Hacker News Reader

Comment by 2ndorderthought | Hacker News Reader

[flagged]

Not to mention the comment agents. The true AI bro/sis won't waste her/his time commenting when an agent can do it for him/her.

Have to efficiently trash the internet of course.

AI bros and crypto bros. One and the same thing. Same optimism. Same arguments. Same blind faith. Same zero knowledge of how economy, society or even the technology they are evangelizing works under the hood and what are its shortcomings that are impossible to overcome because physics won't allow.

> what are its shortcomings that are impossible to overcome because physics won't allow.

Have you intended to say "because reasons"? There should be a long chain of reasoning connecting "LLMs will never be able to strictly follow instructions written in natural language (as agreed by 90% consensus of experts or some such, because you can't formally verify adherence to informal natural language instructions)" and "physics doesn't allow that." And I can't find it anywhere. Neither in your comment history, nor in literature.

Because it is next token predictions and there is no logical part into that. It is like saying I cannot find any research paper on the fact that car collusion might cause fire because gasoline is in the tank.

But the fact is that there's plenty of literature out there on hallucination and unreliability of LLMs already. If you know otherwise, let us inform Dario before next funding round.

Next token prediction is a pretraining objective that doesn't tell anything about behavior and activation structure of the resultant network. The literature that explores a hallucination problem has little to do with your claim about physical impossibility.

If you want it to be sarcasm, treat it like sarcasm. What's the worst thing that could happen?

Well, the last time I assumed everyone was being sarcastic and ironic, until they actually elected the guy president.

Haha, ah yes, good old GWB.

Reply on news.ycombinator.com