Ankesh Bharti
@_feynon
applied researcher and technologist. he/him/25/blr. building https://tiles.run. stewarding https://userandagents.com. publishing https://tracingneurons.com.
Dropped the Virtual Cell Challenge Primer on HF. We are shipping transformers support for STATE (the SOTA model for predicting perturbation response) very soon!
Crustaceans evolve into crabs U.S. companies evolve into banks Websites evolve into short form video feeds And apparently, mammals evolve into anteaters
“The arc of history is long, and it bends toward” *flips through history book* Oh. Oh no. Oh no no no
The new skills gallery in @diabrowser gives you access to the expertise of a very talented community of creators. Here's how it works:
This is the most thoughtful thing I’ve read on LLMs and coding, very much resonates with my experience. Thank you @nicholasbs @recursecenter. recurse.com/blog/191-devel…
there’s a reason we purchased os.ai from @dharmesh. the roadmap is to make comet feel like your own mini customized computer within your existing computer or phone. and compute running across client and server with ability to run local models too. lot to do!
shortcuts for repetitive tasks rolling out next week on comet. more invites will be sent next week too. the browser is going to be your personal console for getting work done.
We are working on a tutorial, but its not yet ready. Will publish it more prominently when ready: github.com/ggml-org/llama…
My talk titled "Your GPU is a JavaScript Runtime*" just landed on YouTube! 👨💻 I answer WHY and HOW we compile JS/TS to WGSL, how it improves interop for the whole ecosystem, and the added dimension of customizability it unlocks for WebGPU libraries Video below ⬇️
2/ Strong growth in AI usage across our products and platforms: We’re processing 980 trillion+ monthly tokens across our products and APIs (up from 480T at I/O in May) AI Overviews in Search now has 2B+ monthly users across 200 countries/territories and 40 languages 450M…
Introducing Lucy: 1.7B model that Google for you It's an agentic‑search model that can even run on your phone. - Agentic search on tap - Lucy calls tools (<think></think>‑aware) - Fits in your pocket - runs on CPU or mobile Under the hood: - Built on @Alibaba_Qwen's Qwen3‑1.7B…
New ways to engage with artifacts on mobile: Create interactive tools, browse the gallery, and share your work directly from your phone.
Who will own our digital memory? As AI platforms ask us to entrust our memories to them, we risk endless reintroductions to every AI or (worse) seeing our digital selves held hostage by a single AI overlord.. 1st in a 2-part blog series up now on building an open architecture…
early adopters are generally smarter specifically because they can look at the foundation of the project and extrapolate where it will be in a few months most people get stuck on "it's missing feature A/B/C" and miss what's actually going on
damn, scout is one of my most fav AI apps, which got cloned by Soham, says a lot
It's a true honor to be cloned by @realsohamparekh
Talk is live! Will add as reply b/c... u know. Thx @aiDotEngineer + @swyx :)
Not sure when my @aiDotEngineer talk will be posted, so for those who asked here's a link to my slides (couldn't add notes so if you anything needs explanation just reply here & I'll thread the context): figma.com/deck/0NURDonIV…
“Cursor Friends” Use Dia to send AI assistants off on assignments. Each gets their own cursor and browser tab, just like you. Browsing side-by-side feels super natural. Design is way off, but we built it. Should we finish & ship? Each cursor must feel like a Pixar character!
real, JEE Mains is commodtised now
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
Paper is out. Link: openreview.net/pdf?id=TJjP8d5…
KPOP is a new DL optimizer designed for large scale distributed training on Apple Silicon. KPOP uses a lot more memory but is more efficient per FLOP than AdamW, so it's a better fit for hardware with a high memory:flops ratio. Some hardware numbers: H100: 80GB, 1000TFLOPS…