Overall, this is an impressive proof of capability. But I wouldn't take that proof as anything more than what it is.
If I'm getting paid for the work, I'm happy to leverage the LLMs so I can do more. If I'm paying for the work, I expect more from it.
For hobbyist stuff, where I'm not expecting to receive money? LLMs let me do things I otherwise wouldn't have done.
I've agreed with thoughts like "the LLM wrote the code, surely it's not worth sharing" or "I could just have the LLM write my own version of that". I'd also wondered about my own personal projects, surely "an LLM could have written all of this". -- But how I feel about that changes a bit based on how much it'd take the LLMs to get the same output.
Of course money in this situation is a bit of a funny measurement, right, because if I was able to take the rest of the week off as soon as I had solved the one-week problem, then I would have no problem at all throwing even $100 worth of tokens at it, so I could enjoy a nice 4-day "mini-vacation".
How cheap "cheap" is, is indeed "in the eye of the beholder".
In Scrum terms my personal velocity grows by a factor of four or more with access to agentic AI workload, but if it means that I will just be asked to "consume" 4*X more Story Points per sprint, I'm not the winner in the end, my employer is. If they asked me to complete X Story Points per sprint regardless of my velocity, and they let me take the days off when I was done, I would be the winner. But that's not how it works.
AI is "Cheap" for the person/organization that gets more product for less money, not for the individual person building the product faster.
it went from not having a price, to having one, and we are trying to retroactively transpose economic viability or economic existence to it from some parallel and prior time.
But trying to maintain this distinction leads to insuperable difficulties. Our conceptual framework for understanding the world are always value-laden. There is no "view from nowhere", no historically unconditioned set of values or concepts. Your framing, in which "values" are external to "intelligence" and must be imposed on it (on pain of intelligence being "value-neutral"), leads inevitably to the dead end of "AI Alignment", "superintelligence", etc. Which is a kind of pseudo-theology.
"We humans better [be] refocusing our energy on our core values/principles, given most of our skills are becoming irrelevant."
In light of the untenability of a strong fact/value or intelligence/ethics distinction, I would suggest this alternative advice: humans should focus on critical appropriation and extension of the received wisdom, whether that comes to us directly from human beings or indirectly through an LLM. Perhaps this is compatible with the spirit of your original suggestion.
They will, however, get there as well either directly or as interfaces to models that do, and your core point stands.
If there was a deep fundamental inability, we wouldn't see things like newer generations of LLMs consistently improving on ARC-AGI series (heavy spatial reasoning loading) and SimpleBench (a lot of commonsense + spatial reasoning components).
In a way, it's a surprise that LLMs, notoriously lacking any sort of embodied experience, can even get this close to human baselines on tasks like this.
My takeaway is that text is a far richer modality than anyone has expected - and that high end LLMs are often sharp and flexible enough to recognize their weak points and substitute their strengths. I.e. all the LLMs implementing A* to optimally solve pathfinding in ARC-AGI-3 tasks, often unprompted.
There might still be unrealized gains there from true depth-unbounded recurrence, or maybe from finding better ways to integrate modalities in training. But clearly, a "fundamental limit" it ain't.
Yeah, that's fair.
> My takeaway is that text is a far richer modality than anyone has expected - and that high end LLMs are often sharp and flexible enough to recognize their weak points and substitute their strengths. I.e. all the LLMs implementing A* to optimally solve pathfinding in ARC-AGI-3 tasks, often unprompted.
I agree and disagree with this. I think we've learned a lot of humans are more text based than we thought, but conversely I'm not persuaded what non-textual task reasoning LLMs are doing is necessarily text based, just that models have grown large enough for other reasoning modes to conceivably be hiding in the parameter space.
As I mentioned elsewhere, like many others I find LLMs work entirely by example, and reaching for A* when pathfinding is the single obvious thing to do. In cases where the magic key word is not mentioned and the problem cannot be identified as "pathfinding" (or some other trigger with a highly specific widely documented solution) they will struggle, yet the moment the trigger is hit they get there very fast. This is why prompting remains such an art form.
Fable is the first one I've encountered that is capable of serious open ended 3D programming in ways that suggest it has some grasp of the spatial aspects of the problem (not merely symbolic manipulation of the vectors etc.), but it still misses optimization opportunities a human will find glaringly obvious based on spatially predictable bounds etc.
Basic LLMs don't reason in text, and never did. They use it as an interface - for input, output and some of the intermediate products. Heavy use of those "pseudo-recurrence" intermediates in "reasoning models" is a relatively late post-training adaptation. But the process that happens between those endpoints is not at all text-based. What happens in the hidden dimension is part "output logit domain", tied to probability distributions over possible output tokens, and part "incomprehensible concept-space madness".
The latter being where things like latent world models live. LLMs develop partial world models, right in pre-training, despite not being explicitly forced to - because it brings them closer to heaven of accurate next token prediction.
And yes, larger models like Fable seem to be better at spatial reasoning. Maybe because their large size increases the sample efficiency and improves generalization, allowing them to absorb the sparse signal of "spatial reasoning" in the training text better. Maybe because this extra size means more layers, allowing for deeper latent space reasoning in lieu of true recurrence. Maybe because the default "next token prediction" reward underrates rare spatial reasoning challenges, and the model only starts to "get good" at them once the other sources of loss reduction are heavily depleted. Maybe because no true recurrence is suboptimal for spatial reasoning architecturally. But it is what it is. Spatial reasoning gains in LLMs are extractable, but extracting them is nontrivial.
You don't have to do much statistical analysis to figure out what is meant by the token string "cat under a tree". However you need to do an enormous amount to encode any permutation of pixels that show a cat under a tree from the set of all possible pixels arrangements that illustrate that (along with the massive fringes of ambiguity).
Basically current gen LLMs apparently do spatial reasoning the way they seemingly do everything else: by reference to previous example. I didn't see them work out which known example to use for a given problem until specifically prompted, in my case by accident.
That's not what I said happened though. It didn't solve the problem (for weeks) until I (accidentally) told it which example it happened to know was relevant, and then it solved it in hours.
Only a fraction of the games can be solved by Sol, generally at sub-human efficiency in terms of turns, AND at a cost of >$10,000 per game.
I’ve been doing more math as a hobby in the past few weeks — working on lesser-known conjectures and exploring proofs of hard theorems — than I could have managed over the previous several years. It’s an exciting time.
Stored potential.
But will that potential be converted that contributes to the economy..? That requires other traits.
Might be focus, might be discipline, might be the need to get revenge lmao.
This is what the llm-boosters miss. Progress is willed into existence.
At the end of the day it is still making a best guess at what the user wants based on data it has seen before.
It still requires someone smarter than the output to be able to evaluate if the result is any good, or just hand waving.
This is basically what LLMs do on really hard tasks. Prompt it a million times on a really hard problem and it might output the correct answer once.
Given the tokenizers have a vocabulary in the 10k-100k range, "a million attempts" will generally still only get the first token of the answer correct.
Even really rubbish models, e.g. talkie, the "what if we only use pre-1930s data to train a model?"** model, had to be almost all the way to the right answer to reach the really low HumanEval pass@100 score of ~0.04 (I'm only eyeballing the relevant chart).
* Actual monkeys not being like this is, while amusing, irrelevant
Even if every atom in the universe were a supercomputer generating a trillion trillion random characters every second since the Big Bang, the chance of producing Hamlet would still be essentially zero.
Even when you've got an interesting idea, if you're an enthusiastic amateur who don't yet know enough to phrase the question right but does actually know the basics, they'll put you in the same category as the people who think healing crystals can power hyperspace telepathy with Anubis: "oh no not another one".
LLMs have infinite patience, but unfortunately come (came?) with too much sycophancy, giving even more people far too much confidence.
AI hasn’t even taken the class of jobs associated with customer service lmao
This is what the whole https://people.csail.mit.edu/brooks/papers/elephants.pdf is about.
You mistyped it.
Best I've come up with is we'll need to be adopted by technofeudlaist overlords to be our patrons like in the roman days
Continually progressing AI (combined with our current socioeconomic systems) throws a lot of uncertainty into our mid to long term future, but I don't think this is going to be what happens.
There are billions more of "us" than of "them", people don't respond well en masse to a drastic worsening of their societal status and "they" are lagging very far behind on building their robot armies.
If we poorly navigate this transition the outcome should be worrying them more than it worries us.
Fortunately, yes, those robot armies do seem to be rather behind schedule*.
However, even if Musk dies of old age before anything like the Optimus can be connected to useful artificial intelligence, it can still be driven by the common joke a few years back that "AI" really stood for "Actually Indians".
When all the people currently upset about "immigrants coming here taking our jobs" discover those same people are now staying home and remote-working those same jobs over a VR headset and a Starlink connection... my guess is that by this point, Musk will have no political allies left.
> If we poorly navigate this transition the outcome should be worrying them more than it worries us.
It can screw everyone over. Literal communism was invented in response to Laissez-faire capitalism, and while Laissez-faire died with the Great Depression, the form of capitalism which succeeded it in the USA came into conflict with the USSR and gave us the Cold War, the Cuban Missile Crisis, etc.
* 2022 "next year": https://www.cnbc.com/2022/04/08/elon-musk-says-tesla-is-aimi...
Fwiw I was mostly joking. I agree that the techno overlords have no reason to keep us, unlike in Roman times.
Reminds me of Wigner's Unreasonable effectiveness of mathematics in natural sciences [0].
[0]: https://en.wikipedia.org/wiki/The_Unreasonable_Effectiveness...
I don’t know if LLMs will kill the working-mathematicians but at least seem like that it doesn’t seem absurd to imagine LLMs will be good at math…