Not to be a luddite, but large language models are fundamentally not meant for tasks of this nature. And listen to this:
> Most notably, it provides confidence levels in its findings, which Cheeseman emphasizes is crucial.
These 'confidence levels' are suspect. You can ask Claude today, "What is your confidence in __" and it will, unsurprisingly, give a 'confidence interval'. I'd like to better understand the system implemented by Cheeseman. Otherwise I find the whole thing, heh, cheesy!
I've spent the last ~9 months building a system that, amongst other things, uses a vLLM to classify and describe >40 million house images of number signs in all of Italy. I wish I was joking, but that aside.
When asked about their confidence, these things are almost entirely useless. If the Magic Disruption Box is incapabele of knowing whether or not it read "42/A" correctly, I'm not convinced it's gonna revolutionize science by doing autonomous research.
How exactly are we asking for the confidence level?
If you give the model the image and a prior prediction, what can it tell you? Asking for it to produce a 1-10 figure in the same token stream as the actual task seems like a flawed strategy.
I’m not saying the LLM will give a good confidence value, maybe it will maybe it won’t, it would depend on its training, but why is making it produce the confidence value in the same token stream as the actual task a flawed strategy?
That’s how typical classification and detection CNNs work. Class and confidence value along with bounding box for detection CNNs.
Because it's not calibrated to. In LLMs, next token probabilities are calibrated: the training loss drives it to be accurate. Likewise in typical classification models for images or w/e else. It's not beyond possibility to train a model to give confidence values.
But the second-order 'confidence as a symbolic sequence in the stream' is only (very) vaguely tied to this. Numbers-as-symbols are of different kind to numbers-as-next-token-probabilities. I don't doubt there is _some_ relation, but it's too much inferential distance away and thus worth almost nothing.
With that said, nothing really stops you from finetuning an LLM to produce accurately calibrated confidence values as symbols in the token stream. But you have to actually do that, it doesn't come for free by default.
Yeah, I agree you should be able to train it to output confidence values, especially integers from 0 to 9 for confidence should make it so it won’t be as confused.
CNNs and LLMs are fundamentally different architectures. LLMs do not operate on images directly. They need to be transformed into something that can ultimately be fed in as tokens. The ability to produce a confidence figure isn't possible until we've reached the end of the pipeline and the vision encoder has already done its job.
The images get converted to tokens using the vision encoder, But the tokens are just embedding vectors. So it should be able to if you train it.
CNNs and LLMs are not that different. You can train an LLM architecture to do the same thing that CNNs do with a few modifications, see Vision Transformers.
> If the Magic Disruption Box is incapabele of knowing whether or not it read "42/A" correctly
Are you implying that science done by humans is entirely error-free?
There exists human research that is worse than AI slop. There is no AI research worthy of the Nobel prize
yet.
Yes and no at the same time, depending on what you intend to get from asking. I don't know what you were doing with this project, obviously, so I don't speak to that, but science (well, stats in general, but science needs stats) has a huge dependency on being sure the question was the correct one and not just rhyming.
Reading hand-written digits was the 'hello world' of AI well before LLMs came along. I know, because I did it well before LLMs came along.
Obviously a simple model itself can't know if it's right or wrong, as per one of Wittgenstein's quote:
That said, IMO not (as Wittgenstein seemed to have been claiming) impossible, as at the very least human brains are not single monolithic slabs of logic: https://www.lesswrong.com/posts/CFbStXa6Azbh3z9gq/wittgenste...In the case of software, whatever system surrounds this unit of machine classification (be it scripts or more ML) can know how accurately this unit classifies things in certain conditions. My own MNIST-hello-world example, split the test set and training set, the test set tells you (roughly!) how good the training was: while this still won't tell you if any given answer is wrong, it will tell you how many of those 40 million is probably wrong.
Humans and complex AI can, in principle, know their own uncertainty, e.g. I currently estimate my knowledge of physics to be around the level of a first year undergraduate course student, because I have looked at what gets studied in the first year and some past paers and most of it is not surprising (just don't ask me which one is a kaon and which one is a pion).
Unfortunately "capable" doesn't mean "good", and indeed humans are also pretty bad at this, the general example is Dunning Kruger, and my personal experience of that from the inside is that I've spent the last 7.5 years living in Germany, and at all points I've been sure (with evidence, even!) that my German is around B1 level, and yet it has also been the case that with each passing year my grasp of the language has improved, so what I'm really sure of is that I was wrong 7 years ago, but I don't know if I still am or not, and will only find out at the end of next month when I get the results of an exam I have yet to sit.
A blind mathematician can do revolutionary work despite not being able to see
Here's a logical step you skipped: A blind matematician can do revolutionary work in mathematics. He is highly unlikely to do revolutionary work in agriculture.
Interesting example, as there was an article on HN front page 10 days ago about exactly that - a blind person doing revolutionary work in agriculture. [0][1]
[0] https://www.bbc.com/news/articles/c4g4zlyqnr0o — "I used Lego to design a farm for people who are blind - like me"
[1] https://news.ycombinator.com/item?id=46502269
> large language models are fundamentally not meant for tasks of this nature
There should be some research results showing their fundamental limitations. As opposed to empirical observations. Can you point at them?
What about VLMs, VLAs, LMMs?
Old "agged Technological Frontier" but explains a bit the challenge https://www.hbs.edu/faculty/Pages/item.aspx?num=64700 namely... it's hard and the lack of reproducibility (models getting inaccessible to researcher quickly) makes this kind of studies very challenging.
That is an old empirical study. jadenpeterson was talking about some fundamental limitations of LLMs.
Finding patterns in large datasets is one of the things LLMs are really good at. Genetics is an area where scientists have already done impressive things with LLMs.
However you feel about LLMs, and I say this because you don't have to use them for very long before you witness how useful they can be for large datasets so I'm guessing you're not a fan, they are undeniably incredible tools in some areas of science.
https://news.stanford.edu/stories/2025/02/generative-ai-tool...
https://www.nature.com/articles/s41562-024-02046-9
In reference to the second article: who cares? What we care about is experimental verification. I could see maybe accurate prediction being helpful in focusing funding, but you still gotta do the experimentation.
Not disagreeing with your initial statement about LLMs being good and finding patterns in datasets btw.
This is also true of lots of human research, there's always a theory side of research that guides the experimental side. Even if just informal, experimental researchers have priors for what experimental verification they should attempt.
Yeah, there’s an infinite numbers of experiments you could run but obviously infinite resources don’t exist, so you need theory to guide where to look. For example, computational methods in bioinformatics to guess a protein function so that experimental researchers can verify the protein function (which takes weeks to months for a given protein function hypothesis) is an entire field.
You need to search in both likely and unlikely places. This is pretty common in high dimensional search spaces. Searching only in the most likely places gets you stuck in local minima
As a scientist, the two links you provided are severely lacking in utility.
The first developed a model to calculate protein function based on DNA sequence - yet provides no results of testing of the model. Until it does, it’s no better than the hundreds of predictive models thrown on the trash heap of science.
The second tested a models “ability to predict neuroscience results” (which reads really oddly). How did they test it? Pitted humans against LLMs in determining which published abstracts were correct.
Well yeah? That’s exactly what LLMs are good at - predicting language. But science is not advanced by predicting which abstracts of known science are correct.
It reminds me of my days in working with computational chemists - we had an x-ray structure of the molecule bound to the target. You can’t get much better than that at hard, objective data.
“Oh yeah, if you just add a methyl group here you’ll improve binding by an order of magnitude”.
So we went back to the lab, spent a week synthesizing the molecule, sent it to the biologists for a binding study. And the new molecule was 50% worse at binding.
And that’s not to blame the computation chemist. Biology is really damn hard. Scientists are constantly being surprised at results that are contradictory to current knowledge.
Could LLMs be used in the future to help come up with broad hypotheses in new areas? Sure! Are the hypotheses going to prove fruitless most of the time? Yes! But that’s science.
But any claim of a massive leap in scientific productivity (whether LLMs or something else) should be taken with a grain of salt.
> Finding patterns in large datasets is one of the things LLMs are really good at.
Where by "good at" you mean "are totally shit at"?
They routinely hallucinate things even on tiny datasets like codebases.
I don't follow the logic that "it hallucinates so it's useless". In the context of codebases I know for sure that they can be useful. Large datasets too. Are they also really bad at some aspects of dealing with both? Absolutely. Dangerously, humorously bad sometimes.
But the latter doesn't invalidate the former.
> I don't follow the logic that "it hallucinates so it's useless".
I... don't even know how to respond to that.
Also. I didn't say they were useless. Please re-read the claim I responded to.
> Are they also really bad at some aspects of dealing with both? Absolutely. Dangerously, humorously bad sometimes.
Indeed.
Now combine "Finding patterns in large datasets is one of the things LLMs are really good at." with "they hallucinate even on small datasets" and "Are they also really bad at some aspects of dealing with both? Absolutely. Dangerously, humorously bad sometimes"
Translation, in case logic somehow eludes you: if an LLM finds a pattern in a large dataset given that it often hallucinates, dangerously, humorously bad, what are the chances that the pattern it found isn't a hallucination (often subtle one)?
Especially given the undeniable verifiable fact that LLMs are shit at working with large datasets (unless they are explicitly trained on them, but then it still doesn't remove the problem of hallucinations)
I made a toy order item cost extractor out of my pile of emails. Claude added confidence percentage tracking and it couldn't be more useless.
This is what Yan Le Cun means when he talks about how research is at a dead end at the moment with everyone all in on LLMs to a fault
I'm just a noob but lecun seems obsessed with the idea of world models, which I assume means a more rigorous physical approach, and I don't understand (again, confused noob here) how are t would help precise abstract thinking.
LLMs do typically encode a confidence level in their embeddings, they just never use it when asked. There were multiple papers on this a few years back and they got reasonable results out of it. I think it was in the GPT3.5 era though
Can't LLMs be fed the entire corpus of literature to synthesise (if not "insight") useful intersections? Not to mention much better search than what was available when I was a lowly grad...
I use Gemini almost obsessively but I don't think feeding the entire corpus of a subject would work great.
The problem is so much of consensus is wrong and it is going to start by giving you the consensus answer on anything.
There are subjects I can get it to tell me the consensus answer then say "what about x" and it completely changes and contradicts the first answer because x contradicts the standard consensus orthodoxy.
To me it is not much different than going to the library to research something. The library is not useless because the books don't read themselves or because there are numerous books on a subject that contradict each other. Gaining insight from reading the book is my role.
I suspect much LLM criticism is from people who neither much use LLMs nor learn much of anything new anyway.
I never suggested I want an LLM to be the definitive answer to a question but I'm certain that there are a lot of low hanging fruit across disciplines where the limit is the awareness of people in one field of the work of another field, and the limiting factor was the friction in discovery - I can't see how a specialised research tool powered by LLMs and RAG wouldn't be a net gain for research if only to generate promising new leads.
Throwing compute to mine a search space seems like one of the less controversial ways to use technology...