Lots of world-models papers exist today but there is almost no map of how they connect 🧬 @ylecun said, “The world model is a kind of simulator of the relevant aspects of the world. It predicts future world states from observations, prior state, and action.” So I built Engram, a second brain for world-models research: → Timeline: 57 papers across 8 paradigms distributed by year of release → Neuron: force-directed clusters across the paradigms inspired by Karpathy's version of second brain. → Random: land somewhere unexpected and get started on learning something new today → Every paper has arXiv and github links baked into it. There is also @askalphaxiv link attached to each paper so you can learn better and use their incredible features. PlaNet → Dreamer → DreamerV3. MuZero → EfficientZero. JEPA, Genie, GAIA and much much more. 🌐 ⭐ Launching today! Please try it and tell me what’s broken, missing, or inaccurate. Have fun with it!!

I've spent this year on computer use agents learning to operate a world made of pixels. Skild's S1 learns to operate a world made of atoms. A 10-minute task it has never seen. No fine-tuning, and no pre-written programs. The model outputs the motion directly. AI has finally transitioned from pixels to atom.

I've spent this year on computer use agents learning to operate a world made of pixels. Skild's S1 learns to operate a world made of atoms. A 10-minute task it has never seen. No fine-tuning, and no pre-written programs. The model outputs the motion directly. AI has finally transitioned from pixels to atom.

I built a second brain to cut my token bill. It didn’t. A 2026 study found LLM-written context files cost 20% more, with no measurable gain. Your agent can already read the repo. It doesn’t know what you decided, what failed, or what must not change. That’s what I carry between harnesses. No harness has to learn it twice.

I built a second brain to cut my token bill. It didn’t. A 2026 study found LLM-written context files cost 20% more, with no measurable gain. Your agent can already read the repo. It doesn’t know what you decided, what failed, or what must not change. That’s what I carry between harnesses. No harness has to learn it twice.

Grok Bot shouldn't write the code. It should run the team that does. I gave mine Claude Code, Codex, OpenCode, and pi. You define the outcome. It decides who plans, who builds, who reviews, and when you need to step in. One interface. An army of coding agents. Let it route the work. Never let it rewrite what has already been decided.

Grok Bot shouldn't write the code. It should run the team that does. I gave mine Claude Code, Codex, OpenCode, and pi. You define the outcome. It decides who plans, who builds, who reviews, and when you need to step in. One interface. An army of coding agents. Let it route the work. Never let it rewrite what has already been decided.

I have five coding harnesses installed. Cursor, Claude Code, Codex, OpenCode, Prime Agent. I still can't tell you which one deserves which task. Claude Code will spin up a swarm of sub-agents to rename a variable. pi will bring four tools to do a job that needed the swarm. So I stopped choosing. One setup, and every prompt lands on whichever one is actually best at it

this might be one of the most useful things I've read about AI coding agents all week everyone is building bigger and bigger memory systems but the real problem isn't that agents forget everything it's that every time you switch agents, the next one has to figure out what the previous one already learned this article explains why giving every coding agent the same second brain might be a much better idea

I have five coding harnesses installed. Cursor, Claude Code, Codex, OpenCode, Prime Agent. I still can't tell you which one deserves which task. Claude Code will spin up a swarm of sub-agents to rename a variable. pi will bring four tools to do a job that needed the swarm. So I stopped choosing. One setup, and every prompt lands on whichever one is actually best at it

grok bot + grok 4.6 can literally build any idea from 0 to 100 [here is the exact grok bot setup that does that] 1. extract all your claude code and codex sessions to your chief of staff. it will help structure the team 2. give it an auth.md for credentials and paste the video below to make the entire team

Your Claude hits the session limit. You switch to Codex. And you paste three hours of context into a blank window like none of it happened. You didn't lose the chat. You lost the decisions. I've paid that tax all month. So I wrote down the architecture that ends it.

Your Claude hits the session limit. You switch to Codex. And you paste three hours of context into a blank window like none of it happened. You didn't lose the chat. You lost the decisions. I've paid that tax all month. So I wrote down the architecture that ends it.

KV-Cache hits missing is the biggest TRADE-OFFS with these tools !! TLDR: Explained in the tldraw workflow image attached !! As soon as I saw this article, I went onto see if kv-cache has been talked about or not and apparently it was. It cites a paper which reads "KV-cache transfer in LLM families" - But if you would have given this paper a read anytime, it is actually not transfer but just REPRESENTATION TRANSFER" - There are several more studies on this but there is no solution to this yet, anytime you switch a model mid-task, KV cache will de optimise and you won't be getting the same results until you spend an enormous amount of tokens again to develop or tune the same context for the model and even after that some representations will differ and go off the birdge. - I have been working and researching in this space from few months now and this is the most interesting work I have been doing tbh. -As per what I have studied and found out with my study and research, SSMs are a great solution to this compared to LLMs. Simple and Concise explanation- - Instead of transferring KV cache state from: K_A, V_A to K_B, V_B for 2 different models, we will just have to transfer the nth state from SSMs. -And hence the state-n becomes the 1st state of the new model which has been switched. There is a lot more underlying concept and research to this and very less resource is available. Happy to have a conversation with people involved in KV CACHE and its transfer and topics around the similar corner.

AGENT EVALS is going to be a big thing and it’s a big white space to work on !! I have been working with agents from the past 6-8 months now and I have been shipping it in production pipelines as well. The thing about agent evals o have found out is - They don’t exist !! Yes, there is no proper eval that can be used in all the production grade agents or pipelines. Almost every big company has a blog on this - -Open Ai -Anthropic -Langchain -Langfuse -Data bricks -IBM -Hugging face articles So many repos as well but all are scope bounded. 2 days back I had to score few checkpoints in my agentic harness. How would I ? Simply prepared 6-8 mathematical scores and combined them with some ground truth and that became my agent eval for judgement. Now from the last 2 days any new model or pipeline or experimentation I do over those checkpoints, I can instantly scrape them on and off just via the scores. Simple and easy. Learn to build custom agent evals and you have a long way to go, the need for the same is gonna get even bigger !! For resources check this research paper or research book as you may call it: > GENERAL AGENT EVALUATION

RIP to every "which framework should I use" thread. harness on top, framework in the middle, runtime at the floor. one stack, three heights, fully composable. paste these four images into claude and ask which layer your agent belongs on.