Dorsa

[Dorsa @dorsar]

We made Claude Code multiplayer!

Use it with other people in the same terminal, and connect your Claudes so they can all talk to each other.

[Dorsa @dorsar]

@jparkjmc lmao

[LaurieWired @lauriewired]

I love that randomness can, in of itself, be a form of computation.

Superoptimization tries to find the best possible machine-code for a given function by searching. A lot.

But what if you just…randomly changed the assembly of a program. Sure, 99+% would be incorrect, but there’s a chance you find something interesting!

This was the idea behind STOKE, some Stanford research that, on occasion, would beat out human-written x86 assembly!

For a little snippet of OpenSSL’s big-number library, STOKE found an 11 instruction solution.

LLVM -O0 was 116 instructions. Even GCC -O3 was 27 instructions.

The tradeoff of course is, insane computational cost. They gave 40 dual core Opteron a budget of 30 minutes of search for each small function. The OpenSSL optimization took about 17 minutes, resulting in a ~1.6x speedup over GCC.

Not a bad optimization, but compilation was about uhh…10,000x slower. I do think it’s (at least theoretically) interesting for hot-path optimization, where you might be willing to spend the compilation budget to save a few nanoseconds.

I like to imagine a very unusual (but possible!) processor. Custom hardware that does a randomized search for machine-code optimizations with extremely quick formal verification.

A chip solely designed to make the software on…other chips faster!

[jpark @jparkjmc]

Introducing DeepRSI, a new industry standard for measuring recursive self-improvement.

On preexisting benchmarks, tasks are contributor-based and short-horizon (~12 hrs). In DeepRSI, we push agents to work 48 hours on ML tasks actually derived from ongoing proprietary research.

[Dorsa @dorsar]

@hackermondev What are some of your favorite writeups?

[Dorsa @dorsar]

@jaivinwylde Lmaooo

[Dorsa @dorsar]

@devahaz Yes, this relies on our definition of superintelligence. Here I assume superintelligence implies know/inspect

[Dorsa @dorsar]

Will standard game theory break when two superintelligences interact because they can simulate each other's code?

[Dorsa @dorsar]

@MehmetMars7 lovely

[Dorsa @dorsar]

@benrayfield Are you referring to recursion?

[Dorsa @dorsar]

@silver__tsuki @Kaweees Lol

[Dorsa @dorsar]

@harjassgambhir Would you recommend watching Person of Interest?

[Dorsa @dorsar]

@EricBeltt Hm, I'm not sure if human <-> human modelling is analogous to agent <-> agent modelling provided the agents are capable of representing each other from scratch

[Dorsa @dorsar]

@sudarshk_ 👀

[Dorsa @dorsar]

@GabeSpound Thinking about this

Even if p=np, they still hit the Halting Problem once they try to run mutual recursion

[Dorsa @dorsar]

For me, this question illustrates the scale and structural weirdness of what superintelligence may be, more than it does a tractable game-theoretic puzzle. Though the latter certainty informs the former

[Shubham Patil @shubhampatilsd]

today im biking to every lyft bike station in sf

t-minus 8 hours to finish

[Dorsa @dorsar]

@theybanjan haha

[Dorsa @dorsar]

sim2real gap is closing on human emotion

[Dorsa @dorsar]

@DavidSHolz three times a year