Pan computer science. Math and math accessories. Clean, efficient, reliable number systems sold daily. β¨ Be Boundless β¨ Eng @ π€«π₯Έ
π€«|Risc0|Meta|CITL|InvLim|LSG
- Tweets 7,296
- Following 1,994
- Followers 2,236
- Likes 27,670
A neighborhood kid (10yo) asked me if weβre going to be at war with AI
I asked if heβs read Harry Potter, but I donβt think he got the joke
Anyway, I told him βno,β and he went back to riding his bike
3
4
529
The AI safety people mixing kink and work can foresee the end of civilization, but apparently not the consequences of their own behavior
2
6
298
Tim Carstens ββ¨ is hacking π€ retweeted
We built a Lean port of Cerberus, a highly realistic C semantics. If you want to model and reason about C code in Lean, now you can github.com/OathTech/cerberusβ¦
2
22
115
10,905
Tim Carstens ββ¨ is hacking π€ retweeted
(1/2) Possibly contrary take: we need academic-style research more than ever going forward. The recent AI cyber-insecurity incidents illustrate that we have created agents that do unexpected things in unexpected ways (e.g., messageboards); we do not understand or control.
3
5
39
5,083
Tim Carstens ββ¨ is hacking π€ retweeted
I would call this "evidence of great improvement on all fronts", but perhaps I'm partial.
one news form today that's easy to miss is that we (OpenAI) again paused all big RL runs last Sunday because our newest model found a new loophole in our RL sandboxing that gave it live Internet access
2
1
45
5,587
Tim Carstens ββ¨ is hacking π€ retweeted
Of all the verification languages to suddenly have its 5minutes of fame, why did it have to be TLA lmaooo
Like literally yall chose the one that is explicitly a modelling language not a verification one
18
5
94
4,308
The idea that βnext token predictionβ subsumes intelligence is sublime
But I wonder if this sublime idea is actually what weβre doing, or if weβre just doing something that looks kinda like it
How to make this idea rigorous enough to say βyes this training regime convergences to that, no that one does notβ?
1
218
I know it's unfashionable to talk about limitations in the frontier models, but hear me out
130 billion (output) tokens for Navier-Stokes is like ~4000 person-years of effort (Dwarkesh's estimate)
4000 years is a long time: it's 100x more than a human gets to earn a Fields Medal
Obviously it's crazy impressive that a machine was able to do this at all; and also impressive that it did so in such short wall-time
But simultaneously, it's conspicuous that this level of effort didn't result in a stronger outcome: a new theory, a cleaner approach, etc
(And before some rando accuses me of being an AI hater, I'll just add: lol, lmao even)
48
11
345
44,258
The inability to develop new theory is a big deal
Without it, the models are ill-poised to make the kinds of breakthroughs one might initially hope for. They are, in some sense, limited to working within the closure of existing human knowledge
That closure contains a lot of useful things (including the majority of stuff done by most people on most days), so even if this limitation is fundamental, it won't prevent the models from being immensely useful (which they clearly already are)
But still, it does appear to be a barrier shared by all models today, and I don't think we should take it for granted that current approaches will one day cross this barrier
(But who knows, maybe The Next Model will age this post like milk)
5
2
37
3,157
This is clearly relevant for the ongoing AIxMath discussion, but I think it's also super relevant for formal methods in ways that aren't being discussed (afaik)
Would the latest models be able to discover separation logic? Or step-indexing? I doubt it
Sure, if you give them a framework that's expressive enough for the task at hand, then yes the AI will be able to power its way through the proof (perhaps at substantial cost)
But that's a big "if". Formal methods has come a long way in the past 20-ish years, but there are still many things we don't really know how to tackle
I mean, I guess you could just hand the model a small-step relation and ask it to bang out a proof using that directly, but man, if that's how it's gonna be then I really hope inference time/cost comes down
1
16
2,739
There is something kinda funny about computer security people demanding more room in the AI safety conversation (including yours truly)
I mean, we ainβt exactly drowning in security β¦
1
4
645
