A place for developers to talk about building on Farcaster.

/dev

I don't think I'm leaning into the probabilistic nature of LLMs enough...

Codex can now generate multiple responses simultaneously for a single task, helping you quickly explore possible solutions and pick the best approach. 

Best-of-N is rolling out to Pro, Enterprise, Team, Edu, and Plus users starting today.

Here’s Joey, an engineer on Codex https://t.co/D1pOJ1AmpH

OpenAI Developers

I don't think I'm leaning into the probabilistic nature of LLMs enough...

https://x.com/OpenAIDevs/status/1933617093938692164

I see stuff like this, which is an agent that uses "darwinian exploration" / open-ended algorithms to generate variations of itself. And then it runs the coding benchmarks against them to determine performance

I see stuff like this, which is an agent that uses "darwinian exploration" / open-ended algorithms to generate variations of itself. And then it runs the coding benchmarks against them to determine performance

https://sakana.ai/dgm/

And this, where a guy translated a compression library from C to Rust by having agents generate un-guided fuzz tests over all possible inputs until the Rust output was the exact same as the C output. 

I thought his bottom-up approach was neat, porting all the symbols at the bottom of the call graph first, and gradually working up through the codebase's levels of abstraction. Normally your instinct might be to start with the high level function, and work your way down through the children until it works

Or this guy, a vulnerability researcher who set up his own harness to run a prompt 100 times. He found that o3 discovered a legitimate vulnerability in 8 of the 100 executions, which like, setting up this infra was probably less work than raw dogging the investigation? Right?

(although his experiment also found 28 false positives 😵‍💫)

Thematically these example feel like this https://minimaxir.com/2025/01/write-better-code/

You're asking the computer to "Enhance" over-and-over, through hundreds of variations, programmatically working through the entire input space

Or this guy, a vulnerability researcher who set up his own harness to run a prompt 100 times. He found that o3 discovered a legitimate vulnerability in 8 of the 100 executions, which like, setting up this infra was probably less work than raw dogging the investigation? Right?

(although his experiment also found 28 false positives 😵‍💫)

https://sean.heelan.io/2025/05/22/how-i-used-o3-to-find-cve-2025-37899-a-remote-zeroday-vulnerability-in-the-linux-kernels-smb-implementation/

And this, where a guy translated a compression library from C to Rust by having agents generate un-guided fuzz tests over all possible inputs until the Rust output was the exact same as the C output. 

I thought his bottom-up approach was neat, porting all the symbols at the bottom of the call graph first, and gradually working up through the codebase's levels of abstraction. Normally your instinct might be to start with the high level function, and work your way down through the children until it works

https://rjp.io/blog/2025-06-17-unreasonable-effectiveness-of-fuzzing