Thoughts

2025

  1. AI will not takeoff fast

  2. Work in the office

  3. AI should create unicorns

  4. The asymmetry of AI research

  5. Fact-checking good AI researchers

  6. The description-execution gap

  7. RL environment specs

  8. The inverse 80-20 rule

  9. Method-driven vs problem-driven research

  10. AI research is a max-performance domain

  11. ATLA is the ultimate benchmark

  12. Measurement is all you need

  13. AlphaEvolve is thought-provoking

  14. Binary-choice questions for AI research taste

  15. The craziest chain-of-thought

  16. The best hard-to-solve easy-to-verify benchmark

  17. Flavors of AI for scientific innovation

  18. Debugging-prioritized AI research

  19. When scientific understanding catches up with models

  20. Butterfly effect of AI researchers’ backgrounds

  21. Benchmarks quickly get saturated

  22. Deep browsing models

  23. Unstoppable RL optimization vs unhackable RL environment

  24. Dopamine cycle in AI research

  25. Find the right dataset

2024

  1. Biggest lessons in AI in past five years

  2. Solving hallucinations via self-calibration

  3. Cooking with AI mindset

  4. OpenAI o3

  5. Value of safety research

  6. RL all the time

  7. Transition to AI for science

  8. Information density & flow of papers

  9. CoT before and after o1

  10. SimpleQA

  11. The o1 paradigm

  12. Inspiring words from a young OpenAI engineer

  13. Levels and expectations

  14. Bet on AI research experiments

  15. History of Flan-2

  16. When I don’t sleep enough

  17. Thinking about history makes me appreciate AI

  18. Advice from Bryan Johnson

  19. Sora is like GPT-2 for video generation

  20. A typical day at OpenAI

  21. Yolo runs

  22. Uniform information density for CoT

  23. Inertia bias in AI research

  24. Compute-bound, not headcount-bound

  25. Magic of language models

  26. Why you should write tests

  27. Co-founders who still write code

2023

  1. Hyung Won

  2. Read informal write-ups

  3. Relationship board of directors

  4. Reinventing myself

  5. Good prompting techniques

  6. 10k citations

  7. Manually inspect data

  8. Language model evals

  9. Amusing nuggets from being an AI resident

  10. When to use task-specific models

  11. Benefits of pair programming

  12. Many great managers do IC work

  13. Why I’m 100% transparent with my manager

  14. My girlfriend is a reward model

  15. Better citation metrics than h-index

  16. My strengths are communication and prioritization

  17. Emergence (dunk on Yann LeCun)

  18. UX for researchers

  19. My refusal

  20. The evolution of prompt engineering

  21. Prompt engineering battle

  22. Incumbents don’t have a big advantage in AI research

  23. Potential research directions for PhD students

  24. Best AI skillset

2022

  1. Add an FAQ section to your research papers

  2. Prompt engineering is black magic

  3. What work withstands the bitter lesson

  4. A skill to unlearn

  5. Advice on choosing a topic