> The Jacobian conjecture is notorious for the large number of published and unpublished proofs that turned out to contain subtle errors.
https://en.wikipedia.org/wiki/Jacobian_conjecture#cite_note-...
There's no shortage of folks doing that in software, right now.
For someone that use Claude Code every day, this is obvious, but for some reason many scientists refuse to accept that it's truly reasoning; perhaps not in the human sense, but in a very profound and real sense. These powerful results are devastating to their point of view.
I can sympathize, because I too called LLMs "fancy Markov chains" in the GPT 3 era. But there comes a time where you have to update your world view to match reality, or be stranded in fantasy land.
Similarly I've run into several "you just gotta know" type problems where LLMs still just fail. My favourite was a quirk in how async relationships work inside emberjs. Three different models gave a variation of the same incorrect answer. When I searched myself, I initially came up with nothing and eventually found what I think might be the only example of the same issue on the internet, a single stack overflow question wth two responses. The first is what Claude, gemini and chatGPT said, the second response was the OP saying it was wrong. I asked in the Ember discord and a core team member responded instantly with the answer.
There's a video I love on Youtube, where a guy uses ML to assemble a blank jigsaw. It performs amazingly, the jigsaw being blank is of no consequence and if it did have an image, it'd perform worse. That's all LLMs do. Just because the jigsaw pieces are smaller, they're still just getting assembled in whatever way fits, there's no mechanic for interpreting the image on the front.
Why do LLMs still have trouble on floating point math without forking out to a tool, but they can perform symbolic manipulation just fine?
Because symbolic manipulation is just rote work and textual stepping through symbols. A side poster commented on the number of prior attempts on this problem which were close, but not quite.
Starting from a known "close" solution (which this did), and using exploration to search around the space is exactly something an LLM would and could be good at (clearly).
The "transformer LLMs are next-token predictors with some in-GPU processing of bounded complexity with respect to token count" remains undefeated. Both because that is mathematically what they are, and also because we don't have counterexamples to that effect that don't require some higher-order tooling wrapping the systems.
Symbol manipulation is what LLMs are good at.
What still confuses people is the insane inefficiency of deep learning, and that those emergent capabilities require such an immense training corpus compared to the only other architecture that we know of.
But this already is an optimization problem. If the machine gets super human at symbolic reasoning, and at the same time, can solve the symbol grounding problem to real world data and sensors, what prevents you from saying it thinks? Can it not solve real world problems? Can it not redefine its tasks and display some form moral agency - even if a totally foreign morality for us humans? Can it not use these abilities to reproduce and expand, create ships and turn the universe into paperclips, if it finds it worthwhile?
Math is basically just a playground that is perfectly suited for these emergent capabilities, so of course we will see the first progress here; but there is no firewall separating math problems from general cognition.
And the "insane inefficiency" of deep learning is fully to be expected from how it works. As well, there are provably no—literally no—emergent properties in these models. The choice of metric was a convenient, sloppy, and embarrassing fault of the field. It should be discredited; the field should be embarrassed; expectations on messaging should have changed; and it did not.
Why? Because the industry is full of charlatans, and this is a highly profitable enterprise telling people that this would lead to AGI.
Multiply two floating point numbers without a tool call. Still can't, because it's a curve fit.
So, in summary, nothing you just said is relevant. There are no emergent capabilities, simply 1) search, + 2) the original set of learned feature vectors from throwing tons of data at this.
I mean, lol.
We can replace multiplication with any class of problems which should go from 0->100% solution almost immediately if there was actually a concept learned.
There is not. Because they are plain ol' fits. And there are no "emergent" features that pop out without having a sufficient set, where "sufficient" is absolutely gigantic and equivalent to memorizing enough of the space to compress the problem. LLMs are Rain Man.
They interpolate within a known distribution. Search allows places outside of distribution to be explored.
This paper should be required reading [1]. You can explore the curves yourself. You can see exactly what it's doing. And you also have this nagging thing—which you know and I know—that all these models converge and do not diverge upwards. An "emergent" "hyperintelligence"—a characteristic that could be found if something was actually learned and combined with a new concept—would not have this problem.
Exponentials on exponentials added to compute and data and the problem classes still sit at not great places, and require agents, feedback loops, and trial and error to solve. The models are the problem, but more importantly, the people selling things these models could never do are the problem.
[1] https://hai.stanford.edu/news/ais-ostensible-emergent-abilit...
Edit: It should be mentioned, if there's some scary neural architecture that's super-de-duper and doing something beyond the very obvious next string prediction that LLMs clearly do, it can't do what absolutely ancient ML models could do; a network to multiply two floating point numbers should pop out somewhere without symbolic computation, no?
It does not. There's no magic other than the run-of-the-mill SV fake it til you make it magic. And that magic has failed.
So you are arguing here that LLMs should just "know" the result of a multiplication when the operands are in context, ie, that a hidden multiplier circuit should emerge in their weights.
Is this how you do multiplication, if I give you two 12 digit numbers, does the 24 digit multiplication result just pop in your head? Don't you have to follow a learned algorithm through a tedious system 2 effort? Don't you need to write down the results on paper because you can't actually hold in your head the dozen partial results, each with a dozen digits? How many visual, tactile and reasoning tokens does this consume, moving the hundreds of muscles that make up your hand to draw each number under visual feedback, then reading all those numbers back and transforming ocular activation data into numeric symbols?
It seems to me your system 2 is just following a symbolic algorithm for multiplication, and it does roughly the same steps as the LLM trace I showed you previously.
So if you think this is the mark of your intelligence, why wouldn't it apply to the machine too? Why is it implausible that, following a similar algorithm learned from some mathematical paper, the LLMs has reasoned a new solution to a problem in another math field? Why couldn't the machine combine and morph these algorithms for symbolic manipulation, to yield entirely new and original results? How would those results differ from results human mathematicians generate, using recipes they learned in university?
The emergent behavior that we talk about isn't that the machine can do multiplication in its "head" after knowing the multiplication algorithm. What emerges is the ability to follow any other algorithm, even algorithms that were not in the training set, even algorithms to create other algorithms, which it then executes. This is the emergent behavior that matters for AGI; once it can do that, it's a trivial exercise to create a non-AI tool to automate and accelerate the mechanical tasks - just like we humans do it.
And yes, if "emergence" and the "there's something else going in there!" ridiculousness was true, there would be some other architecture which can do more than its symbol prediction.
Everyone can see the code. Everyone can see the math. Everyone can see the failures of the claims. And there's now about 5 years of everyone watching the lies come undone one-by-one.
> ((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3
I'm yet to see anyone cite the prior work. If it's true I think some credit is due for the authors Fable/Sol are branching from.
This specific counterexample really is trivial. There's nothing to cite. People have wasted hours and hours on a question whose answer you could give as a homework problem in Calc II.