research taste
The quality of judgment that lets researchers run the right experiments, learn efficiently from data, and avoid flailing; identified as a key bottleneck input to AI R&D.
As defined in AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo
Other definitions
The judgment and intuition required to select which scientific experiments are worth running and which research directions are likely to be fruitful; requires understanding not just how to execute experiments but which experiments are most valuable
What Will AI Look Like in 2027? | Interview
Where it comes up · 4
- Dwarkesh Patel Podcast - The avalanche that hit math is coming for AI development4 claims
- AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo3 claims
- What Will AI Look Like in 2027? | Interview3 claims
- AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo2 claims
Most important claims · 10 of 12Search all claims for “research taste” →
- Once AI research is fully automated with millions of parallel agents running at high serial speed, the quantity of researchers and serial speed cease to be bottlenecks (with diminishing returns past ~90x serial speed); the binding constraints become research taste (ability to learn from experiments and pick the right ones) and compute for running experiments.AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo
- Contrary to the view that progress is mostly random fortuitous discovery, the recent history of AGI shows small startups with a coherent vision and a few cracked researchers repeatedly crushing giant companies with more money, more researchers, and more compute — which would not happen if progress were mostly random, implying vision and taste matter more than raw experiment-throwing.AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo
- Even serendipitous discoveries (like Ozempic from Gila monster venom) are not randomly distributed but come disproportionately from intelligent researchers with good technology in first-world countries; you get more discovery 'lottery tickets' with intelligence, good technology, and good work, so superintelligence with vast compute and good research taste would systematically out-discover human institutions.AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo
- The bottleneck to AI progress is not primarily researcher headcount (only 20-30 people on core pre-training teams at major labs, and companies don't aggressively hire more PhDs despite having capital), but rather a combination of compute resources, research taste, and serial speed of thinking; parallel multiplication of researcher count has diminishing returns.AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo
- Superhuman coding capability alone is insufficient for automating all AI research because systems need additional skills like research taste (judgment about which experiments to run), long-horizon planning, and coordination ability to work in large distributed teamsWhat Will AI Look Like in 2027? | Interview
- Before the 2017 transformer paper, attention, residual networks and other ideas were already known, and the transformer was only obviously better once half a dozen additional tweaks were combined, which required research taste and many small experiments.Dwarkesh Patel Podcast - The avalanche that hit math is coming for AI development
- Because models generalize far beyond what they were trained on, models trained on less ambitious verifiable math problems can solve problems no human had cracked, so the same generalization should make models excellent at unverifiable parts of AI research like research taste.Dwarkesh Patel Podcast - The avalanche that hit math is coming for AI development
- Choosing which objective to optimize is part of research taste, but paradigm-agnostic outer-loop proxies such as Millennium Prize solves, time to rediscover general relativity from pre-1905 data, and speed of mastering unseen games can verify progress across paradigms.Dwarkesh Patel Podcast - The avalanche that hit math is coming for AI development
- The scenario assumes superhuman AI researchers will face serious diminishing returns when scaling up parallel agents, and the main bottleneck will shift from researcher quantity to research taste (ability to learn from experiments efficiently) and available compute, not to having more minds in parallel.AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo
- Superhuman coding ability is a distinct milestone from the complete automation of AI research because AI research requires judgment about which experiments to run (research taste), not just coding ability.What Will AI Look Like in 2027? | Interview