pranav

🚨 Bengaluru-based @Airbound_Aero has conducted 700 flights for Narayana Health since January 2026 with a zero failure rate.

@abinayaaaa @physical_int congrats!!

for PI day, i'm thrilled to share that I'll be joining @physical_int in a few weeks to work on accelerating robot learning research through fast, observable runtimes! i started playing with robots relatively late - my sophomore year of college - and have learned almost entirely through personal projects. its a dream come true to collaborate w the brilliant folks at PI to make better robot software. stay tuned for more !!!

A reward model that works, zero-shot, across robots, tasks, and scenes? Introducing Robometer: Scaling general-purpose robotic reward models with 1M+ trajectories. Enables zero-shot: online/offline/model-based RL, data retrieval + IL, automatic failure detection, and more! 🧵 (1/12)

For the better part of a decade, I've been writing that the US government's restrictions on Chinese tech companies were unstrategic, poorly explained, and based in caprice (foreignaffairs.com/united-states/…). Now it looks like the administration is turning these controls on American firms

Standard Intelligence's @devanshpandey responds to @tszzl's tweet that "text is the universal interface," and explains why their new foundation model is trained on video: "At some point in the arbitrarily long future, if we only use text models, we could force most things to be text. But I think there are just a lot of things that are much more native when done from a computer-use [perspective]." "GUIs are designed for humans to use. We have this massive long tail of things on the internet that are entirely undoable by LLMs." "For example, when I do ML engineering most of my time is spent doing the grunt work of engineering. It's a lot of looking at graphs, analyzing, and comparing loss curves. You can do this in text, but it's a much larger pain than doing it in the native interface." "There's a reason humans don't interact with a computer purely through text, it would kind of suck."

@_neelr_ @comma_ai scary

@_pranavnt @comma_ai shhhhhh

chat i made a computer use model drive a car @comma_ai

Incredibly impressive! VPT-style approach to computer use at 30fps, done entirely by a team of 5(!)

computer use is too important to relegate to post-training. this has been many months in the making, I'm super proud of what we've achieved as a team and excited to scale!

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.

We’ve made two main advances: the ability to train on our 11M+ hour computer action dataset and understand long-context video. Our video encoder can fit nearly two hours of 30FPS, high-resolution video into a 1M token context window, ~50x more efficient than existing SOTA.

improve AI discourse about 5% just by renaming evals accurately Humanity's Last Exam: PubQuizFromHell MATH: RemedialMath FrontierMath: QuarterFrontierMath SWE-Bench: DjangoBench MMLU Virology: NoiseBench Terminal Bench 2: NoiseBench METR HCAST: GreenfieldCodeGigworkBench

new @OpenAI — in collaboration with @paradigm, we developed an evaluation to measure AI on critical smart contract security capabilities

SUPER BOWL LX CHAMPIONS ‼️

SEAHAWKS ARE THE CHAMPIONS!

Came out to play.

@eumycotan amazing

I spent a bunch of time a year ago thinking about the data wall. A blackpill at the time for me was when I realized that the total stock of natural text data is depleting much faster than Chinchilla's infamous 20 tokens per param compute optimal ratio suggested. Here is a naive BOTEC from back then: Famously, Chinchilla showed that using about 20 tokens per param was compute optimal, measured at 6*10^23 FLOPs. It turns out that even though MoEs are more compute efficient than dense models, training them compute optimally needs a lot more data! In fact, at a 1:32 (97%) sparsity it uses ~6x more tokens per active params (see [1]). The Llama 3 405B report measured 40 token per param to be optimal with their data at 4*10^25 FLOPs. And for a 1:32 sparse MoE model such as DeepSeek v3, this suggests 240 tokens per param could well end up being optimal! At this ratio, things would break down. A 4*10^27 FLOPs model (a pretraining run that might be planned e.g. for 2026) will need 400T tokens. A 5*10^28 FLOPs model would require O(1400T) tokens. These are insane numbers, and they only get worse into the 2030s! The totally unfiltered Common Crawl is about 240T tokens. People have been offsetting this to some extent by training for multiple epochs or repeating the same data a la "Scaling Data-Constrained Language Models" by Muennighoff et al. (2023). Of course, this is a naive BOTEC, and I'm happy to dive into more details, e.g. how much compute might be put into other uses, such as long-horizon RLVR which could well require a lot of those 5*10^28 FLOPs. But we are casually talking about hundreds of trillions to over a quadrillion tokens as compute-optimal! It makes one question whether these numbers are actually necessary for the kind of capability gains we want. We are working on this question at @flappyairplanes, and we're excited to be advised by @karpathy. I will end here with this @ilyasut quote from the @dwarkesh_sp episode with him: "The data is very clearly finite. What do you do next? Either you do some kind of souped-up pre-training, a different recipe from the one you’ve done before, or you’re doing RL, or maybe something else. But now that compute is big, compute is now very big, in some sense we are back to the age of research. [...] Up until 2020, from 2012 to 2020, it was the age of research. Now, from 2020 to 2025, it was the age of scaling—maybe plus or minus, let’s add error bars to those years—because people say, “This is amazing. You’ve got to scale more. Keep scaling.” The one word: scaling. But now the scale is so big. Is the belief really, “Oh, it’s so big, but if you had 100x more, everything would be so different?” It would be different, for sure. But is the belief that if you just 100x the scale, everything would be transformed? I don’t think that’s true. So it’s back to the age of research again, just with big computers." [1] arxiv: 2501.12370

There might be fast takeoff at SFO, but people are forgetting about it in AI. We're building Flapping Airplanes to train models radically differently and fly over the data wall. We can’t wait to show you what we’ve been working on soon.

Today, we release SERA-32B, an approach to coding agents that matches Devstral 2 at just $9,000. It is fully open-source and you can train your own model easily - at 26x the efficiency of using RL. Paper: allenai.org/papers/opencod… Here’s how 🧵