I think my thoughts are coalescing into "AI as an imprecise patternmatcher". An analogy that works for me is that AI will happily paint by numbers, and can do a decent job of it if the number colours and amount of shapes to paint remains low enough. It sees a pattern, it grabs its paint and goes to town. It's often imprecise - it might not always have good hand-eye coordination (model size), it might cross lines here and there (hallucinate), it might need something to keep its hands steadier (grounding). You can try and wring more out of it? And it will, from a distance, look good. Might even be hung on a wall and make for a lovelier room. But any painter with an eye for detail will look right through it. And if it's too inprecise, non-experts will notice, too (slop). I think the current AI boom is resting entirely on finding ways to reduce the imprecision through more patternmatching. Now there's a lot of strides made, but there are fundamental limits to the nails we can hit with the transformer-hammer and I feel like we need fundamental discoveries if we want AI models to actually develop a world model, to actually reason instead of pattern-matching what reasoning looks like in an imprecise roughly-good direction. I mean, did you see what Apple concluded? "nuanced understanding" you mean, it sucks at reasoning. just say it sucks Now if your job is just painting by the numbers? I'd be worried, yeah. I mean, let's be real, how many average people just clock in and out of their average jobby job? Answering emails, sitting in meetings, writing a document here or there, all with enough corpospeak and jargon and shibboleths thrown in to make it obtuse to normal people. How much of that isn't pattern matching, really? I recenly started tracking my time again, basically divided into "deep work" (expertise), "email" and "meetings" (patternmatching, basically). The former is only 45% of my entire week. You mentioned Graeber before, but I'm not sure if that's fair, because the patternmatcher doesn't care if it's matchin' patterns for the next TPS report or at a charity to cure rare cancers. "We've always done it that way" tasks and jobs now have their head put under the guillotine, and they are not few or far between I think. Journalism...yeah, not looking great.
We hypothesize that this decline is because current LLMs cannot perform genuine logical reasoning; they replicate reasoning steps from their training data. Adding a single clause that seems relevant to the question causes significant performance drops (up to 65%) across all state-of-the-art models, even though the clause doesn't contribute to the reasoning chain needed for the final answer. Overall, our work offers a more nuanced understanding of LLMs' capabilities and limitations in mathematical reasoning.