Why could the same AI rebuild a 16,000-line program without seeing its code, yet fail to customize a library for METR? The answer shapes what you hand it.
Watch now (21 mins) | Are we developing a more general intelligence, or getting much better at producing extraordinary abilities that transfer unevenly?
TL;DR
- AI's progress might not be uniform, with capabilities becoming 'spikier' and gaps widening.
- Extraordinary performance in one domain, like mathematics, does not automatically imply equal progress in others, such as engineering judgment.
- OpenAI's internal model producing mathematical research with machine-checked proofs exemplifies narrow superintelligence.
- The inference from mathematical breakthroughs to broad competence is a key assumption in public AI narratives that needs closer examination.
- Uneven AI development could result in increasingly astonishing successes and harder-to-reconcile failures.
- Training feedback concentration, where useful feedback is plentiful, can accelerate progress in specific abilities faster than others.
- A system that notices requirement conflicts with project purpose without prompting would be evidence against uneven development.