The burden of proof is very much on the people who say a computer is a person (and get paid for it).

Your comment inspired me to actually listen to it and there's a lot more evidence for LLM-mindbrain convergence than I thought. Still seems like we invented something too complex to characterize directly so we're interrogating it the way we do our own cognition... I wonder what kind of biases this introduces.

The discussion is about cognition

Ah yes the foundation of "AI" research - retreating from obvious meanings to redefining words in order to make absurd claims.

We have many very plausible, evidence-driven models

"Plausible" counts for jack shit in science.

There are many models because there's no convincing evidence for any of them.

the problem is figuring out which have the best predictive power. And that problem is being very actively worked.

Yes, these models have no experimental or "predictive" use even after decades of "research".

You can look at a pair of meshed gears and understand it just as readily as you can understand a single transformation by running the math on paper with a pencil, but I don't think anyone will ever be able to comprehend the fullness of any useful-scale model for the same reason they would never comprehend a clockwork with the same number of gears as an LLM has transformers or indices. It's a non-intuitable space with too much complexity which cannot be simplified without oversimplifying.

Their areas of expertise and research are not areas that offer insight into mechanisms of cognition, they are not a neurobiologist or a computer scientist involved in machine learning. Their high standing in related areas is being used to give credence to an argument that is no better than a gut feeling.

You also have no ideas what conflicts of interest exist here, there may be none or the professor could be heavily invested in LLM companies.

LLMs dont think and the "trains of thought" is just strapping a bunch of LLMs together to try and tidy up and correct if one of the models in the chain outputs nonsense which they are wont to do. They are useful tools, they don't spontaneously generate output or request input. It's a very complicated auto complete that is capable of generating plausible valid output (when considered by humans) to a given natural language input.

What a dumb question. "Cognition" is an abstract concept that just leads to debate over its definition, except that it's commonly understood that humans have cognition.

This entire interview also seems predicated on the belief that neural networks work fundamentally the same as human brains, which is an oversimplification of brains. To begin with, neural networks cannot train themselves at inference time (context windows aren't training). Also, humanity has barely a concept of how the brain works and regularly learns new things about it. What we do know is that they are more complex than a bunch of connected neurons, and that we have no way of modelling something we don't yet fully understand.

So do neural networks have cognition? If they did, then we've had cognitive AIs for longer than I've been alive.

The rest of the interview is imaginative fiction based on what they think the model is doing predicated on what Anthropic likes to advertise their models do.

As basic evidence, if I tell you that the word "strawberry" has 3 "r"s, you can remember that. You can recall that no matter how many conversations we have in between when I told you that and when I asked you again.

If I teach you how to count how many "r"s there are in the word "strawberry", you learn a skill. In the future, if I ask how many "r"s there are in the word "strawberry", you can count them using your new skill.

If I teach a LLM how many there are or how to count the letters, it can only use that skill or knowlege if it can effectively search for it. Its ability to execute skills (which, to begin with, isn't it executing the skills but it asking something else to do them) is diminished the more skills it gains access to, and its "knowlege" (aka long-term storage) becomes diluted when it has access to more of it.

This is blatantly untrue of humans, who can process knowlege and skills they learn, efficiently and quickly search them, and actually get faster and more accurate as they develop related skills and learn related knowlege.

If i could prove it, it would be proof that we do. Our brains aren't logic machines, they're a wet mess where cognition emerges irrationally.

Can you prove that humans don’t?

This, this this. This is something the prevailing armchair opinion is really missing. Their blatantly-wrong assumptions about human cognition.

It feels as proudly ignorant as my grandmother's generation who proudly declared Evolution to be patently ridiculous ("my grerat-gramma was no monkey swingin' from a tree!")

I am not a researcher in the area, but I am the kind of nerd who watches university colloquia and sometimes does in fact read the paper.

"I don't hallucinate all the time"

YES, YES YOU DO, do you not understand how the sensory system works? You interpolate between inputs. This is why optical illusions work! Your fovea is a tiny tiny spot and is the only area that's actually in focus!

Learning by statistical sampling, reward/repress feedback loops, production system-based computation giving rise to "fast" (System I) thinking, "hallucinating" to interpolate between sensory inputs ... the list goes on and on

I already posited that LLMs may perform lower-level thinking. Also, in the reference you linked to read their third stated limitation. Perhaps one day some kind of AGI will do "higher-level" aka more realistic thinking... but not yet. Which, I want to add, is not entirely relevant to their utility to us as humans.

Talking about "thinking" gets us distracted from what can actually be done, though if we do want to do it then we'd need to come up with a working definition of what "thinking" actually is.

To reiterate the point, we also share that cognitive commonality with most birds and mammals too, so it doesn't really mean much.

Model output can look like human reasoning however they often ignore these intermediate steps and the output tokens are often filler designed to allow more context to load. There something there that's half way to reasoning because it's loading that related training data but it not a connected chain of thought as we do. I read this article about it a few weeks ago and it summarizes the current research.

https://www.quantamagazine.org/is-ai-reasoning-right-for-the-wrong-reasons-20260731/