Links

Selected references worth standing behind.

This is the broad index: papers, essays, posts, repos, tools, benchmarks, docs, hubs, and useful references.

AI Safety

• Emergent Misalignment
  Research writeup. Narrow finetuning can produce broad behavioral shifts.

• Narrow Misalignment is Hard, Emergent Misalignment is Easy
  Essay. Why broad misalignment may be an easier solution than local harmful behavior.

• Training Large Language Models on Narrow Tasks Can Lead to Broad Misalignment
  Paper. Journal version of the narrow-training-to-broad-misalignment result.

• Weird Generalization and Inductive Backdoors
  Paper. Related work on unexpected generalization patterns and hidden failure modes.

Evaluation

• Medmarks
  Benchmark suite for medical capabilities in language models.

• LAB-Bench
  Paper. Benchmark for language models doing biology research tasks.

• FOMO26
  Challenge. Foundation model challenge for brain MRI.

• Open Graph Benchmark
  Benchmark suite. Standardized graph ML datasets, loaders, and evaluators.

• RoboTwin
  Paper. Dual-arm robot benchmark using generative digital twins for scalable task and data generation.

Tool Use And Agents

• Harness Engineering
  Essay. Building products with agents through environments, specs, and reliability loops.

• Code Mode
  Post. Tool use through code interfaces rather than repeated chat-level tool calls.

• Context Mode
  Post. Pattern for keeping agent context usable when tools produce large outputs.

• Agents Learn Their Runtime
  Paper. Persistent versus reset Python interpreters in CodeAct-style training.

• AI Gave Birth to the 100x Engineer
  Essay. Case study on compounding agent workflows with test harnesses and supporting tools.

ML Systems

• Frontier Model Training Methodologies
  Survey. Open frontier training recipes and implementation choices.

• Scaling LLMs with JAX
  Book. Distributed training practice.

• Beyond Language Modeling
  Paper. From-scratch multimodal pretraining study.

• How to Train the Best Embedding Model in the World
  Essay. Embedding model training, label noise, verification, and dataset scale.

• GPU MODE
  Community and resource hub for GPU programming.

• CUDA Writeups by Tushar Gautam
  Posts. Implementation-forward notes on CUDA kernels and optimization.

Research Engineering

• Distill
  Essays. A high bar for clear technical exposition.

• You and Your Research
  Essay. Hamming on choosing important problems and organizing a life around serious work.

• An Opinionated Guide to ML Research
  Essay. Practical advice on developing taste in machine learning research.

• Principles of Effective Research
  Essay. Research as a skill that can be deliberately improved.

• Fast
  Essay. Examples of ambitious work happening faster than conventional expectations.

• Design Docs for Machine Learning Systems
  Essay. Clear technical writing as part of engineering.

• Machine Learning Design Patterns
  Essay. Recurring patterns in applied ML systems.

• Reproducing Deep Reinforcement Learning
  Essay. Experimental fragility in deep RL.

• ML Productivity
  Essay. Making ML experimentation less ad hoc.

• Harvard CS197: Communicating Computer Science Research
  Course notes. Turning technical work into a clear research artifact.

• Your LLM-assisted scientific breakthrough probably isn’t real
  Essay. Caution against mistaking plausible machine-generated research narratives for evidence.

• The Missing Semester
  Practical computing tools that make technical work less fragile.

• How to Read a Paper
  Compact guide to reading research papers deliberately.

Programmable Biology

• Evo 2
  Paper and code. Long-context genomic foundation model for sequence modeling and design.

• HyenaDNA
  Paper. Long-context sequence models at nucleotide resolution.

• AlphaFold
  Paper. Foundational protein structure prediction.

• AlphaFold 3
  Paper. Structure prediction for biomolecular complexes and interactions.

• OpenFold
  Open implementation and training stack for AlphaFold-style systems.

• Rosalind
  Bioinformatics algorithms through concrete programming problems.