Yeah I'm aware of that. I'm not sure what I said that made you think otherwise. I don't think it's very similar to fuzzing but the op does

the faster we find vulnerabilities the less time there is to exploit them? if you don't find them fast enough vulnerabilities will pile up in the codebase faster than they are closed

yeah, i understand that. the claim is that this tool speeds up the process dramatically.

do you not find the linux local privilege escalation vulnerability interesting? i guess the details there are more sparse?

a lot of the tools we use already are not deterministic, right? that's what fuzzers are? randomness is also used to make a more minimal trigger

if the tools were good enough wouldn't these old vulnerabilities have been found by them? a lot of resources have been spent hardening these platforms.

Aren't most security vulnerabilities logic bugs?

What would a bug bounty for this kind of exploit look like? I would've thought 20k would be in range

And yeah I do not doubt the process could be optimized

They found remotely exploitable dos and local privilege escalation bugs. I think everyone would care about such exploits. The maintainers decided them worth fixing. Academic and industry security researchers often point their fuzzers at major code bases in the hopes of finding stuff like that

If that's so why did they find decades old bugs that previous researchers couldn't find. Do you think they had way more human triagers?

At least it's an impressive fuzzer if it found many previously unfound issues?

Clearly an April fools prank

It's just nice to imagine the governments are capable of cooperating and getting things done

If a cat has nine lives does it game over after nine or ten deaths?

Isn't it just crepes

Work with us – Wikimedia Foundation Make the internet a better place for free knowledge.

maybe wikimediafoundation.org/jobs/#sectio... i work there is you have questions

I can see you aren't interested in an actual conversation. That's fine. Have a good weekend

Yes I dropped the word "by" thanks for the correction

i'm aware of and appreciate your work. but the o1 chart you showed does not prove what you seemed to say. there are benchmarks not created or revealed to the major labs that are also going up. i believe you say that the private benchmarks are leaked via the APIs but that doesn't explain open models

Yes. But that's with fixed techniques and architecture. o3 outperformed o1 at the same or lower cost on this and similar tasks

But model performance on AIME and similar tasks has continued to improve and the price per token has not gone exponential. I assume you attribute that to over fitting or lies? But you're ignoring the possibility of fundamental architecture improvements

Work with us – Wikimedia Foundation Make the internet a better place for free knowledge.

wikimediafoundation.org/jobs/#sectio... ? i work there if you have questions

yeah categories are a good resource here. i was gonna read the code but the file was long so i thought i'd ask. so many people keep building things like this e.g. wikitok.vercel.app

recs are challenging when you want to preserve privacy and have a small userbase (and have a small budget).

What does the heart button do? I work at the wikimedia foundation and we're exploring stuff like this.

Your brain is being mean to you

Like you can talk. We've been waiting for Quest 65 for years

What happens if we get between them?

When are we getting a creepy centipede wiki plushie?

Crazy that they know your secret hacker alias

When talking about LLMs where do you draw the line for "large"? 200 million params? I think this question greatly changes the answer

According To Wikipedia YouTube video by Daniel Erenrich

according to @wikipedia.org www.youtube.com/watch?v=cT6y...