It brings me no pleasure to report that completing a minor task you've been avoiding (1) is not very hard and (2) makes you feel better afterwards

Logo drop! 🇮🇳 This is what Team India will wear for its historic first appearance at the International AI Olympiad!

The theme: 8 feathers for our 8 incredible Olympians. Let's cheer them on!

#IOAI2025 #TeamIndia #AI

I wrote a very long blog post about AI writing. I hope you'll read it.

meresophistry.substack.com/p/the-mental...

I want to share my latest (very short) blog post: "Active Learning vs. Data Filtering: Selection vs. Rejection."

What is the fundamental difference between active learning and data filtering?

Well, obviously, the difference is that:

1/11

rlhfbook also available on arxiv for SEO 😀 happy friday
arxiv.org/abs/2504.12501

DeepSeek-R1 Thoughtology: Let’s <think> about LLM reasoning

142-page report diving into the reasoning chains of R1. It spans 9 unique axes: safety, world modeling, faithfulness, long context, etc.

Very happy to see "Can AI writing be salvaged? Mitigating Idiosyncrasies and Improving Human-AI Alignment in the Writing Process through Edits" get a Best Paper Honorable Mention and is in the Top 5% of submissions for #CHI2025! 🎉 @chi.acm.org

Check it out here: arxiv.org/pdf/2409.14509

Graham's Scan (1972) is an O(n log n) algorithm for finding the convex hull of a set of 2D points. It sorts points by polar angle, then builds the hull by pushing points onto a stack, popping them when a clockwise turn is detected. en.wikipedia.org/wiki/Graham_...

AM-DeepSeek-R1-Distilled-1.4M: Massive reasoning dataset for LLM training

- 1.4M high-quality reasoning problems with verified solutions
- 900K entries distilled from DeepSeek-R1-671B
- Covers math, code, and complex reasoning tasks
- Bilingual (Chinese/English)

huggingface.co/datasets/a-m...

A reinforcement learning system to beat Pokémon Red. The system successfully completes the game using a policy under 10M parameters, PPO, and a few novel techniques.

drubinstein.github.io/pokerl/

InternLM v3

- Performance surpasses models like Llama3.1-8B and Qwen2.5-7B
- Capable of deep reasoning with system prompts
- Trained only on 4T high-quality tokens

huggingface.co/collections/...