tc

Post ideas

  • The middle path: AI safety and x-risk, AI ethics, and progress studies
  • The state of the AI safety, ethics, and governance landscape
  • AI ethics startups: no differentiation
  • The relationship between mechanistic interpretability and causal inference
  • Talent, luck, and precision of evaluation in interview settings -- producing very certain but wrong outcomes
  • Prompt engineering for diffusion models
  • Simulating gains from AI progress, AI risk, and losses from slowing down
  • The role of progress in strengthening institutions and reducing risk
  • Signals of change (an idea from Jane McGonigal)
  • Restrictive institutions vs. enabling institutions
  • "Why not run the hurry-up offense all the time?" and other similar analogies
  • Heterogeneous treatment effects, quantile regression, and thinking beyond the mean
  • LLMs used as research tools will have the same downsides as metaanalyses when used for truth discovery
  • Philosophical underpinnings of both AI ethics and AI safety are driven by a very small number of philosophers
  • Movements whose aims are undermined by their own supporters beliefs/tactics
  • Applying structural hole theory to become influential
  • Fun things you didn't know you could do with AI
  • Folk certainty / overconfidence in how interventions, laws, predictive models will behave
  • The relationship of maximizing expected value, quantile regression, and utilitarianism
  • Rationalist / EA vibes and how they misunderstand influence
  • Quantification bias
  • Quality and testing in machine learning
  • Using AI to keep up with the AI literature
  • Why do so many "data-driven" initiatives fail?
  • Being a ratio thinker, not an ordinal thinker
  • Most of the AI x-risk community relies on a single study of AI timelines
  • Believing that a bad outcome can only be alleviated by even more of the cause
  • Learning reinforcement learning using C64 games
  • Explainable AI isn't super explainable unless you already understand the AI
  • Wondering why it is fashionable in tech to opine on perceived overstaffing and get excited about layoffs
  • Designing a metric for goal-setting
  • Diversity of optimization methods creates interestingness
  • P(A | B) ≠ P(B | A)
  • Using your childhood interests as a guide to making decisions as an adult is a terrible idea
  • The fundamental problem of causal inference in plain English