Lucas Chen

on the atlas · 26

• A Mathematical Framework for Transformer Circuits40 savers
• Mail - Lucas Chen - Outlook1 savers
• Four LLM loss functions → four flavors of LLM misalignment — LessWrong4 savers
• What just happened? Pragmatism and Pessimization — LessWrong8 savers
• nn-notes.pdf4 savers
• An OpenAI model left notes about how to evade containment; we need more details — LessWrong1 savers
• [0906.4325] Finite element exterior calculus: from Hodge theory to numerical stability1 savers
• A Mechanistic Explanation of Prompt Injection (and why you should study roles) — LessWrong6 savers
• google/hackable_diffusion ·1 savers
• A Visual Guide to DiffusionGemma - by Maarten Grootendorst1 savers
• google/diffusiongemma-26B-A4B-it · Hugging Face1 savers
• A (Slightly) Mechanistic Theory for Exponentially Increasing AI Time Horizons? — LessWrong2 savers
• ProblemsILike.com — Daniel Litt1 savers
• Natural Language Autoencoders \ Anthropic10 savers
• Notes from the front - by Michael Harris - Silicon Reckoner1 savers
• Maybe I was too harsh on deep learning theory (three days ago) — LessWrong5 savers
• A Technical Introduction to Solomonoff Induction without K-Complexity — LessWrong1 savers
• Deep learning as program synthesis — LessWrong1 savers
• [2604.21691] There Will Be a Scientific Theory of Deep Learning5 savers
• [2503.02113] Deep Learning is Not So Mysterious or Different1 savers
• Introducing talkie: a 13B vintage language model from 193020 savers
• [2410.12101] The Persian Rug: solving toy models of superposition using large-scale symmetries1 savers
• pdf1 savers
• [2502.17424] Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs6 savers
• The paper that killed deep learning theory — LessWrong1 savers
• Lucas's Bookshelf / Curius1 savers

highlights · 3

• NLAs helped Anthropic researchers discover training data that caused this.

  Natural Language Autoencoders \ Anthropic

• When this round is over, they will no longer have any need for mathematics (within 5 years, according to several people who were participating in this morning’s conversation)

  Notes from the front - by Michael Harris - Silicon Reckoner

• Maybe I was too harsh on deep learning theory (three days ago)

  Maybe I was too harsh on deep learning theory (three days ago) — LessWrong