Comment by pixl97 | Hacker News Reader

Comment by pixl97 | Hacker News Reader

Parent Full thread

>AI didnt break through stavier-nokes until a human set the initial direction. That's not going to change.

Why isn't this going to change?

Right now it's easy to see why don't set AI lose on every problem, someone human or AI has to delegate rare resources between competing interests. But if we look at general compute it was no different in the past. In 1980 you had to ask for permission to get CPU compute time. In 2026 you run it on your own computer, or maybe pay for it on Amazon. The constraints are much different.

If we keep pumping out chips and increasing efficieny someone will just make the LLM into an agentic loop (build the harness in) and set it lose on problems.

It’s like asking why we don’t just run all the programs instead of making users decide what to run. There’s an infinite number of problems that could be solved, and the question of which ones are important to solve is a question about us and what we want.

Model collapse

If you solve it then yes humans are toast mental usefulness wise

Model collapse as an insurmountable wall is a fiction dreamed up by Luddites, actual frontier labs understand the failure mode and avoid it.

Nobody can avoid it. You can RL things in a box until the box is mapped perfectly but that doesn't mean you can go forever on out of distribution data. Our number of boxes will increase but model collapse is an open problem.

Astra goes farther than any model before but it cannot go forever.

Models have been trained on synthetic inputs for over a year now. Collapse only happens in very specific circumstances, it's not a limit in practice.

Yeah, that's why fully autonomous robots are so easy to train on novel tasks. Or why figure is strapping cameras to human workers for cleaning.

Sythetic data works in a bounded box we have mapped you can't just make up novel data and it works.

Model collapse is completely unsolved.

Reply on news.ycombinator.com