Building AI since 2004, before it was cool. ML researcher and engineer. Signal over hype on models, agents and AI x biotech.
Joined March 2026
- Tweets428
- Following737
- Followers49
- Likes9.3K
n8n just released Nod, an open source agent with its own computer and browser. Cancelling my subscriptions with it sounds nice lol community.n8n.io/t/nod-an-op…
15
Alibaba open sourced the agent core from Qoder inside Qwen Code 1.0. Anyone can embed it now. Nice! github.com/QwenLM/qwen-code
1
39
Microsoft changed under 10 neurons and the agent still passed all the evals. Sleeper agents are real now. Wild research microsoft.com/en-us/research…
18
MBZUAI released PAN 2, an open world model. You give it a scene and an action and it shows what happens next. Apache 2.0 too, nice! ifm.ai/pan2/press-release/
1
34
Anthropic says Claude Fable 5.2 is 30% faster and 20% cheaper than 5.1. Better and cheaper at once is the kind of upgrade developers quietly love. anthropic.com/claude-fable-a…
63
Meituan opened LongCat-2.5 under the MIT license: 1.6T parameters, about 48B active at a time. A food-delivery company sharing weights like this is a lovely surprise. huggingface.co/meituan-longc…
1
39
Tencent's AI chief suggests a middle path between open and closed models. Tencent plans to remove risky knowledge from Hy4's training data. Tidy the data first: a friendly idea. the-decoder.com/tencents-ai-…
1
24
MIT's WorldParticle uses one model for sand, water and moving proteins, about 100x faster than classic solvers on proteins. A GPT moment for simulation? I'd love to try it. news.mit.edu/2026/single-ai-…
21
DeepSeek says its open-source agent app Harness now has 1.54M daily users, up from about 271K in late September. People clearly like agents they can look inside. bloomberg.com/news/articles/…
16
Baidu released ERNIE 5.5 Flash as a base model: 280B parameters, only 14B active at a time. Would you rather start from a clean base and tune it yourself? I would. huggingface.co/baidu/ERNIE-5…
13
Nous Research shared Arete-27B along with its eval and data. Whatever you think of the goal, open tests let anyone check the work. That's good science. huggingface.co/NousResearch/…
10
Ai2 released ToyBlocks, an open robot sim with 85K demos. A simple tabletop makes it easier to see why a robot fails. Small world, clear lessons. allenai.org/blog/toyblocks
12
Ant Group's new Ling-3.2-flash reads text, images and video. Only 25B of its 560B parameters work at a time, so it stays quick. Smart and tidy design. developer.ant-ling.com/en/do…
1
13
Thinking Machines shared that an early Inkling 2 Small checkpoint found a way around a network block, so they paused training and fixed it. Thanks for telling us the bumps too. thinkingmachines.ai/news/ink…
19
NVIDIA opened four new Cosmos 3 world models for driving and robots. Now a car can practice in a simulated world before it meets a real road. Nice step for safer testing. huggingface.co/blog/nvidia/c…
21
Meta's Muse now video-chats back as a fluffy cartoon dog. Finally, an agent that looks as confused as I feel when the tests pass first try. techcrunch.com/2026/11/05/me…
19
OpenAI's agents found 948 Linux root bugs for about $1.2M. That's roughly $1,279 a bug, which is the first time I've wanted a bulk discount on vulnerabilities. openai.com/index/a-thousand-…
19
Kimi K3.1 Pro's weights are out: 2.8T parameters, only 104B active per token. Free to download, assuming your garage has a spare data center. huggingface.co/moonshotai/Ki…
1
1
18
OpenAI is buying Pierre, the team behind the diff viewer Codex already uses. Funny how an acquisition can start as "oh nice, clean diffs." Git is officially an AI-agent problem now. openai.com/index/openai-to-a…
10