Hussein Lezzaik

Built clawsweeper, which runs 50 codex in parallel around the clock, scans issues/prs deep and closes what is already implemented or what makes no sense. Closed around 4000 issues today, a few thousand are in the pipeline. (rate limits are rough) github.com/openclaw/claws…

1/4 LLMs solve research grade math problems but struggle with basic calculations. We bridge this gap by turning them to computers. We built a computer INSIDE a transformer that can run programs for millions of steps in seconds solving even the hardest Sudokus with 100% accuracy

Today, we're launching the world's largest open-source dataset of computer-use recordings. 10,000+ hours across Salesforce, Blender, Photoshop and more, to automate the next level of white-collar work. Link in the comments :) @markov__ai

Computer use models shouldn't learn from screenshots. We built a new foundation model that learns from video like humans do. FDM-1 can construct a gear in Blender, find software bugs, and even drive a real car through San Francisco using arrow keys.

Introducing OneContext. I built it for myself but now I can’t work without it, so it felt wrong not to share. OneContext is an Agent Self-Managed Context Layer across different sessions, devices, and coding agents (Codex / Claude Code). How it works: 1. Open Claude Code/Codex inside OneContext as usual, it automatically manages your context and history into a persistent context layer. 2. Start a new agent under the same context, it remembers everything about your project. 3. Share the context via link, anyone can continue building on the exact same shared context. Install with: npm i -g onecontext-ai And open with: onecontext Give it a try!

@mehul @maticrobots congrats @mehul well deserved!

Announcing $60M for @maticrobots. We didn't ask "What's the most impressive robot we can demo?" We asked "What's the most useful robot we can ship? What comes after Roomba?" Customers answered with their wallets: It's Matic.

🥝 Meet Kimi K2.5, Open-Source Visual Agentic Intelligence. 🔹 Global SOTA on Agentic Benchmarks: HLE full set (50.2%), BrowseComp (74.9%) 🔹 Open-source SOTA on Vision and Coding: MMMU Pro (78.5%), VideoMMMU (86.6%), SWE-bench Verified (76.8%) 🔹 Code with Taste: turn chats, images & videos into aesthetic websites with expressive motion. 🔹 Agent Swarm (Beta): self-directed agents working in parallel, at scale. Up to 100 sub-agents, 1,500 tool calls, 4.5× faster compared with single-agent setup. - 🥝 K2.5 is now live on kimi.com in chat mode and agent mode. 🥝 K2.5 Agent Swarm in beta for high-tier users. 🥝 For production-grade coding, you can pair K2.5 with Kimi Code: kimi.com/code - 🔗 API: platform.moonshot.ai 🔗 Tech blog: kimi.com/blogs/kimi-k2-… 🔗 Weights & code: huggingface.co/moonshotai/Kim…

My implementation of the Recursive Language Model (RLM) paper by @a1zhang , Kraska, and @lateinteraction . Key insight: "Treat long context as an external environment, not something to stuff into a context window." Applied to video understanding — instead of encoding 38K frames into a prompt, the agent: → Treats video as an environment → Writes code to explore segments → Uses recursive LLM sub-calls for analysis Tested: 20+ min video, 7 steps, $0.002 Paper: arxiv.org/abs/2512.24601 Code: github.com/mohammed840/RL…

NEO’s Starting to Learn on Its Own

casually driving around on google maps this is also opensource and can be found on my github

@bcherny why does this happen and how do i fix it? @bcherny

run this: /mobile

this took around 30mins of training only

for more info check: - blog: husseinlezzaik.com/tess/flow-matc… - model: huggingface.co/TESS-Computer/… - code: github.com/HusseinLezzaik… - data: huggingface.co/datasets/TESS-…

the model produces trajectories in action chunks of detlta_x/delta_y

@trq212 now that CC codes autonomously — what are the odds that a hacker group gets CC to install a key-logger without it knowing it did?

If you started using Claude Code over the holidays, you might be curious about how AI actually works, the benefits and risks, and where it's headed. Here are some of my favorite papers on alignment, interpretability, and societal impacts 🧵

@modal serverless infra is a perfect fit for Claude Code. I used to always use Lambda Labs, but given how easy it is to give CC access to a GPU, monitor training runs, benchmarks, run certain inference tests, push code fixes .. all on demand & in the background it's effortless.

Much like the switch in 2025 from language models to reasoning models, we think 2026 will be all about the switch to Recursive Language Models (RLMs). It turns out that models can be far more powerful if you allow them to treat *their own prompts* as an object in an external environment, which they understand and manipulate by writing code that invokes LLMs! Our full paper on RLMs is now available—with much more expansive experiments compared to our initial blogpost from October 2025! arxiv.org/pdf/2512.24601

None of the SoTa Computer use models can output a smooth high-frequency cursor path in order to draw a circle. I trained a small DiT action head + Qwen2.5-VL on 10k samples to generate continuous point trajectories using action chunking and flow matching for 30mins on an H100.

OpenAI defines 5 stages of intelligence w/ level 5 AI running organizations. The best economically human institutions all run on custom UIs with limited API support. Therefore agents wrapping 10k APIs will always be at a disadvantage and limited compared to ones that could.

Figure's Helix VLA to click screens uses the same VLM backbone of computer use models to control it's hands. General computer use is a robotics problem, we need better digital hands!