@__avik
Account based inUnited States!
About this account
- Account based in
- United States
- Connected via
- United States App Store
X says this location may be affected by a proxy or VPN.
Account-level information from X, not a live location or the device used for a specific post.
Pinned Tweet
Episode 1 of a new video series on #PyTorch Compiler topics is now out!
In future episodes I'll be hosting other people on the team to share tips and tricks that help you get the max out of torch.compile, torch.export, and related technologies, while enjoying a glimpse into all the cool engineering work that goes on behind the scenes.
Stay tuned! We will drop new videos pretty frequently on this channel. Lots of great content incoming.
We're kicking off a new PyTorch Compiler video series 🎬 In the first episode, Avik Chaudhuri discusses a few simple principles that drive the programming model for export & illustrates them via a series of examples.
🔍 Export creates ahead-of-time IR representations of PyTorch models that provide higher safety and coverage than previous technologies like TorchScript and FX. Exported IR can be further lowered and optimized to run on heterogeneous hardware by runtimes such as ONNX, TensorRT, ExecuTorch, and AOTInductor.
📺 Watch now: hubs.la/Q03plFhR0
I quite liked the nuances in this post by @wtgowers. It also made me re-read The Two Cultures of Mathematics (dpmms.cam.ac.uk/~wtg10/2cult…).
It is interesting to see the internal struggle that mathematicians are currently going through right now, but it is not unlike the effects that the revolution in programming has already had on programmers.
I do not think the answer is as gloomy as being feared, and it is a common one across fields: the value of human mathematicians is more or less analogous to the value of human programmers, human physicists, etc.
Namely, if problem solving is the end goal, then accept that AI is a valuable automation tool towards it, and that is not the thing to focus on. Rather, the focus should be on asking the right questions that lead to new problems, and defining the right abstractions to generalize solutions.
So far, my daily use of AI for coding has not been in conflict with the higher-level human exercise at pointing it to new problems and thinking how to generalize its solutions. Likewise, all the examples of AI use in mathematics I've seen so far have been in the sub-area of finding counterexamples or proofs. I do not think that is where the real value of mathematicians lie.
One only needs to look at the cliched physics analogy of Einstein coming up with general relativity to explain deviations from Newton's model. We could have instead had a series of corrections to Newton's model that fit experimental data better and better and be none the wiser.
I've written a blog post responding to the letter about maths and AI signed by 25 Fields medallists. As with the Leiden Declaration, I didn't sign it, but I agree with much of it and welcome its existence.
gowers.wordpress.com/2026/09…
This sort of thing confirms my hope that we’re about to explode. @THeFullCo’s toothpaste analogy in the recent @OhSoSpurs video sealed it.
If you've watched Spurs this season and felt simultaneously encouraged and frustrated, the numbers agree.
Four games in, the data from @Gradient_Sports is already showing some clear patterns.
1. Tottenham are one of the league's most possession-dominant sides.
Avik Chaudhuri retweeted
Pyrefly 1.3 is out: better control over errors, experimental tensor-shape checking for JAX/NumPy/PyTorch, and DataFrame schemas for Polars.
Python has always had shapes and schemas. Now they can stop living in comments and prayers.
pyrefly.org/blog/v1.3/
Avik Chaudhuri retweeted
🚨 Introducing The 82nd Percentile (@82ndPercentile)
Alongside @JimOhSoSpurs and @TheFullCo, we're launching a Spurs podcast & analysis platform.
Our aim is: combine data, tactical expertise and informed fan perspective to provide an independent view of Spurs performances. 👇
My wife returns today after her yearly African wild life pilgrimage. Her favorite coconut curry is waiting.
Funny thing I'm learning about Codex that I hadn't observed with Claude: context compaction can make it apply previous instructions out of context.
This is scary.
[State A]
Me: "Looks good, let's amend this into the current commit and plan out next steps."
[State B]
Codex: "Amended. Here's the plan. (1) ... (2) ..."
Me: "Looks good. Go."
[State C]
...implements (1), context compacts
[State D]
...wakes up, reads my message from [State A], amends, then continues with (2)...
WTF?
Avik Chaudhuri retweeted
Compute Polynomials Twice as Fast -
Here's a fun result, done entirely without AI.
That because we proved it years ago, but it ran a hundred pages and we weren't quite sure enough to publish. Now AI verified it in Lean.
A fun fact about polynomials is that
P(x) = x⁴ + a₃x³ + a₂x² + a₁x + a₀
can be evaluated in just two(!) multiplications!
The trick is to write
y = (x + b₀)x + b₁
P(x) = (y + x + b₂)y + b₃
where b₀…b₃ are easy to calculate from a₀…a₃.
Polynomials are everywhere from computing exp/sin, to cryptographic hashes and codes. Over finite fields multiplications are particularly expensive, so a 2x speedup matters.
Donald Knuth and others showed ~n/2 multiplications suffice for any degree n, but their preprocessing needs complex roots: numerically unstable and useless over finite fields. Rabin & Winograd fixed that with rational preprocessing, but 2logn extra multiplications, which hurts at the small n we most care about.
Our new method solves this. It uses ⌈(n+1)/2⌉ multiplications, and we prove this is optimal.
You can try it out on your own polynomials at thomashale.com/fast-polynomi…
Avik Chaudhuri retweeted
Just landed: /diff is now a persistent pane that you can scroll and click. It updates in real-time.
For the times when you want to see the code without having to switch windows.
Enjoy!
Avik Chaudhuri retweeted
Am teaching grad complexity theory at CMU; about 1/3 of the lectures will be new (vs. last time), 'modern' results. Videos are going onto youtube.com/@ComplexityTheor… which will later also feature student videos.
We did Williams (/Cook-Mertz/Shalunov) TIME(t) in SPACE(~√t) today.
Dulce!
Lean has a new checker: con-leche, a CONsistent LEan CHEcker. This is an external checker for Lean that is proven (in Lean) to be consistent, meaning it does not accept a proof of False.
Joachim Breitner (@nomeata) is the mastermind behind the project. github.com/leanprover/con-le…
Avik Chaudhuri retweeted
Hello, world. We’re Pyrefly: a fast, open-source type checker and language server- built in Rust for first-time typers, large codebases, IDEs, and coding agents.
Follow for approachable typing, benchmarks with receipts, and fewer bugs disguised as vibes.
pyrefly.org
Tensor shapes will soon be a mainstream feature of type systems!
PyTorch is just the beginning. Join us on this revolution as we work together to bring this to other tensor libraries like JAX, numpy, Triton, CuTe, etc. in the future.
Ever wondered why type systems don't automatically track tensor shapes across your neural network modules?
In his upcoming talk at PyTorch Conference North America 2026 (October 20–21 in San Jose, CA), Avik Chaudhuri from @Meta breaks down how PowerFly - a static type checker for Python - enables end-to-end static shape coverage for PyTorch models.
Register for PyTorchCon North America today: bit.ly/4qStJvL
#PyTorchCon @__avik
There's a new version of this post
For the past several months I've been working on a performance prediction model for Triton that can be used to dramatically reduce autotuning overhead. Excited to get a chance to finally talk about it publicly in this month's Triton community meetup, check it out! (Timestamp: 26:35, ignore the thumbnail.)
TL;DR: Powering this approach is a static analysis that computes symbolic formulas for bytes, FLOPs, etc. using chain-of-recurrences and other badass-ery. Automating precise "pen-and-paper" roofline analysis that an expert would do. A simple regression model then chews on these formulas and spits out predicted running times, learning hidden physics of the hardware.
youtube.com/watch?v=KcjctHAt…
Avik Chaudhuri retweeted
Oh my goodness…
eprint.iacr.org/2026/1630