[poutine]

Have you met my cousin Barry? He does pretty much the same thing, though in a less formal way.

Barry is still sentient and can perform quite a few tasks quite admirably. I still wouldn't use him as a sole source reference for obscure facts however.

(I don't actually have a cousin Barry, this is for illustration).

[crazygringo]

No, actual human beings don't generally lie to your face like this about whole sets of facts. That would be pathological behavior.

They usually just say they don't know, or they think it might be X but they're not totally sure.

People sometimes lie when it's in their self interest for various reasons, e.g. where they were last night, or when writing an Op-Ed or on the campaign trail, but not just lying willy-nilly about regular facts when asked a normal question.

[kenjackson]

They may not do it to your face (though some will). But on the internet there are a bunch of people that will lie to you. Even worse they often do it for financial or social gain.

ChatGPT's motivation is simply that it was trained to do so. Huamns usually have more nefarious motives for their lies.

[poutine]

You haven't met a bullshitter then. ChatGPT can be a bullshitter. We built a bullshitter.

[dmonitor]

Notice that we don’t have Barry fielding a public Q&A so that he can share his knowledge with the world.

[poutine]

Lots of Barry's out there, one of them even got elected president and shared his knowledge with the world quite widely. Half the people took him credulously.

Point being, yes the LLM loves to make shit up. Lots of people dismiss it as a result. It's still bloody impressive, we just need to be aware of its limitations.

[f6v]

> Half the people took him credulously.

I get that the current US president is senile. But that sets a low bar. Why do we need to pretend something is good if it’s as shitty at facts as some people? People want something that’s better and more trustworthy.

[kilgnad]

A low bar? You're saying if someone recreates an AI that imitates biden or trump perfectly that we failed to do anything impressive?

Talk about self denial. chatGPT literally passed the turing test and people are still literally just thinking it's just a probability word generator.

It's more then just a word generator.

[f6v]

I went to Twitter and read what ML scientists say. They don’t think it’s anything like a layperson imagines it to be(I’m also a layman when it comes to LLMs). But it’s an impressive technology IMO. I just think we don’t know all the limitations and strength yet because there’s a vocal majority that suffers from survivorship bias.

[kilgnad]

They're just telling you how LLMs work. Anyone can understand the underlying algorithm with a bit of study. It's trivial.

Nobody is understanding the high level emergent effects of LLMs plus training. What the ML scientists say has as much credibility as a lay person in this regard.

[poutine]

You think ChatGPT isn't good because it bullshits some times?

[f6v]

I think it’s a wrong question to ask. It’s an impressive piece of technology for which people are trying to find use cases. But the right questions are: 1. What are the tasks where LLMs out-perform humans, or, at least deliver similar performance? 2. Are LLMs more efficient in these tasks?

[mrtranscendence]

I'd rather a senile president than sub-100 IQ megalomaniacal sociopath, but YMMV.