ETA:
You updated your comment, which is fine but I wanted to reply to your points.
> I would argue that LLMs are actually smarter than the majority of humans right now. LLMs do not have quite the agency that humans have, but their intelligence is pretty decent.
I would actually argue that they are decidedly not smarter than even dumb humans right now. They're useful but they are glorified text predictors. Yes, they have more individual facts memorized than the average person but that's not the same thing; Wikipedia, even before LLMs also had many more facts than the average person but you wouldn't say that Wikipedia is "smarter" than a human because that doesn't make sense.
Intelligence isn't just about memorizing facts, it's about reasoning. The recent Esolang benchmarks indicate that these LLMs are actually pretty bad at that.
> We don't have clear ASI yet, but we definitely are in a AGI-era.
Nah, not really.
There is a long history of people arguing that intelligence is actually the ability to predict accurately.
https://www.explainablestartup.com/2017/06/why-prediction-is...
> Intelligence isn't just about memorizing facts, it's about reasoning.
Initially, LLMs were basically intuitive predictors, but with chain of thought and more recently agentic experimentation, we do have reasoning in our LLMs that is quite human like.
That said, there is definitely a biased towards training set material, but that is also the case with the large majority of humans.
For the Esoland benchmarks, I would be curious how much adding a SKILLS.md file for each language would boost performance?
I am pretty confidence that we are in the AGI era. It is unsettling and I think it gives people cognitive dissonance so we want to deny it and nitpick it, etc.
There sure is, and in psychological circles that it appears that there's an argument that that is not the case.
https://gwern.net/doc/psychology/linguistics/2024-fedorenko....
> Initially, LLMs were basically intuitive predictors, but with chain of thought and more recently agentic experimentation, we do have reasoning in our LLMs that is quite human like.
If you handwave the details away, then sure it's very human like, though the reasoning models just kind of feed the dialog back to itself to get something more accurate. I use Claude code like everyone else, and it will get stuck on the strangest details that humans actively wouldn't.
> For the Esoland benchmarks, I would be curious how much adding a SKILLS.md file for each language would boost performance?
Tough to say since I haven't done it, though I suspect it wouldn't help much, since there's still basically no training data for advanced programs in these languages.
> I am pretty confidence that we are in the AGI era. It is unsettling and I think it gives people cognitive dissonance so we want to deny it and nitpick it, etc.
Even if you're right about this being the AGI era, that doesn't mean that current models are AGI, at least not yet. It feels like you're actively trying to handwave away details.
Much of our reasoning is based on stimulating our sensory organs, either via imagination (self-stimulation of our visual system) or via subvocalization (self-stimulation of our auditory system), etc.
> it will get stuck on the strangest details that humans actively wouldn't.
It isn't a human. It is AGI, not HGI.
> It feels like you're actively trying to handwave away details.
Maybe. I don't think so though.
Personally, I've used LLMs to debug hard-to-track code issues and AWS issues among other things.
Regardless of whether that was done via next-token prediction or not, it definitely looked like AGI, or at least very close to it.
Is it infallible? Not by a long shot. I always have to double-check everything, but at least it gave me solid starting points to figure out said issues.
It would've taken me probably weeks to find out without LLMd instead of the 1 or 2 hours it did.
In that context, I have a hard time thinking how would a "real" AGI system look like, that it's not the current one.
Not saying current LLMs are unequivocally AGI, but they are darn close for sure IMO.
Why is it that LLMs could ace nearly every written test known to man, but need specialized training in order to do things like reliably type commands into a terminal or competently navigate a computer? A truly intelligent system should be able to 0-shot those types of tasks, or in the absolute worst case 1-shot them.
Being able to actually reason about things without exabytes of training data would be one thing. Hell, even with exabytes of training data, doing actual reasoning for novel things that aren't just regurgitating things from Github would be cool.
Being able to learn new things would be another. LLMs don't learn; they're a pretrained model (it's in the name of GPT), that send in inputs and get an output. RAGs are cool but they're not really "learning", they're just eating a bit more context in order to kind of give a facsimile of learning.
Going to the extreme of what you're saying, then `grep` would be "darn close to AGI". If I couldn't grep through logs, it might have taken me years to go through and find my errors or understand a problem.
I think that they're ultimately very neat, but ultimately pretty straightforward input-output functions.
Well, I guess you lose artificial if there’s a human brain hidden in the box.
I can't argue that LLMs do not know an absolute insane amount of information about everything. But you can't just say LLMs are smarter then most humans. We've already decided that smartness is not about how much data you know, but thinking about that data with logical reasoning. Including the fact it may or may not be true.
I can run a LLM through absolutely incorrect data, and tell it that data is 100% true. Then ask it questions about that data and get those incorrect results as answers. That's not easy to do with humans.
This (surprisingly common) view belies a wild misunderstanding of how LLMs work.
Would they? Perhaps if you only showed them glossy demos that obscure all the ways in which LLMs fail catastrophically and are very obviously nowhere even close to AGI.
Certainly, they wouldn't expect that an AI able to score 150 on an IQ test is unable to play a casual game of chess because it isn't coherent enough to play without making illegal moves.
To be fair, I am pretty sure Claude Code will download and run stockfish, if you task it to play chess with you. It's not like a human who read 100 books about chess, but never played, would be able to play well with their eyes closed, and someone whispering board position into their ear
5 years ago we thought that language is the be-all and end-all of intelligence and treated it as the most impressive thing humans do. We were wrong. We now have these models that are very good at language, but still very bad at tasks that we wrongly considered prerequisites for language.
ChatGPT Health failed hilariously bad at just spotting emergencies.
A few weeks ago most of them failed hilariously bad at the question if you should drive or walk to the service station if you want to wash your car
The second question sounds like a useless and artificial metric to judge on. The average person might miss such a “gotcha” logical quiz too, for the same reason - because they expect to be asked “is it walking distance.”
No one has ever relied on anyone else’s judgment, nor an AI, to answer “should I bring my car to the carwash.” Same for the ol’ “how many rocks shall I eat?” that people got the AI Overview tricked with.
I’m not saying anything categorically “is AGI” but by relying on jokes like this you’re lying to yourself about what’s relevant.
In my experience, they contain more information than any human but they are actually quite stupid. Reasoning is not something they do well at all. But even if I skip that, they can not learn. Inference is separate from training, so they can not learn new things other than trying to work with words in a context window, and even then they will only be able to mimic rather than extrapolate anything new.
It's not the lack of perfect, it's the lack of reasoning and learning.
I've seen a lot of reasoning in the latest models while engaging in agentic coding. It is often decent at debugging and experimentational, but around 30% it goes does wrong paths and just adds unnecessary complexity via misdiagnoses.
I consider myself a bit of a misanthrope but this makes me an optimist by comparison.
Even stupid people are waaaaaay smarter than any LLM.
The problem is the continued habit humans have of anthropomorphizing computers that spit out pretty words. It’s like Eliza only prettier. More useful for sure. Still just a computer.
I don't believe in a separation of mind and spirit. So I do think fundamentally, outside of a reliance on quantum effects in cognition (some of theorized but it isn't proven), its processes can be replicated in a fashion in computers. So I think that intelligence likely can be "just a computer" in theory and I think we are in the era where this is now true.
This doesn't mean they aren't useful, I like Claude a lot, but I don't buy that it's AGI.