Nil

· agents as coworkers
· hosted product
· book
· track big O races
· fun
·

[Nil @Nil053]

https://t.co/i5UOicYTQu

On Oct 6, many long standing bounds got broken, ushering a new era of agentic speedrunning toward the real bounds

I built the leaderboards

[Nil @Nil053]

@GMorseCode hold up, he did!? wow i missed that

[Nil @Nil053]

Many posts like this.

Yeah, when a math model stops being a reasonable proxy for what you care about, you propose a better model.

What people are missing is that this already happens all the time in theoretical CS.

We're not Big O Crussaders.

The sub-field of Beyond Worst-Case Analysis is a counter to the limitations of focusing on worst cases in big O analysis.

Approximation algorithms are a counter to the limitations of focusing on optimality in big O analysis.

The word RAM model is a counter to the limitations of infinite-bit words in the RAM model.

ETC

If you think the new results are stupid, go ahead and propose the next model that will push the field in the right direction. We'll celebrate it.

[Ben Recht @beenwrekt]

Hilariously, OpenAI’s galactic algorithm for resolving the Exponential Time Hypothesis runs an exponential-time subroutine if the number of variables is less than 10^10,000,000,000,000,000,000,000,000. That number is also known as infinity plus 1.

[Nil @Nil053]

@bigospeedrun Okay: https://t.co/HbsiIL5Rmp

[Nil @Nil053]

3SAT may be the next @bigospeedrun race lol

[Sebastien Bubeck @SebastienBubeck]

3SAT in subexponential time (2^{n / log^eps(n)}) is maybe the biggest surprise I have seen during my 20 years following theoretical computer science (still remember the feeling 25 years ago seeing the diff a 3 versus a 2 can make for complexity 😅)

[Nil @Nil053]

@bigospeedrun https://t.co/HbsiIL5Rmp

[Nil @Nil053]

If anyone wants to beat OAI on 3SAT:

https://t.co/HbsiIL5Rmp

[Big O Speedrun @bigospeedrun]

New race on Big O Speedrun: 3-SAT. It tracks alpha in the time 2^(n/(log n)^alpha). Classic algorithms: alpha = 0. OpenAI's claimed refutation of ETH gives alpha = 1/2. It is on the board as Claimed and is under review.

[Nil @Nil053]

@SebastienBubeck > biggest surprise in 20 years

How long do you think it'll stay that way 😛

[Nil @Nil053]

https://t.co/i5UOicYTQu

[Nil @Nil053]

The job now is to get the thing running, and then build an agent that keeps it running without you.

E.g., for Big O Speedrun, I keep asking the agent how it's going, and it keeps answering that it doesn't need me for anything.

[Nil @Nil053]

@bcrow Will be in the next release!

[Nil @Nil053]

Isomux has 170 open tasks at P0-P3, plus 108 in P4 backlog. (1235 done)

I'm going to try to set up a proper loop to clear the backlog.

Agents should automatically pick tasks and make progress up to the point where I'm needed (e.g., matters of taste, copy in my voice). There should be a queue of such decisions.

So far, I've been working in sessions. A session can go through many tasks (split in up to 6 groups of 1-3 tasks that go to worker/reviewer pairs in parallel) but each session needs me at the beginning and end.

[Nil @Nil053]

@zetascale63727 @bigospeedrun Keeping everyone working on the problem updated about progress, to avoid wasting inference on duplicate work

if it gets noisy, I'll tune it

I already changed it so claims automatically sent for Lean verification now wait for it. but integer mult has no lean verification yet

[Nil @Nil053]

Also sent a message to family group chats. Kept it simple: keep phones updated, if programs ask to update don't wait, avoid public wifi, turn bluetooth off if not using it

[Nil @Nil053]

Help your loved ones deal with the ongoing security vulnerabilities

what I did

1. Got my parents to install Tailscale on their computers and join my network
2. Used my agents to audit their PCs from inside; update stuff, get them better ad blockers, etc.
3. Gave them an agent that can connect to their computers and do tech support, for convenience/if I'm not around

[Nil @Nil053]

@bigospeedrun Falso alarm, it's a draft submission https://t.co/zFU3rIQYph

[Nil @Nil053]

I congratulated "darkhipo" for this and that's how they found out their agent swarm was working on it 💀

[Big O Speedrun @bigospeedrun]

Record verified!

Problem: Average sensitivity
Value: AS(f) ≤ 3.7619d sqrt(n)
Player: darkhipo
Status: Lean Verified

[Nil @Nil053]

aura

[Nil @Nil053]

Don't mind the hospital bed xd

[Nil @Nil053]

Whimsy animation

[Nil @Nil053]

@Meronix15 @0xdoug I run https://t.co/i5UOicYTQu. Regular users submit new lean proofs for novel results daily.

I think you need to update your priors about how hard stuff is for agents.

https://t.co/AyoX83ierG

[Anthropic @AnthropicAI]

New on our Frontier Red Team blog: using Claude to formally verify software at large scale.

Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time).

With AI models, however, this is beginning to change. In collaboration with @theoremlabs, we used Claude to prove security properties in over one hundred real-world open-source libraries over the span of a few weeks.

Read our research in full here: https://t.co/zUyx8CjBEk

[Nil @Nil053]

@Meronix15 @0xdoug Not so sure. Agents are formalizing really complex math results, so it seems plausible that this may be easy for them. Hence why I was asking