Reasoning LLMs are still fundamentally flawed. The flaw may surface less often, but it does. I don’t use any LLM day to day, but from time to time I try to see if they got better. Yesterday I tried to find out how to connect an ipv4+ipv6 vps to a wirguard vpn that’s only on ipv4. GPT 4o-mini (AFAIK, reasoning) created an invalid response and when I asked for sources it gave me 4: 2 were general WG tutorials and 2 were 404s.
I fail to see any real progress in the field. It’s cheaper - sure. But still not production ready.
I’m not sure “it can get the answer wrong” is a solvable problem. Even humans can get the answer wrong. The best we can maybe do is tune the personality to be a bit less confident…
Oh look, a blog on reasoning LLMs, written by an expert from a different domain. Let’s give it a chance, maybe it will not be yet another room temperature take .. oh.
Totally agree and it’s bigger than just LLMs. Consciousness is not unique or as complicated as we had hoped. I use hope because we keep clutching onto the idea that consciousness is mysterious. All we need to do is merge the two modes, training and inference, then remove the human from the equation. That’s it.
“whatever is may be” is exactly the mysticism I’m talking about. There’s nothing special about consciousness, we are just hopelessly biased by our own egos.
Cognitive science was always one of my pet subjects. So, there was this book I read, years back: The User Illusion. And then there was this other popular book, which presaged the Deep Learning explosion: On Intelligence. Long before that, Daniel Dennett’s Consciousness Explained got a lot of attention.
I don’t want to spoil any of these for anybody, but I suggest checking out the responses that their arguments received in serious academic reviews. Might learn something!
All we need to do is merge the two modes, training and inference, then remove the human from the equation.
“Look, all we need to do is merge the two modes, tractor beams and faster-than-light drives, and then we could build the Starship Enterprise. That’s it.”
No, consciousness is a different phenomenon, and we don’t even have a robust definition of what would count as consciousness in non-humans.
We recycle terms of human cognition for behaviors we observe in ML models, but that doesn’t mean they’re the same behaviors. It doesn’t really matter whether that’s true “reasoning” like humans do, we just need some word for the extra output they generate that improves their ability to find the correct answers.
The question of whether a computer can think is no more interesting than the question of whether a submarine can swim — Dijkstra
Well, at least you’re not even wrong! I don’t see how you get that from what is posted except for the overloading of terms between LLMs and neuroscience or cognitive research or something. We have no idea what consciousness is.
Reasoning LLMs are still fundamentally flawed. The flaw may surface less often, but it does. I don’t use any LLM day to day, but from time to time I try to see if they got better. Yesterday I tried to find out how to connect an ipv4+ipv6 vps to a wirguard vpn that’s only on ipv4. GPT 4o-mini (AFAIK, reasoning) created an invalid response and when I asked for sources it gave me 4: 2 were general WG tutorials and 2 were 404s.
I fail to see any real progress in the field. It’s cheaper - sure. But still not production ready.
I’m not sure “it can get the answer wrong” is a solvable problem. Even humans can get the answer wrong. The best we can maybe do is tune the personality to be a bit less confident…
I need my tools to be more reliable than humans, and I don’t work with humans who spout bullshit.
Lucky
Oh look, a blog on reasoning LLMs, written by an expert from a different domain. Let’s give it a chance, maybe it will not be yet another room temperature take .. oh.
Totally agree and it’s bigger than just LLMs. Consciousness is not unique or as complicated as we had hoped. I use hope because we keep clutching onto the idea that consciousness is mysterious. All we need to do is merge the two modes, training and inference, then remove the human from the equation. That’s it.
I am not into mysticism but I don’t think what LLMs do is remotely close to consciousness whatever it may be.
“whatever is may be” is exactly the mysticism I’m talking about. There’s nothing special about consciousness, we are just hopelessly biased by our own egos.
Admitting something is not properly understood at the moment is not mysticism
Cognitive science was always one of my pet subjects. So, there was this book I read, years back: The User Illusion. And then there was this other popular book, which presaged the Deep Learning explosion: On Intelligence. Long before that, Daniel Dennett’s Consciousness Explained got a lot of attention.
I don’t want to spoil any of these for anybody, but I suggest checking out the responses that their arguments received in serious academic reviews. Might learn something!
“Look, all we need to do is merge the two modes, tractor beams and faster-than-light drives, and then we could build the Starship Enterprise. That’s it.”
No, consciousness is a different phenomenon, and we don’t even have a robust definition of what would count as consciousness in non-humans.
We recycle terms of human cognition for behaviors we observe in ML models, but that doesn’t mean they’re the same behaviors. It doesn’t really matter whether that’s true “reasoning” like humans do, we just need some word for the extra output they generate that improves their ability to find the correct answers.
This is exactly my argument, it isn’t a different phenomenon.
You’re not making an argument, you’re just saying a ridiculous thing without any concrete evidence.
Well, at least you’re not even wrong! I don’t see how you get that from what is posted except for the overloading of terms between LLMs and neuroscience or cognitive research or something. We have no idea what consciousness is.