> That’s not true at all. The biggest issue is that it doesn’t work. You can’t actually trust ai systems and that’s not a product issue.

I don't know about that, it seems to work just fine at creating spam and clone websites.

It works fine for some things. You just need a clearly defined task where LLM + human reviewer is on average faster (ie cheaper) than a human doing the same task themselves without that assistance.

Given the fact that you need to review, research, and usually correct every detail of AI output, how can that be faster than just doing it right yourself in the first place? Do you have some examples of such tasks?

Yes. There's the one that $employer built a POC app for and found did in fact save time. There's also github copilot which apparently a large chunk of people find saves time for them (and which $employer is reviewing to figure out if they can identify which people / job functions benefit precisely enough to pay for group licensing).

Ilya is one of the Founders of the original nonprofit. This is also an issue. It does look like he was not the Founder or in any control of the for profit venture.

if the ai is the product, and the product isnt trustable, isnt that a product issue??

It’s a core technology issue.

The AI isn’t the product, e.g. the ChatGPT interface is the main product that is layered above the core AI tech.

The issue is trustworthiness isn’t solvable by applying standard product management techniques on a predictable schedule. It requires scientific research.

> You can’t actually trust ai systems

For a lot of (very profitable) use cases, hallucinations and 80/20 are actually more than good enough. Especially when they are replacing solutions that are even worse.

What use cases? This kind of thing is stated all the time, never any examples.

All the usecases we see. Take a look at perplexity optimising short internet research. If I get this mostly right its fine enough, saved my 30 minutes of mindless clicking and reading - even if some errors are there.

You make it sound like LLMs just make a few small mistakes when in reality they can hallucinate on a large scale.

Any use case where you treat the output like the work of a junior person and check it. Coding, law, writing. Pretty much anywhere that you can replace a junior employee with an LLM.

Google or Meta (don't remember which) just put out a report about how many human-hours they saved last year using transformers for coding.

What are examples of these (very profitable) use cases?

Producing spam has some margin on it, but is it really very profitable? And else?