> I mean LLMs crush most humans (even many professionals) at coding/legal/medical/etc. exams.
Exams, yes. Not the work yet. This is also why they've not made doctors and lawyers obsolete: even in the fields where the models perform the best, you're getting the equivalent of someone fresh out of university with no real experience.
I suspect you're right about your point with generalise vs. memorise. Not absolutely sure, but I do suspect so.
I also suspect we'll get transformative AI well before we can train any AI with as few examples as any organic brain needs. Unless we suddenly jump to that with one weird trick we've been overlooking this whole time, which wouldn't hugely surprise me given how many radical improvements we've made just by trying random ideas.
Understanding in this sense seems different from the memorizing+flexible retrieval we know LLMs exceed at because it extends much further beyond its training distribution. If I ask a (good) medical professional a question unlike anything they’ve ever seen before, they’ll be able to draw on their “understanding” of anatomy to give me a decent guess. LLMs are inconsistent on these kinds of questions and often drop the ball.
We can also point to training data requirements as a discrepancy. In brain organoid experiments, we observe a much lower quantity of examples are required to achieve neural-network-like results. This isn’t surprising to me. Biological neurons have exceedingly complex behavior; it takes hundreds of neural network nodes to recreate the behavior of a single neuron, and they can reorganize their network structure in response to stimuli, form loops, operate in non-discretized time, etc. We don’t know how far transformers will be able to go, but I think that if you want to make a model that holds a candle to that sort of complexity, you’ll need at least many, many orders of magnitude more scale or, more likely, a different, less limiting architecture.