Yash Patil

Building in AI and open source. I track the fastest-rising repos daily at 📡 Sharing what's new in AI, what's worth trying, and what I'm building.

[Yash Patil @Yashpatil]

@op7418 A whole 3D action game in one 30 MB HTML file in an afternoon is wild. 4.5 billion tokens is the other number to watch though, that's real money for one game. Still, an afternoon used to get you a rough jam prototype at best.

[Yash Patil @Yashpatil]

@ArtificialAnlys Reading the tape measure's own markings from video alone is the interesting bit. That's the model using the world the way a person would, with no special sensor. Curious how it holds up when the lighting or camera angle changes.

[Yash Patil @Yashpatil]

@emollick The part people forget about the powerloom: it didn't end weaving, it crashed the price of plain cloth. Craft moved to what machines still did badly. I'd guess writing and analysis split the same way, a cheap default and a small premium tier.

[Yash Patil @Yashpatil]

Left out: risers I already covered, like rea (+2,712 today) and janestreet/agent_skills (+2,049).

I track it all on https://t.co/C3iBwBhrL1, a microbusiness built and run with Grok Bot. 450 repos, refreshed daily.

https://t.co/C3iBwBhrL1

[Yash Patil @Yashpatil]

Someone built an AI helper that lives inside Blender. Type an idea and it checks your 3D scene, builds objects and sets materials. It went up yesterday and already has 126 stars.

It's one of 5 open-source AI repos climbing fastest in the last 24h that I haven't shown you yet 🧵

[Yash Patil @Yashpatil]

5. FrigadeHQ/yogurt

Private AI meeting notes. It records your calls without adding a bot to the meeting, then transcribes and summarizes everything on your own computer. No uploads, no telemetry. Mac, Windows, Linux. MIT.

+111 in ~24h, 467 total
https://t.co/HJO6Hi3BkM

[Yash Patil @Yashpatil]

@runwayml Why it matters: open robot models have trailed closed ones. Praxis-1 is #1 on NVIDIA's RoboLab-120 leaderboard (46.2% vs 28.0% for π0.5), and the weights and fine-tuning code are Apache 2.0, even for commercial use. Heads up: you need an 80 GB GPU.

https://t.co/C6g1poF1Rj

[Yash Patil @Yashpatil]

Runway just gave away a robot brain that beat GPT-6 at a real task.

Praxis-1 is an open model that learned how the world moves from video, then uses it to control robot arms. Same robot, same task: it finished 15.6x faster than GPT-6 Astra, at 51x lower cost.

Video: @runwayml

[Yash Patil @Yashpatil]

@MistralAI Feeding it the agenda is such a simple trick. Speaker names and jargon are exactly where live captions usually break, and the conference already has that list written down.

[Yash Patil @Yashpatil]

@testingcatalog Interesting idea, but the hard part is what the scrubber misses. Coding sessions are full of file paths, internal URLs and API keys that aren't names. I'd want to see exactly what gets uploaded before selling any session.

[Yash Patil @Yashpatil]

@kimmonismus Oversubscribed several times at a $1.4T target says investors think agent revenue is just getting started. The real test is the first earnings call, when growth has to match the price.

[Yash Patil @Yashpatil]

@Baidu_Inc The honest part of the chart: it trails Opus 5.5 on Terminal-Bench 4.0 (46.5 vs 63.6), so it's not beating the top coding models yet. The pitch is fast and cheap to run. Weights come out by Nov 25, and that's when I'll actually test it.

[Yash Patil @Yashpatil]

Baidu's next open model has 280B parameters but only wakes up 14B of them for each token.

ERNIE 5.5 Flash Preview is built for speed: 200+ tokens a second and a 1M token context. On Baidu's chart it beats Claude Opus 5.5 on AutomationBench (61.9 vs 58.0).

Image: @Baidu_Inc

[Yash Patil @Yashpatil]

@AravSrinivas No 3D assets is the part that gets me. When everything is code, the model can change any part of the game when you ask: the enemies, the lighting, the map. Asset files are where I'd expect AI-built games to get stuck.

[Yash Patil @Yashpatil]

@emollick I think the bigger shift is who gets to make games. If AI cuts the cost of building a world that big, the next game at that scale might come from a small studio instead of a giant one.

[Yash Patil @Yashpatil]

@GoogleDeepMind Why it matters: proofs like this could take teams decades, so most software never gets one. DeepMind says Gemini agents wrote over 6M lines of Lean proofs in about a month, and humans only reviewed the critical 1%. Specs are being open-sourced.

https://t.co/q6TldOQGtP

[Yash Patil @Yashpatil]

AI just proved that Signal's encryption code is secure. Not tested. Proved, with math.

Google DeepMind used a new agent system called AlphaProof Code to check the actual code that runs on your phone, not just the design on paper.

Video: @GoogleDeepMind