PinnedEugène Rastignac·Sep 4, 2024It has gone viral! Check 3Blue1Brown’s latest video on how LLMs store facts.When typing “Michael Jordan plays the sport of ___” into a Large Language Model (LLM), it will likely predict the next word to be…
Eugène Rastignac·Oct 19What the DeepSeek-v4.1-Flash-1017 Update Changes for Engineers Building AgentsThe same 552B base model got a post-training update. Three parts of the announcement say more than its slogan
Eugène Rastignac·Oct 19Rules vs Examples for In-Context Learning Across 5 Task TypesYou can state the rule or show worked examples. One study measured which one the model follows more accurately.
Eugène Rastignac·Oct 18How the OpenCode Breach Drained Claude and GPT Credits OvernightStolen provider API keys ran up bills while developers slept. Only one fix actually shuts a leaked key down
Eugène Rastignac·Oct 18How OpenAI’s Model Wrote 722 Math Manuscripts With Lean ProofsAn unreleased frontier model produced the papers and much of the proof code. Mathematicians now have to read them
Eugène Rastignac·Oct 17Can Anyone Check 722 Math Manuscripts Written by an AI Model?What OpenAI released, what counts as a proof artifact, and why mathematicians are pushing back on how it was published
Eugène Rastignac·Oct 14LLM Interview Note: How to Fix an Agent’s Overflowing Context?6 compression techniques worth knowing when a bigger context window falls short
Eugène Rastignac·Sep 22Jev Beaten by an Open-Source Model? Here Is How to Read That BenchmarkWhat the Jev and Laya comparison actually measures, and which numbers survive when you run the same decisions yourself.
Eugène Rastignac·Sep 20Two Agents Can Argue Forever. Here Is Where the Stop Button Actually Belongs.Why a simple prompt can never end a loop, and what an orchestrator has to own instead.
Eugène Rastignac·Jan 14, 2025Why are LLM Sizes Commonly Designed as 6/7B, 13B, and 130B?Analysis of Scaling Laws in Large Language Models