I think the easiest way to get the gist of what he (and I) are arguing is a thought experiment. Take an LLM back to the dawn of mankind, and train it on all of the collective knowledge of man. This would actually be impossible, because language hadn't yet been invented, so let's just imagine it could somehow read the collective mind of humanity. So cutting edge technology is poke 'em with the pointy side, science is nonexistent, and knowledge isn't that far behind.
Now run this system through our LLM for as long as you want with as much power as you want. Where is it going to take you? Nowhere really, it's still just going to be constrained to its pool of data, recombining things in primitive ways and be essentially permanently stuck in the present, until somebody gives it some new data to train on it and mix and match. Yet somehow humanity, starting from that exact same basis, would soon (relatively speaking) put a man on the Moon, unlock the secrets of the atom, discover quantum mechanics, invent/discover mathematics, and much more.
This is what is fundamentally meant by LLMs cannot create "new" knowledge. They absolutely can mix and match their pool of training statements in ways that can generate new statements. But if it's not an extremely simple 'remix' and we're in a domain where there are right and wrong answers, there's good chance it's just a nonsensical hallucination.
You are comparing a species going from developing language to building spaceships on the order of hundreds of thousands of years, to a technology that was invented 7 years ago and only started to really be taken seriously 4 years ago. Working memory systems required for learning and iteratively developing ideas are in their infancy (on the scale of recent developments), but the recall technology (vector databases a la RAG) is quite well proven. I see no reason that a language model couldn't do iterative science in the same way humans have been with the resources available to it (APIs) using the composition of current technologies.
The current technology for LLMs is still a rather complex guess the next word algorithm. But this leaves the software with no real room to ever move beyond its training, regardless of how much time or training you give it. You could give them literally infinite processing power and they would not suddenly start developing new meaningful knowledge - it would still be little more than simple recombinations of its training dataset, until somebody gave it something new to train upon.
> I see no reason that a language model couldn't do iterative science in the same way humans have been with the resources available to it (APIs) using the composition of current technologies.
I think this belies a common refrain from people in non-tech feeling like people in tech tend to oversimplify their fields and claim they can "fix any problem" without any actual specific knowledge in that field. This just feels like the next iteration of that.
This seems like a lot of words to just say "you are wrong". Can you explain why? What features of the scientific method cannot be achieved through the composition of existing technologies? Feel free to be specific and use technical terms.
I think you're misreading this. I'm not saying "you are wrong". I'm saying there's a common trap that a lot of people in tech (and therefore a lot of people on HN fall into), where they believe that they can solve any problem from what they know of basic tech problems.
> What features of the scientific method cannot be achieved through the composition of existing technologies?
^This statement, sums up the above trap perfectly. "I know how to create a workflow, therefore, I know how to create a workflow of the scientific method, therefore I can replace a scientist with an AI robot."
But okay, how do we know what quantitatively, is a significant result? R=0.05 is the most common cut-off for this, but like, that's a number we picked, and isn't even agreed on.
LLMs can absolutely create synthesize new knowledge out of existing knowledge. They can't easily do so iteratively because we haven't quite figured out memory yet. Until we figure that out you won't have an LLM discover a new theory of quantum gravity.
And even once we solve that, LLMs - just like human scientists - absolutely need new data from the outside world. Very few breakthroughs were achieved by just thinking about it long and hard, most were the the result of years of experimentation. Something LLMs simply can't do
Recursive self improvement requires the ability to objectively select the "prime" of your own knowledge. From an LLM's perspective a hallucination and a correct answer are the same thing. It does not have any beliefs about what is true or false, because it has no concept of what it's even outputting. It's all just guess the next word. So even if the hallucination completely contradicts countless things it ostensibly "knows" to be true, it is unable to correct itself or realize that what it's outputting is unlikely to be correct.
Pause tokens [1] are directly dependent on the model recognizing whether the answer it arrived at is one it should "commit" to or whether it should abandon it and output a pause token instead.
Similarly if you ask ChatGPT about the current president of Numbitistan it will tell you that it doesn't know about a county with that name, rather than just hallucinating an answer. So it can at least in this circumstance tell the difference between knowing something and not knowing something.
I don't think this is true - the main difference is internal consistency. Humans adopt a series of views and values, true or false, and tend build up from those. The accuracy doesn't really matter so much as the internal consistency, because it tends to turn out that trying to build from an invalid foundation eventually causes you to stop moving forward, and so the more factually supported values tend to win out over time.
But it's the internal consistency that really matters. LLMs have no internal consistency, because they have no way of 'adopting' a view, value, fact, or whatever else. They will randomly hallucinate things that directly contradict the overwhelming majority of their state, and then do so repeatedly in a single dialogue. If there were a human behaving in such a fashion, we would generally say they had schizophrenia or some other disorder that basically just ruins your ability to think like a human.
The biggest mistake I see people make when criticizing LLMs is that they take the best possible modes of human thought from our best thinkers, and compare that to LLM edge cases.
Accuracy vs consistency isn't really a delineator. There's so much low-hanging fruit atm, like world models for LLMs improving drastically if you just train them longer. I'll believe the naysayers if say in 5 years GPT-4 is still near state of the art. Until then, there doesn't seem to actually be any theoretical limitations.
Hallucination is not an LLM "edge case." It is their normal and only state of operation. It just so happens that 'guess the next word' algorithms are capable of a reasonable frequency of success owing to the fact that a lot of our language is probably mostly redundant, making it possibly 'hallucinate' reasonable statements quite regularly.
Take what I wrote above. If you were given the context of what I have already written, then you could probably fill in most of what I wrote, to a reasonable degree of accuracy, after "It is their normal..." Because the issue is obvious and so my argument largely writes itself. To some degree even this second paragraph does.
IDK, I think it's kinda important for LLMs to get the simple stuff correct in order to justify looking into the rest of the hallucinations.
Like, if you can't tell me "what day is it today?" (actual failed prompt I have seen) then there's no world where I'm going to have a more complicated follow-up conversation with you. It's just not worth my time or yours.
I agree with this, but find it a poor mode of debate. Because it results in a hole-plugging which is then called goal shifting, even though it's not - but rather a lack of precision in the goal to begin with. For example imagine it goes viral that 'wow look LLMs can't even play a half decent game of chess.' So OpenAI or whoever decides to dedicate an immense amount of post-training, hard-coding, and other fun stuff to enabling LLMs to play a decent game of chess.
But has anything changed? Well no, because it's obviously trivially possible for them to play a decent game of chess (or correctly assess the date), but it's an example of a more general issue of LLMs being generally incapable of consistently engaging in simple tasks across arbitrary domains. So you have software that can score some high thing on the LSAT or whatever, but can't competently engage in a game children can play.
The over-specialization for the sake of generating headlines and over-fitting benchmarks is, IMO, not productive. At least not in terms of creating optimal systems. If the goal is to generate money, which I guess it is, then it must be considered productive.
I'm not asking for someone to overfit to being able to properly answer the question of "What day is it today?" I'm giving an example of a simple question that all LLMs need to be able to answer correctly.
But like, people are on here saying that this will make scientific improvements, and until it can get past the basic stuff, it's not in the ballpark of anything more complicated. Right now, we're basically at the stage of 10 million monkeys on 10 million typewriters for 10 million hours. Like, maybe we'll get Shakespeare out of it, but are we willing to sort through all of the crap it will generate along the way, when it can't actually create a useful answer to simple questions?
Why are humans capable of doing that, but it's categorically impossible for LLMs? That's a very high level capability. What's the primitive it is built on that humans have but machines can't have?
Quantum biomechanics. Human brain has built-in indeterministic pattern seeking. Computers are Turing based which 100% dependent on their programming and data inputs. All current LLMs operates on that very deterministic hardware. In order for LLMs to break thru existing glass ceiling, the hardware need to change.
There's more to AI than LLMs though, like Alpha Fold figuring the structure of 200 million proteins for example. You can link those systems to your LLM so you can ask it what do you think of this protein and what experiment could I do on it and so on. There are many such avenues to explore. I don't see things getting stuck at a maxima for quite a while.
Explaining a point is part of an argument. And the premise is unsound and illogical and if we had an LLM firewall their post would have been blocked until they cleaned up their reasoning.
Now run this system through our LLM for as long as you want with as much power as you want. Where is it going to take you? Nowhere really, it's still just going to be constrained to its pool of data, recombining things in primitive ways and be essentially permanently stuck in the present, until somebody gives it some new data to train on it and mix and match. Yet somehow humanity, starting from that exact same basis, would soon (relatively speaking) put a man on the Moon, unlock the secrets of the atom, discover quantum mechanics, invent/discover mathematics, and much more.
This is what is fundamentally meant by LLMs cannot create "new" knowledge. They absolutely can mix and match their pool of training statements in ways that can generate new statements. But if it's not an extremely simple 'remix' and we're in a domain where there are right and wrong answers, there's good chance it's just a nonsensical hallucination.