it can actually solve problems though, its not just an illusion of intelligence if it does the stuff we considered mere years ago sufficient to be intelligent. But you and others keep moving the goalposts as benchmarks saturate, perhaps due to a misplaced pride in the specialness of human intelligence.
I understand the fear, but the knee jerk response “its just predicting the next token thus could never be intelligent” makes you look more like a stochastic parrot than these models are.
It solves problems because it was trained with the solutions to these problems that have been written down a thousand times before. A lot of people don't even consider the ability to solve problems to be a reliable indicator of human intelligence, see the constantly evolving discourse regarding standardized tests.
Attempts at autonomous AI agents are still failing spectacularly because the models don't actually have any thought or memory. Context is provided to them via prefixing the prompt with all previous prompts which obviously causes significant info loss after a few interaction loops. The level of intellectual complexity at play here is on par with nematodes in a lab (which btw still can't be digitally emulated after decades of research). This isn't a diss on all the smart people working in AI today, bc I'm not talking about the quality of any specific model available today.
You're acting like 99% of humans aren't very much dependent on that same scaffolding. Humans spend 12+ years in school, their brains being hammered with the exact rules of math, grammar, and syntax. To perform our jobs, we often consult documentation or other people performing the same task. Only after much extensive, deep thought can we extrapolate usefully beyond our training set.
LLM's do have memory and thought. I've invented a few somewhat unusual games, described it to Sonnet 3.5 and it reproduces it in code almost perfectly. Likewise its memory has been scaling. Just a couple years ago context windows were 8000 tokens maximum, now they're reaching the millions.
I feel like you're approaching all these capabilities with a myopic viewpoint, then playing semantic judo to obfuscate the nature of these increases as "not counting" since they can be vaguely mapped to something that has a negative connotation.
>A lot of people don't even consider the ability to solve problems to be a reliable indicator of intelligence
That's a very bold statement, as lots of smart people have said that the very definition of intelligence is the ability to solve problems. If fear of the effectiveness of LLM's in behaving genuinely intelligently leads you to making extreme sweeping claims on what intelligence doesn't count as, then you're forcing yourself into a smaller and smaller corner as AI SOTA capabilities predictably increase month after month.
The "goalposts" are "moving" because now (unlike "mere years ago") we have real AI systems that are at least good enough to be seriously compared with human intelligence. We aren't vaguely speculating about what such an AI system might be like^[1]; we have the real thing now, and we can test its capabilities and see what it is like, what it's good at, and what it's not so good at.
I think your use of the "goalposts" metaphor is telling. You see this as a team sport; you see yourself on the offensive, or the defensive, or whatever. Neither is conducive to a balanced, objective view of reality. Modern LLMs are shockingly "smart" in many ways, but if you think they're general intelligence in the same way humans have general intelligence (even disregarding agency, learning, etc.), that's a you problem.
^[1] I feel the implicit suggestion that there was some sort of broad consensus on this in the before-times is revisionism.
> but if you think they're general intelligence in the same way humans have general intelligence (even disregarding agency, learning, etc.), that's a you problem.
How is it a me problem? The idea of these models being intelligent is shared with a large number of researchers and engineers in the field. Such is clearly evident when you can ask o1 some random completely novel question about a hypothetical scenario and it gets the implication you're trying to make with it very well.
I feel that simultaneously praising their abilities while claiming that they still aren't intelligent "in the way humans are" is just obscure semantic judo meant to stake an unfalsifiable claim. There will always be somewhat of a difference between large neural networks and human brains, but the significance of the difference is a subjective opinion depending on what you're focusing on. I think it's much more important to focus on the realm of "useful, hard things that are unique to intelligent systems and their ability to understand the world" is more important than "Possesses the special kind of intelligence that only humans have".
> I think it's much more important to focus on the realm of "useful, hard things that are unique to intelligent systems and their ability to understand the world" is more important than "Possesses the special kind of intelligence that only humans have".
This is a common strawman that appears in these conversations—you try to reframe my comments as if I'm claiming human intelligence runs on some kind of unfalsifiable magic that a machine could never replicate. Of course, I've suggested no such thing, nor have I suggested that AI systems aren't useful.
I understand the fear, but the knee jerk response “its just predicting the next token thus could never be intelligent” makes you look more like a stochastic parrot than these models are.