Thoughts2025Find the right dataset2024Biggest lessons in AI in past five yearsSolving hallucinations via self-calibrationCooking with AI mindsetOpenAI o3Value of safety researchRL all the timePeople who influenced meTransition to AI for scienceInformation density & flow of papersCoT before and after o1SimpleQAThe o1 paradigmInspiring words from a young OpenAI engineerLevels and expectationsBet on AI research experimentsHistory of Flan-2When I don’t sleep enoughThinking about history makes me appreciate AIAdvice from Bryan JohnsonSora is like GPT-2 for video generationA typical day at OpenAIYolo runsUniform information density for CoTInertia bias in AI researchCompute-bound, not headcount-boundMagic of language modelsWhy you should write testsCo-founders who still write code2023Hyung WonRead informal write-upsRelationship board of directorsReinventing myselfGood prompting techniques10k citationsManually inspect dataLanguage model evalsAmusing nuggets from being an AI residentWhen to use task-specific modelsBenefits of pair programmingMany great managers do IC workWhy I’m 100% transparent with my managerMy girlfriend is a reward modelBetter citation metrics than h-indexMy strengths are communication and prioritizationEmergence (dunk on Yann LeCun)UX for researchersMy refusalThe evolution of prompt engineeringPrompt engineering battleIncumbents don’t have a big advantage in AI researchPotential research directions for PhD studentsBest AI skillset2022Add an FAQ section to your research papersPrompt engineering is black magicWhat work withstands the bitter lessonA skill to unlearnAdvice on choosing a topic