory cannot do that. Contradicting facts overwrite each other silently. pip install semantica, pick a graph backend you already run, and your agents read it through the MCP server. MIT licensed, repo is semantica-agi/semantica

xAI ran a 12 hour AI hackathon and here is what won. Not one of them built an app. What they built instead is not what you think AI is for. Builders got a room in San Francisco, the latest Grok models, and one day. 1st, Nova. Old software ships as a file the computer can read and a person cannot. Nova hands that file to Grok and gets back source code you can actually read. They started with Game Boy games, then pointed it at the computer that runs a 1995 car engine, which nobody has the source code for. To check the answer they ran the original game and the rebuilt one at the same time and compared what happened on screen, so a wrong guess showed up straight away. 2nd, Signal. It reads your followers, sorts them into groups, then tests your post on a fake copy of your audience before you publish anything. 3rd, ThinkVoice. A headset reads faint signals from your head. Grok follows the conversation, offers what you would probably say next, and you pick it with a tiny movement. You never open your mouth. Nobody built a chatbot. Nova took three tries: first run broken, second run broken, third run everything worked. So stop building the landing page and the dashboard the course sellers keep pointing you at. Point the model at something nobody has tried yet. Cursor and ElevenLabs both started as somebody's odd side experiment.

Anthropic pays $3,850 a week to people with no ML experience, and over 50% of the first cohort ended up working there It's called the Anthropic Fellows Program. Four months, paid weekly, with compute funding on top. An Anthropic researcher pitches you project ideas, you pick one and shape it together, and you spend four months on real empirical AI safety research. Over 80% of the first cohort published a paper out of it. The barrier is almost nothing. No PhD, no prior ML experience, no published papers. They want Python, comfort with vague problems, and the ability to learn fast. Six tracks are open right now, and it's not US only. London, Ontario, San Francisco, and remote-friendly across the States. Applying is free and costs you an afternoon. Four months later you could be inside the company everyone else is reading about. Apply now:

Karpathy's LLM wiki concept is good, but it is not perfect. There is one thing missing from it Everyone built the filing cabinet. Documents go in, the model reads across all of them, answers get better. That part is real. The missing thing is action. Nothing in that folder ever moves until you sit down and ask it to. The harness is the other half. It is the rules that let something act on that cabinet while you are asleep: what it may open, what counts as finished, what it writes back, and who checks it. What the pair looks like: the notes folder holds what you know. Four more files hold what is being done. A spec, a plan, a progress log, and a list of decisions already made. The assistant reads all of it before it starts and updates it when it stops. What that gets you: close the laptop mid-task, open it tomorrow, and it carries on from where it stopped instead of asking you to explain again. The one rule: if it is not in a file, it does not exist. Full breakdown in the article below

Boris Cherny manages tens of thousands of agents everyday. Eight months without writing a line of code, and even the prompts He created Claude Code. Fortune asked what his job looks like now. You run one Claude Code and it has subagents that are other Claudes. His words, the user is no longer prompting Claude, another Claude does the prompting. He wakes up and Claude has already acted on ideas it picked by reading GitHub or X or anything else. Claude Code writes itself, writes its tests, runs its own security review, proposes its own improvements. 8x more code written at Anthropic since January. Asked about recursive self-improvement, he called it one of the big risks for AI. The person who built it said that. What he is describing is a job where you build the harness the agents run inside, then read what comes out. Everything upstream of that went to another Claude. Full breakdown in the article below

Figure AI is worth $39 billion. This is what happens to a company that solves one of industrial society's oldest problems Factories have had robots since the 1960s. They are bolted to the floor and repeat exactly one motion forever. Anything that needed a body that walks, reaches and adapts stayed human, because no machine could be trusted to do it unsupervised. Figure put one humanoid on the BMW line in Spartanburg for eleven months. It took sheet metal off a rack and placed it on a welding fixture. 1,250 hours of runtime, over 99% success per shift, zero human interventions, every part within 5 millimeters in 2 seconds. 30,000 cars came off that line. That is what the valuation is priced on. One task, held to a tolerance, for a year, with nobody watching. Robotics startups raised $18.8 billion in the first half of 2026, against $15 billion for all of last year.

THIS GUY RUNS AI MODELS ON HIS DESK THAT USED TO NEED A SERVER ROOM AND A CLOUD BUDGET A year of cloud invoices and he realized the truth. He wasn't building an AI asset. He was financing someone else's data center. So he bought his own What it is: The NVIDIA DGX Spark. A desktop AI machine with 128GB of unified memory. Runs large open models that normally only load inside rented cloud. No rack, no server room, no cooling Why you need this: > Big open models are finally good enough for real work > Owning the hardware to run them just got cheap > Every run is electricity, not another invoice > Your data never leaves the room > Owners move faster than everyone stuck in a cloud queue How you use it: Your stack works out of the box. Ollama, PyTorch, Hugging Face, vLLM, llama.cpp. Point it and build. Migration is minutes. What you can build with it: > Run large models locally with no usage cap > Fine-tune on your own private data > Leave agents running overnight for free > Serve client AI work without their data leaving your machine > Prototype and ship AI products with zero cloud bill It won't beat a top GPU on small models, and production still belongs in the cloud. The point is owning your experiments instead of renting them Bookmark this