Hacker News Reader

Comment by amelius | Hacker News Reader

> This lack of meta-cognition is a big reason why this first generation of LLM-based products – ChatGPT, Stable Diffusion, GitHub Copilot – are expressed as copilots, as opposed to say autonomous assistants that can be relied upon to perform a task correctly. We encountered the same issue in our previous post about doing scientific literature meta-analysis with LLMs.

Is there any indication that a next generation of LLM-based products can solve the meta-cognition problem?

View on HN

Both Sam and Yann have expressed desire to pursue the idea of asymmetric computational load, where "harder" problems emit or recurse through more tokens while "easier" problems (like retrieving tbe current date) use less tokens. This assumes there is a fixed amount of intelligence brought to bear for emitting each token.

And that you can accurately determine difficulty and/or confidence.

Seems like passing the input with a question of "where should this input be handled?" into the LLM would be a good first step.

Why would the LLM be any good at determining where the input should be handled?

There's a long history of "this looks hard, I won't implement/fix it" in GitHub issues the LLM can train on.

And then the LLM can probably know whether an issue can be fixed or not. That’s far from generalizing the problem though.

LLMs have shown some limited ability to generalize.

(Opinion) I think internally, they record how closely 2 words are in meaning based on the training data.

If everyone uses similar language to describe different problems, then the LLM should be able to at least act like it’s generalizing

The first tasks of gpt was to evaluate the emotion of a sentence.

Like determine if a comment/review is good or bad.

GPT4 uses a "mixture of experts" system which is already sort of kind of like doing this.

can probably easily be interpreted from the text?

Did we just discover a new halting problem?

It doesn't seem like this should be the case. Humans can estimate the complexity of a task without performing the task. Why couldn't an advanced LLM?

Isn't a large part of software development's difficulty that people can't actually do that?

>Humans can estimate the complexity of a task without performing the task.

I would argue this ability drops precipitously as the complexity grows. That's part of the reason cost and schedule estimates are notoriously unreliable on big projects.

Accurately or the way humans do it?

Just a meta comment here. When does a metaphor stop being a metaphor and start becoming a lie? Does it have to do with the speakers intention or can it just get away from you and nobody notices?

Anyways. This problem isn't going to be solved.

JEPA architecture seems promising, though my last review was about 4 months back.

this paper explores combining metacognition with LLMs as the application level: https://replicantlife.com/

Not yet.

Not really, and it shows that LLMs just like prior attempts at generalized AI are gonna hit a wall, they're trying to brute force the issue which in my opinion won't work.

Reply on news.ycombinator.com