Carl Ludwig Siegel's genius is one of the best kept secret in mathematics, you might want to check out the publicly available book 'Transcendental Numbers' to see just why Andre Weil called him ''the greatest mathematician of the first half of the 20th century''. You should also check out his 'Lectures on the Geometry of Numbers' and 'Sympletic Geometry'.

DS-STAR: Data Science Agent via Iterative Planning and Verification - arxiv.org/pdf/2509.21825?s=0…

My "The Building Blocks of Today’s and Tomorrow’s Language Models" talk at the PyTorch Conference is now up on YouTube! piped.video/watch?v=nDl6Aj9a… If you have 25 min this weekend, it's a whirlwind tour to catch you up on the key LLM architecture design considerations in popular LLMs this year (plus, an overview of alternative architecture designs). The silver lining of my late arrival and rescheduling: Since there was no talk after mine, it's followed by a 30 min Q&A instead of just the usual 5 min :)

Curated list of system design resources

TIL: Claude Code local sandbox environment is open-source. > native OS sandboxing primitives (sandbox-exec on macOS, bubblewrap on Linux) and proxy-based network filtering. It can be used to sandbox the behaviour of agents, local MCP servers, bash commands and arbitrary processes.

I've been getting a lot of value using coding agents for code research tasks recently - I have a dedicated simonw/research GitHub repo and I frequently have them run detailed experiments and write up the results. Here's how I'm doing that + some examples: simonwillison.net/2025/Nov/6…

The fastest way to ship with AI isn't vibe coding. It's planning. I built an AI image feature by spending 40 minutes NOT coding—just planning with specialized subagents that research, review, and question the spec. By the time I typed "/work," the implementation was already solved. Full breakdown + compounding engineering philosophy: every.to/source-code/teach-y…

absolute masterclass on coding with AI from @kieranklaassen today in @every every.to/source-code/teach-y…

Exciting big Kimi K2 Thinking release! More experts, fewer heads, and even more thinking!

TypeScript’s rise in the AI era: Insights from Lead Architect, Anders Hejlsberg - The GitHub Blog share.google/cKuiSgzzzEqlRTe…

The most intuitive explanation of floats I've ever come across, courtesy of @fabynou fabiensanglard.net/floating_…

Today, we published a paper, “AI and the Future of Learning.” The paper outlines the opportunity of AI to reduce barriers to education and unlock human potential. It also addresses key questions like cognitive decline and cheating, and the need to build tools grounded in learning science.

🏮Async QUIC and HTTP/3 made easy: tokio-quiche is now open-source > Cloudflare's battle-tested > asynchronous QUIC library combining both quiche and the Rust Tokio async runtime > tokio-quiche handles millions of HTTP/3 requests per second with low latency and high throughput blog.cloudflare.com/async-qu…

On-policy distillation is powerful, but @thinkymachines's tinker only supports distilling from a teacher model within the same family, making it impossible for qwen to learn from deepseek, gpt-oss, etc. For the first time, we enabled model-agnostic distillations natively using the tinker api, implemented in our open-source distillation framework spider. The framework now auto-detects differences in model naming and routes to cross-tokenizer KL computation in the training loop. We follow the same algorithm that @huggingface implements in TRL for on-policy distillation, ie. matching token groups by student and teacher many-to-many and computing joint probs for each group. With spider and tinker, model-agnostic distillation can be done in a GPU-free, fully server-side manner, which can empower faster and more budget-friendly iterations of experimentations. With this feature enabled, we will launch distillation runs from a suite of different teacher models, and study how they influence the training dynamics and downstream performances differently. Should be a ton of fun!

We reproduce deepseek-ocr training from scratch, the code, model, results can be found in our website #DeepSeek pkulium.github.io/DeepOCR_we…

Can’t believe it — our Princeton AI^2 postdoc Shilong Liu @atasteoff re-built DeepSeek-OCR from scratch in just two weeks 😳 — and open-sourced it. This is how research should be done 🙌 #AI #LLM #DeepSeek #MachineLearning #Princeton @omarsar0 @PrincetonAInews @akshay_pachaar

You Should Write An Agent fly.io/blog/everyone-write-a… via @flydotio

This aligns perfectly with what Anthropic just published about code execution with MCP Link: anthropic.com/engineering/co… In Claude Code, I've noticed Claude prefers writing and executing code directly in bash rather than using tool calls. I've seen Python scripts appear, execute, and disappear in a single run, Claude knows it needs to minimize token usage and work as efficiently as possible.

I started to learn Lean a few months ago. Now finally complete my first Lean project! With about 10K lines of code, I formalized the a.s. convergence of Q learning and linear TD. Hope AI can formalize more RL theory in the future! github.com/ShangtongZhang/rl…

Here is an example of using a deeply nested Pydantic Schema with +30 fields, Union, Optionals and recursive structures with the new Sturctured Structured Outputs in the Gemini API! 💥 Note: To enable this you need to use the new `response_json_schema` parameter instead of the old `response_schema`. Documentation and Snippet below.