I thought Scott Alexander jumped the gun a bit by declaring victory in this case, just because the prompts used were not the original ones (robot vs. person due

I thought Scott Alexander jumped the gun a bit by declaring victory in this case, just because the prompts used were not the original ones (robot vs. person due to content filters). But Marcus is way off base here and sounding petulant; Alexander is clearly not claiming AI has solved compositionality, his claim is the much narrower one that he won his bet. And the general context to the bet is that usually when he writes an article on AI (at least for the last few years), someone says “we will never get X in the next 5 years”, Alexander makes a bet that it will happen sooner, and X always happens sooner. In this case the X was some loose low bar for the next iteration of compositionality above DALL-E 2 with a multi-year timeframe, and SOTA models at the time of the discussion could (arguably at least) meet that bar. Alexander’s broad claim on compositionality is that simply throwing more scale and/or data at the problem seems likely to solve the problem, to which Marcus counters that these models lack something fundamental and can’t be scaled to human performance. FWIW I find Marcus’ position to be a bit frustratingly ambiguous; he seems to blend two distinct positions: A) NN models are not a model for human intelligence/language B) NN models cannot reach AGI He seems to fluidly switch between these critiques in a way I find a bit irritating. I think it’s quite clear that NN architectures have little to do with the way the human brain does language understanding, lacking the gross structure of the brain, which is certain to affect cognitive capabilities and tendencies. So A) is trivially true. But no AI maximalist cares about using these models as a way to understand or model human language. They care about general intelligence. Even granting A), that does nothing to prove B). Perhaps he simply believes B requires A? That would be odd but would explain his approach.