It’s a perfect example of how “using LLMs for test coverage” can also be harmful. He expected the tests to to prevent introduction of said regressions, probably based on a combination of the quantity of tests and their style (they look like what decent human written tests look like). But the tests are AI slop, and so they give a lot less value per line of code than he expects, hence a significant regression.

It is literally useful to call these tests AI slop, and the problem is in part caused by not calling them AI slop, and having consequent inflated expectations. LLMs are not any better at writing tests than at writing other code! It is merely that the bar for tests can, legitimately, be a lot lower (in projects where there would otherwise be no tests at all). Making an exception to calling AI generated tests “slop” is thus counter productive, because it leads people to act as if LLMs are actually better at writing tests than at writing other code, and not just because the bar for tests is frequently very low.

edit: actually scratch that I looked at the PR and those tests even look like dogshit and worse than the tests I seen claude write at a workplace that was into vibecoding (which i since quit).

Oh, by far. There’s only 80 decimal places in that at most.

It got to be a quantum sweatshop: a quantum computer for AGI (a guy instead)

Oh they are going to charge per token for github copilot? That thing is a money waste for everyone, I’m pretty sure. I get a mix of inane mildly good suggestions, irrelevant stuff, and an occasional suggestion of super evil sabotage. Due to mild OCD about issues, I tend to have to fix said mildly good suggestions, but from the objective perspective that nitpickery is not worth it, everything was fine without, we had compiler warnings, coverity, etc.

edit: the difference being that the old stuff was deterministic and you just ran it on the whole codebase and had it pass. Unlike gh copilot that’ll just make up new shit. And as for the times it caught some bad bug that you made… add more tests instead.

I wouldn’t be too surprised if they really don’t, they’re just advertising the advertising lol.

edit: Basically what if you spent a trillion dollars so that you could beam ads to people’s bathroom mirrors. And better yet, ads reflected from water down in their toilets. Then in the interest of expediency you just take random ads and put them there for free, and your actual product, shares, sells better.

It makes every bad programmer into a 10x bad programmer (equivalent to 10 bad programmers).

I’m afraid they already had that exact idea when they named the startup “oklo”.

I think it’s not very difficult to construct a really shitty small reactor that is horrendously expensive per watt. Can probably be built in a year if you get rid of NRC and just half ass it completely.

I mean, Demon Core was a small reactor. You pretty much have to do a lot of work to ensure you won’t create a small reactor when a truckload of fresh fuel falls into a river.

What’s difficult is making a safe reactor that is actually making electricity at somewhat reasonable price per watt.

Nuclear already makes 9% of world’s electricity.

I think if people are citing in another 3 months time, they’ll be making a mistake

In 3 months they’ll think they’re 40% faster while being 38% slower. And sometime in 2026 they will be exactly 100% slower - the moment referred to as “technological singularity”.

Its also interesting that this is the most conservative, pro “its not just memorizing” estimation possible : they multiplied the probabilities of consequent tokens. Basically it means if it starts shitting out a quote it will not be able to stop quoting until their anti copy the whole book finetuning kicks in after 50 words or so.

It can probably output far more under a realistic test (always picking the top token, temperature =0)