[gerdesj]

Well, you say we underestimate but are you sure you are not the "underestimater"?

Amongst other things, one of the tasks I tried to put ChatGPT to was writing scripts in a not so popular dialect of a very popular language: PowerCLI (PowerShell). Gemini is even worse!

The issue is of course the relative lack of PowerCLI vs the huge body of PowerShell generic stuff. Hallucinations include invented function parameters and much worse. It doesn't help that PowerCLI and MS's Hyper-V effort (whatever that is) both have a Get-VM function etc.

These things are "only" next token/word guessers. They are not magic and they are certainly not intelligent. I do get great results in other domains and with a bit of creativity but you have to be really careful.

No need to feel triggered. Use these tools as best works for you and crack on but do be careful to be an engineer and critically examine the output from the tool.

[cjbprime]

> These things are "only" next token/word guessers.

This is precisely the kind of vacuous "this is technology, I know technology, this is simple" hubristic underestimation that's being called out.

There is no upper bound to the intelligence of a "next token/word guesser". You can end up incorporating an entire world model to your predictions to improve their accuracy, and arguably this has already happened, to a currently-unreliable and basic level. It is possible that no technological advances are required to reach better than human intelligence from this point -- only more compute, bigger models and datasets, and (therefore) better next-token predictions.

[gerdesj]

"This is precisely the kind of vacuous ..."

So I am devoid of anything? Nice. I slapped "only" within quotes to imply that there is more going on and a lot more complexity than implied by a naked reading of my comment. I'm sorry you missed that.

There is no notion of a bound or even intelligence for a LLM. It is a tool and no more - we know how they work - that is defined and we run our own. We can marvel at what looks like intelligence from the outputs but it isn't that. They can be enbiggened ad-nauseam but I very much doubt we'll get intelligence per se.

You might disagree with my arguments but please don't describe me as vacuous.

[cjbprime]

> So I am devoid of anything?

The completion was "vacuous [..] underestimation". Something you appeared to be doing in this one phrase you wrote, not something you are. I don't know anything about you as an entire person. Please try not to take criticism of some of your written thoughts so personally. I continue to take issue with characterizing the LLM as "a word guesser" because it implies a limit to capability that I don't think actually exists.

> There is no notion of a bound or even intelligence for a LLM.

When LLMs are outscoring humans on many/most standardized tests, including for tests where the questions are novel, I also disagree that there is "no notion of intelligence for an LLM". It feels like goalpost-moving, to the extent that I now have no idea what you actually mean when you say intelligence.

> we know how they work

I think this is also a hubristic statement. The researchers working on these systems do not speak like this. They say things like "we did reinforcement learning on question-answering in English and it turns out it answers questions in French too now and we were surprised and can't explain why that happens".