Some teams use sweeps, heuristics, or scaling laws to determine their training LR. At Character, we just have Noam Shazeer dial it to the right value.

I like AI as much as the next guy but I have to admit there's something that feels a bit demeaning about writing commit messages that say 'ask the model nicely for YAML'

Your one-stop-shop for decompiler knowledge, the Decompilation Wiki (decompilation.wiki), is seeking more public applications of decompilation known by the community.

For example, if you know of any public decompiled games (github), please reply, open a PR, or DM me :).

New janitor is working out pretty well

(I think what's happening is: I have an unclosed brace and hit save before I get around to fixing it. Format on Save then re-indents a bunch of things with the wrong brace matching. Then I fix the missing brace and it doesn't restore the other stuff because ???)

Could whoever designed the mysterious VSCode feature that keeps changing the indentation of random parts of this file that I haven't touched please speak up? I just want to talk

What I like about jq is that it really raises the ceiling on just how horrible I can be in my one-liners

Looking at DSPy but all the examples are very focused on information retrieval/question answering- is it useful for more agent like setups with function calling/tool use? Anyone using it in that setting?

Cool that ChatGPT has an achievements feature now

So you want to build a browser engine? Some advice.
robert.ocallahan.org/2024/06/browse…

Me: I have run this script on TWO different files and it worked perfectly. I have extreme confidence that it is bug-free and that I am the greatest programmer alive
Me (after running it on a third file): Ah. Well, nevertheless,

Google AI continues to impress

Something familiar about one of the examples chosen for the GPT4 “lightly incremented timestamps” feature activations…

FUZZING'24 Submission website is open!
📝 fuzzing24.hotcrp.com (21 June AoE).
🔗 fuzzingworkshop.github.io (16 Sep @ Vienna).
🧑‍🏫 Preregistration-based publication process.
⭐ Accepted papers invited as @ACM_TOSEM Registered Reports

//@YannicNoller (RUB/SUTD) @LSzekeres (Google)

Excited to share bug detection tool we've been working on w/ Marco Elver:
github.com/google/gwpsan

It samples unmodified prod binaries with ~1% overhead, can be turned off completely.

Detects data races, use-after-return, uninits. Can do more e.g. UAF/OOB in syscalls.
Cool tech ...

The first _conference_ on offensive technologies is off to a great start with an amazing program. Looking forward to these presentations!

The kid is in the right here, does no one understand opsec these days?? Might as well announce his first grade teacher’s name while you’re at it

v2 of the SWE-agent preprint is out now!

We have substantial speedups to the agent coming in the next week, and additional exciting updates right after!

Thoughtful questions from Nicholas Carlini.