Sasha Rush

Cursor is now on your phone and on the web. Spin off dozens of agents and review them later in your editor.

o3 since this was driving me crazy: A type that implements Rust’s Try trait—like Result or Option—is a “fallible wrapper” that can produce either a success value (Output) or short-circuit with its error/none form (Residual). The Residual itself is that error/early-exit shape, and…

🚨 The era of infinite internet data is ending, So we ask: 👉 What’s the right generative modelling objective when data—not compute—is the bottleneck? TL;DR: ▶️Compute-constrained? Train Autoregressive models ▶️Data-constrained? Train Diffusion models Get ready for 🤿 1/n

20 years ago, this type of person would become an elite math professor. Now they're making AI breakthroughs. This is progress (probably)!

1/N I’m excited to share that our latest @OpenAI experimental reasoning LLM has achieved a longstanding grand challenge in AI: gold medal-level performance on the world’s most prestigious math competition—the International Math Olympiad (IMO).

One ML trend I've been grappling with is that we are post-abstraction. Since everyone is building roughly the same "one big model" there isn't really much need generality. We're roughly converging to giant, straight line, assembly-coded systems that get recoded each year.

Scaling Data-Constrained LMs is now also in JMLR: jmlr.org/papers/v26/24-… Looking back at it 2yrs later, repeating & mixing seem standard now, but maybe another powerful lever to scale data-constrained LMs turns out to have been RL - arguably underrated back then!

haven't made a new blog post in over a year, so here's a new one: justintchiu.com/blog/sftrl/ it's short

Can we build an operating system entirely powered by neural networks? Introducing NeuralOS: towards a generative OS that directly predicts screen images from user inputs. Try it live: neural-os.com Paper: huggingface.co/papers/2507.08… Inspired by @karpathy's vision. 1/5

Can transformers analyze code efficiently? ✅ Yes. We prove transformers efficiently handle real compiler tasks (AST construction, symbol resolution, type infer) using only log size—while RNNs require linear size (in input length). Paper: arxiv.org/abs/2410.14706 #COLM2025

L1 is heading to COLM! We've released 5 new open L1 models and the Massive-Math dataset to celebrate:

two updates: 1. flying to ICML tonight 2. i joined @cursor_ai a month ago come talk to me to learn what makes research at cursor special :)

😅

Can an AI model predict perfectly and still have a terrible world model? What would that even mean? Our new ICML paper formalizes these questions One result tells the story: A transformer trained on 10M solar systems nails planetary orbits. But it botches gravitational laws 🧵

Tokenization has been the final barrier to truly end-to-end language models. We developed the H-Net: a hierarchical network that replaces tokenization with a dynamic chunking process directly inside the model, automatically discovering and operating over meaningful units of data

Tokenization is just a special case of "chunking" - building low-level data into high-level abstractions - which is in turn fundamental to intelligence. Our new architecture, which enables hierarchical *dynamic chunking*, is not only tokenizer-free, but simply scales better.

WOW! 🤯 this groundbreaking dataset from Meta’s Chief AI Scientist has revolutionized the way that we understand vision 👀 🚀 is this one of the highest-impact releases of all time?? ⏳🔥 10 crazy examples below: 🧵

Finally closed our $11M+ funding round! Backed by top Japanese VCs and amazing angel investors including Joi Ito, @Thom_Wolf from @huggingface, @nlpnoah, @LukeZettlemoyer, and @srush_nlp. Now it’s time to focus on commercialization and tech development!!