The recent improvements in practical utility of AI have decreased my probability of near-term recursive self-improvement. I’m updating towards a Chollet-like view. I’m increasingly convinced that LLMs really are a collection of fairly dumb heuristics (surface-level pattern matching) that are intelligent through breadth, and that there’s a qualitative difference from human intelligence. We use fewer heuristics but each one is smarter (more generally applicable, abstract, robust).
In the past it was hard to distinguish whether LLMs were developing smarter heuristics with each improvement or just deploying more and better-selected dumb heuristics. “Needs smart heuristics to solve” and “really hard to do” are highly correlated properties of tasks, so every capabilities gain looks like evidence for model general smartness. Now that we see really hard tasks being done we can differentiate the two properties, so I revised my prior capabilities-based timeline updates down. I worry people are now making the same conflation between “Needs smart heuristics to solve” and “unverifiable.”
Problems that are bottlenecked by deploying smart heuristics won’t be accelerated as much as people expect. AI research is one of them, as evidenced by labs paying extreme premiums for research taste rather than massively expanding headcount.
The recent improvements in practical utility of AI have decreased my probability of near-term recursive self-improvement. I’m updating towards a Chollet-like view. I’m increasingly convinced that LLMs really are a collection of fairly dumb heuristics (surface-level pattern matching) that are intelligent through breadth, and that there’s a qualitative difference from human intelligence. We use fewer heuristics but each one is smarter (more generally applicable, abstract, robust).
In the past it was hard to distinguish whether LLMs were developing smarter heuristics with each improvement or just deploying more and better-selected dumb heuristics. “Needs smart heuristics to solve” and “really hard to do” are highly correlated properties of tasks, so every capabilities gain looks like evidence for model general smartness. Now that we see really hard tasks being done we can differentiate the two properties, so I revised my prior capabilities-based timeline updates down. I worry people are now making the same conflation between “Needs smart heuristics to solve” and “unverifiable.”
Problems that are bottlenecked by deploying smart heuristics won’t be accelerated as much as people expect. AI research is one of them, as evidenced by labs paying extreme premiums for research taste rather than massively expanding headcount.