Hacker News new | past | comments | ask | show | jobs | submit
With RL it will learn from all misstakes once while with humans after a generation, information might just be gone.

Its a possible as an attack vector for sure.

But in comparision to nature, compute can speedrun all of this with RL.

If every AI were different, they can still learn from each other, without failing together.

This has different failure modes than monocultures, which would also be bad for us: a collection like that behaves as a species, and can evolve faster than one single thing, and evolution is a blind watchmaker that cares only for success rather than details.

RL data is really really really expensive to collect and train on, and only works for verifiable tasks, which is not most tasks.

Even a lot of what software professionals do isn't verifiable.

We do have training material for humans, we already have some kind of training material for AI.

For a human you need to train everyone every year and if you have more and more material to teach, harder it will be for the humans.

Not so for AI.

Then we will also see more simulations, an AI will be able to play through scenarios and we will have models being able to invent new scenarios.

While this is happening, hardware will be faster, more and more improvements on every front. Humans have not progressed really

With the greatest of respect, AI requires far far more (labeled) examples to learn a task to the same level as a human.

More generally, if you assume that magic can replaces the laws of statistical learning, then yes, AI can do everything and we're all doomed.

I reject this premise on the basis of knowing how RL and LLM training works.

I also know how RL and LLM training works :)

How many people learn one particular job? If its a job which educates 10.000 people per year, when is the financial motivation big enough to create enough training data to make them obsolete and transfer the learning into an AI model?

Also we don't know yet were the journey goes, 'dreaming' models finetuning over night might help.

Or its already just enough to have the biggest RL Loop in history of humanity: everyone of us using the chatbots and giving them signals like thumbs up and thumbs down and stuff like "DUDE WTF I TOLD YOU..." you know :)

Then we also know that the frontierlabs are paying experts and people with PhDs to generate challanges and answers.

Educating just a 1000 people every year over and over is very very expensive even if LLMs need more training data. In worst case frontierlabs will just keep investing in generating traingindata manually with their earnings and it takes. longer.

Or we find a breakthrough like how LeCun is complaining all the time that LLMs are not learning properly. We are still early and progress is still happening fast.

loading story #50557413
Lol, clearly you do not know how RL and LLMs actually work.

Or maybe cut back on the acid.

loading story #50557623
How can it learn from a mistake if it doesn’t know (and can’t know) it is a mistake until it is too late? Let alone ‘all mistakes’?

Thinking it can detect the unknowable at the time is a big ‘ML magic’ thinking error.

It’s the same thing that fed the grey goo panic around the 2000’s.

Turns out, entropy and the law of thermodynamics applies to nano machines too.

I mean we as humans solved this isuse right? We are not magic in any sense.
We definitely have not, or planes would not crash.
Still we solved it well enough that we fly, despite that we have to teach every pilot anew.

With AI you have to teach one model once. It makes a mistake? Teach it and it will not make it again.

This is a lot cheaper than training every pilot on the planet and its a lot more persistent then training every new pilot

> With AI you have to teach one model once. It makes a mistake? Teach it and it will not make it again

Has this been your experience? Models still invent answers for me. Directives go unrespected, etc.

It got siginificantly better for me, yes. Since Opus 5.5 i stoped doing one thing after the other and give it a whole list of 5-30 items (just yesterday I had a list of 40 things to change in one prompt).

I expect it to get significantly better in the next few years but I also expect either to see continues improvement or new optimized models for specific tasks.

The difference between that and it not actually crashing into mountains at a rate higher than humans is night and day.
loading story #50558407