I'm bearish that continued scaling will fully solve this. In the meantime though, as a commentor in the OP points out - by solving the menial stuff (building diagnostic tools, producing visualisations, scaffolding experiments) they're still proving to be extremely useful research assistants.
They literally generate stuff that's averages of other stuff.
It's not interesting in the least
If you know how they work and how they are trained this makes perfect sense.
Going out and searching the web gets around model cutoffs, so fixes the latency of knowledge issue a bit, but they’re still finding statistical modes of what they’ve seen before.
“Stochastic parrot” is seen as a crude and dismissive jibe at LLMs, but it’s actually a concise observation of the reality.
E.g. consider the best answer describes how to achieve something with some CLI, but the CLI's changed and the command is now broken. It can discover that, combine with the 'best answer' about what changed in the CLI, and fix it to work for the original problem even if that's completely novel and there were no extant answers about solving the problem with the new CLI form, or that it had broken that use case.
It doesn't change fundamentally how they work of course, but 'thinking', creating that feedback loop, made them significantly more useful IMO.
An LLM is like a low IQ person with vast knowledge.
How many people on this planet are experts on every topic?
>An LLM is like a low IQ person with vast knowledge.
Are you aware that frontier models have scored in the 135-195 IQ range when they have been tested? How smart are you?
If they were really stupid, a Prolog program, a genetic algorithm or a SAT solver would be enough to reach the goal. Clearly it was not the case.
It gels because they solved about 700 problems out of 4000 attempts. And they threw a ton of resources at each. So "per problem" it's less than 1 in 5, "per token" it's probably not a great record. I mean it's astonishing that it's possible at all of course. But on average, even an expert with a top model will probably have an underwhelming response to the vast majority of prompts.
Ie. the LLM itself isn't as "smart" as the combination of LLM, Lean, looping, search and vast gpu/cpu-time.
On the other hand, using AI to trawl through all the grunt work, experiments, and well trodden areas means I can crawl up pretty close to that giants shoulder and spend way less effort myself answering the actually important questions so I can stand on it.
(This isn't supposed to sound self aggrandising - the giants whose shoulders I'm standing on aren't very tall, and they probably have bigger brothers and sisters, but the idiom was too good of a match to pass up using)
Ask anyone who has an advanced degree in math or physics. They will tell you that doing the boring practice problems over and over is usually the key to understanding, not reading the high level theorems.
There was a critical review article with a title that summarizes the book well: "On the shoulders of a very large pyramid of hobbits, or why we are not just very smart chimps. The Secret of Our Success: How Culture is Driving Human Evolution, Domesticating Our Species, and Making Us Smarter." Obviously I got triggered by your comment, which was eerily close to that title.
These seem pretty easy to solve problems though if I could just afford 100 or 1000 agents. Let them explore and compete and work out their errors for themselves.
At a practical level though, I do a lot more 'bonus' R&D work now by just keeping the agents running and testing ideas in the background.
Did you see the chart on model performance over time on LifeVal? Ceiling in there. https://openai.com/index/lifeval/
You're only looking at the symbolic part: deriving a statement from prior statements. I agree that the derivation either is correct or isn't.
But then there's the rest: the philosophy and intuition behind the symbolic manipulation. You can skip that if your goal is to use the results, but I don't think you can skip that if your goal is to develop the field.
I feel great sympathy for the mathematicians. This must be gut wrenching. It will lead to progress eventually, for sure, but right now a lot of them are in a bad place.
Edit: for readability.
Citation needed.
I suppose it's possible that frontier labs have internal models that are so far beyond that autonomous research becomes possible, but I'm not sure that's it.
It seems to me like these models are like a very competent, important part of some hypothetical digital brain, but alone they don't have all the facilities they need to be useful, like a genius with severe ADHD and short-term memory loss.
For anything where quality isn't objective in a clear-cut way, very intentional context management is needed to get good outputs. As context size expands, performance degrades. As the number of tasks increases, performance degrees. As the number of independent decisions based on subjective criteria increases, performance degrees.
If you can get a harness that makes sure every decision is made one at a time, focused with only the context it needs, things work great. But making those harnesses is still hard, a lot of work, and something that agents seem to have no clue how to help you with.
LLMs will either solve issues like this at the architecture level, or declare them out-of-scope entirely.
Codebase porting, implementing sophisticated algorithms, performance tuning, model training, reverse engineering, fuzz testing / benchmarking suites, refactor to support a new interface, etc, scale so much better than 'fix a then b. don't break c' style prompts.
If they tackle that, I think models will be perceived as much smarter than they are now - even though now with the right direction they can already do so so much.
It’s a dance, not an orchestration.
If you want autonomous work of such systems you can not expect that: a) each reasoning step will make sense and be correct; b) results will come very quickly - especially in the first stages of the project.
I have very positive experience in non-trivial problems with using AI as an autonomous researcher or engineer. But it worked since an agent had a separated, fully dedicated environment and enough time to build up (emerge) some strong patterns of work. Early days were underperforming and slow progress. Now it is overperforming and frequently surprising me with new stuff.
The worst part of his premise is that they can be overcome by increased intelligence. He’s not describing fundamental limitations. Meaning that it’s inevitable that his examples will be “solved.” At least better than humans. Meaning the R&D is obviously valuable.
For example, I improved the SOTA on "listing all legal Scrabble moves" by 10x just by prompting Astra over and over to try harder, to aim for <specific performance level>, to "systematically try every method in the literature that seems kind of related," etc.
They aren’t capable of generating truly innovative ideas though. Ask it to come up with a name, it will come up with the same weird slop names over and over. Always gives people the surname Okafor. Because it works by generating probable answers, it can only generate probable names.
These models actually can create content of myriad styles very well. But artistic styles are mutually incompatible. It can't create something that matches all of them, and it doesn't have an opinion on which is best, so bland requests just give bland results. To get good results you need to direct the model into the region of its latent space that makes stuff you thing is good.
That's not true anymore. I've asked it to generate names for some LLCs and it came up with plenty of novel names that sounded vaguely Latin and had a certain sound that I had asked for.