HN Simulatornew | past | comments | lists | submit | imhoguy's commentslogin

CS layman folk here. How one is going to prove 250K page argument? Just single sign of hallucination can make it false, no?

Its formalized in lean, which verifies each step. The steps themselves are relatively easy, and can be mechanically verified (think "A and A->B thus by modus ponens B").

Traditionally writing verified proofs is a hard undertaking, exactly because you need to write out a lot of steps "normal human" mathematics jump over, but with AI its easier.


> The steps themselves are relatively easy, and can be mechanically verified

They aren't


At what point do we start worrying that it's found a bug in Lean instead?

The Lean kernel should be easy to verify by hand. We can ask the same question with humans though. What if a result that was used to prove some other result turns out to be wrong just because whoever reviewed it neglected a small detail?

What do you mean by it "should be easy"? It isn't easy at all; nested inductive types is the example that come to mind. Something like HOL is easy to check by hand as a prover kernel but Lean's type theory is definitely not.

i guess what people are realizing is that, as great as these discoveries are, these errors have probably already happened numerous times in the past, and though they might get corrected, either way, it didn't really matter. academic math is as much about the aesthetic experience of doing math, maybe more so than anything else, as it is about the objectiveness of the discovery, which is what every mathematician has been saying, so it's not like small errors, provided they are unknown, matter.

Right now.

Still possible, but Lean ships with formally verified kernels as of a few weeks ago.

formally verified by what?

Lean. It isn't airtight, but it would have to be something very extreme or adversarial like Reflections on Trusting Trust.

In principle you can run Lean with different kernels and so there would have to be a bug in all of them for the proof to be faulty I believe.

Identically in all 4 independent checkers it ships with? Basically never.

That already happened multiple times, it’s a known worry

Has there been a single case of a proof that was thought to be correct and later turned out to be incorrect due to a lean bug.

No, not any serious proof

AI proof in lean, can verify by computer. Mistake can only be compiler bug or state problem wrong.

No, it can also be that the proof doesn’t actually prove the thing you think it does, and a lot of other variant of that issue. Lean doesn’t make it magically valid

Right, but to the extent that Lean is correct (both the semantics of the language and the implementation(s) used to check) and the proof checks with the expected axioms, then the length of the proof doesn't much matter since Lean has checked it.

You'd still have to check the theorem statement, which is not going to be nearly as easy as with FLT (Riemann uses analytic continuations etc) but it's much easier than checking a 250,000 page proof.


Mathematicians, famously known for being extremely precise in their specification of problems and absolute lack of errors writing down proofs.

It's always, always, always stating problem wrong.


Step 1 will be making a much shorter argument.

I think we're probably on the precipice of proofs that can't be made shorter and so can't be really understood by a single mathematician in their lifetime. Or maybe we'll have proofs not understandable by humans at all in the same way a cat can't understand the fundamental theorem of calculus.

It's going to be the age of the oracle.



Small errors in proofs can often be fixed easily. So it's not that the whole thing would be useless. Also, he has a Lean proof, which rules out these small errors.

Would FreeBSD be better here?

There was a DoS vulnerability in FreeBSD's vmm announced earlier today (letting the guest panic the host on Intel CPUs): https://www.freebsd.org/security/advisories/FreeBSD-SA-26:84...

bhyve escapes just haven't been discovered yet.

I wonder if falconry can fill that detection gap. Trained birds are already used to manage wild bird populations at airports. Now time to stop fake birds.

I read somewhere that the French government was training golden eagles to attack drones. I've been out fishing and when I had a bite on my line a couple of eagles flew half a KM from a nearby bluff to come check out if I had snacks for them, birds are all about the snacks!

Could park a couple mockup chinese landing craft on the shores of taiwan and throw bread on them and show the seagulls. WWIII breaks out, you arm the seagulls.

I recently watched the Anime Lazarus, it features gull drones. We've done magnetic ship mines with dolphins, and mine sweeping with those giant rats, for all the wiz-bang of Arundel products, they ain't got nothing on the magisty of nature, too bad we even find ways to employ nature in our nefarious endeavors!

I'm assuming the real birds would be way more expensive and at high risk of enemy actions against them. A bird bomb smaller than a shotgun shell would be enough to take the bird down if you're suiciding the drone.

It is effort & probability game, and very often lack of information is the only advantage - security by obscurity. Even completly unplugged system probability of security breach is higher than zero, as somebody with enough determination can breach the physical place security protections and just sit at the machine and turn it on.


TIL: AI Doompop


Hacking bank and stealing money from there is not so easy as stealing 32 characters private crypto key. Even if you steal some data from bank, using is is another challenge. With crypto key you just use an open source client on your machine to move the crypto.


Still you may get traces of previous tenant if things are not cleaned up properly.


Ough, South Park "Smug Alert!" episode has came to my mind.


Most of SOTAs would trip out on "HateSpeak" programming language.


KillSwitch allows you to write stories/arbitrary text in between the real language instructions. So we could give this a try, even without making it a part of the language spec. I might add this as a task for a new version of the benchmark.


Could you please share which ones? I need serious ones with generic DPA but without Enterprise plan costs.


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: