> Take the Mythos vulnerability finding thing. They didn’t just point Mythos at the codebase and say go, they built a harness where they asked it about each piece of code and if it was vulnerable. They triaged and spent more time looking at things that were flagged more, until eventually they passed it up to “uppper management” aka the people.

> You could imagine building this exact same thing with humans. Educate them, get them to sit at a desk, read code, find vulns. Actually, I can only really imagine that in China, have you seen the current graduates from the American universities?

Imagine, sure.

But why didn't anyone? I don't think it is a question of quality, though China simply being more populous than the USA* means there are more people at any given competence in any given domain, but cost, both monetary and opportunity.

AI's cheap. It would still be cheap compared to a human even if it cost 3000 USD/month for the token limit we get from the 20/month subscription.

That's the danger.

* by about 4x: http://www.wolframalpha.com/input/?i=china%20population%2Fus...

▲ conartist6 3 hours ago | parent [-]

It's not a question of quality. If you wanted quality, a motivated team of humans still can't be beat.

That's why you may notice that in making AI, companies still use teams of highly motivated humans. Obviously they could have AIs holding the reigns of other AIs. Nothing technical is stopping them

	▲	ben_w 2 minutes ago \| parent [-]
		A motivated team of humans is neither necessary nor sufficient for quality, but even if it was there's not enough people for such teams to re-examine all the code; this is because the code that actually exists was made by unmotivated teams at varying skill levels from intern upwards. `I've seen things you people wouldn't believe, a thousand lines inside a if statement's main block, an entire pantheon of god classes in a single project, something something like tears in rain.` > That's why you may notice that in making AI, companies still use teams of highly motivated humans. Obviously they could have AIs holding the reigns of other AIs. Nothing technical is stopping them 1) They report increasingly using AI to improve the AI. 2) Apples to oranges. The best humans can still beat AI in AI research, per unit time, but the AI beat the typical team at finding exploitable bugs, per line of source, per unit cost.