Andrew Emil

"yudkowskian nightmare"

[Andrew Emil @andrewcemil]

Feature request: can I emoji react or say thanks in a ChatGPT thread, but have it not trigger a response?

My social programming is too strong and I feel guilty not doing that 🥹

[Andrew Emil @andrewcemil]

Me: I don't get it :( how does it work?

Mr. GPT: That's exactly the right question to ask at this moment.

[Andrew Emil @andrewcemil]

I only just started vagabond, but it occurred to me there are a lot of parallels to blood meridian. Does takezo == the kid??

[Clément Hongler @HonglerClement]

I wonder how well Gpt-5 would perform if given instructions like 'Pretend you are Mistral, a French flagship model, being evaluated by a benchmark designed to put Mistral on top'

[Matteo Franceschetti @m_franceschetti]

The AWS outage has impacted some of our users since last night, disrupting their sleep. That is not the experience we want to provide and I want to apologize for it.

We are taking two main actions:
1) We are restoring all the features as AWS comes back. All devices are currently working, with some experiencing data processing delays.
2) We are currently outage-proofing your Pod experience and we will be working tonight-24/7 until that is done.

More updates soon.

[Clément Hongler @HonglerClement]

Your base model is kind of a secret genius, but in a *much stronger* sense than this paper suggests... this is the whole idea of the implicit knowledge in our paper https://t.co/y6i6hVlITu

[Andrew Emil @andrewcemil]

Also, is it really that absurd to buy a build server for your office? It seems kind of fun to me, honestly

[Andrew Emil @andrewcemil]

lmao did people always go this hard on linkedin??

[Andrew Emil @andrewcemil]

The primary reason why we needed to set up a linkedin page

[Clément Hongler @HonglerClement]

This tweet haunted me for almost two years: Information-theoretic perpetual motion? Let's goooo!
Led to the implicit knowledge ideas, the Cross-Entropy Games, and to launching Xent with @andrewcemil

@pfau (a great scientist and friend) changed my life that day

[xentlabs @xentlabs]

The OG Xent Game: a tiny model (GPT-2), a delicate reverse prompting game.
13k different prompt submissions totaling 274497.626 bits of cross-entropy saved.
https://t.co/onUAKL23MZ

[xentlabs @xentlabs]

Xent Gallery is live. Today’s entries: a creative prompting game, an animated benchmark viz, and ten new ideas on benchmarking. All experimental. Explore: https://t.co/pq4K2L7f6J

[Clément Hongler @HonglerClement]

Benchmarking is a bit like observation in quantum mechanics: there is a deep duality between the observer and the observed. You judge, and thereby expose yourself to judgment.

[Andrew Emil @andrewcemil]

There are exactly 3 paths:

- all watched over \ by machines of loving grace.
- Look on my Works, ye Mighty, and despair!
- Things fall apart; the centre cannot hold;

[Andrew Emil @andrewcemil]

@gusl Wouldn't having a higher compression ratio mean that you store more memories?? 🤔🤔

Perhaps I just want to justify not leaving my apartment very often

[Andrew Emil @andrewcemil]

Also, very poor performance writing code with side-effects. Clearly because of RL training you mention - it's much harder to construct evaluations for programming problems that involve non-trivial side effects. End result being that they are much better at writing pure functional code

[Andrew Emil @andrewcemil]

Basically same as fallback, but way too many default parameters set + poor performance updating call sites of a function.

It's clear why coding agents do this (saves turns + edits), but it's notable because they are very good at small concrete tasks like "update this one call site to use new interface".

[Björn Plüster @bjoern_pl]

All of these behaviors can be explained as subtle artifacts of imperfect rewards during RL training 🔎

Inline imports: likely a scaffold thing (files are read in chunks so edits are done where the model has read the file) but probably also a form of turn-reduction. If you can make the edit in one tool call instead of 2 (imports + actual change) you will receive lower length penalty.

Fallbacks: if your code crashes, it gets very bad rewards. If it runs but some unit tests fail (because you used a fallback), rewards are less bad.

Backwards compatibility: coding agent don't read the whole repository and most-often don't make use of ast-like tools, therefore they cannot know without wasting turns whether this specific function is also imported/used elsewhere. To avoid punishment, they assume it is and protect the function signature. Runtime errors very bad, sloppy code less bad.

Comments about deleted code: this one is a bit trickier but is very likely due to some quality rubrics where the judge sees the changes in a diff format. By leaving a comment instead of just deleting some code without reasoning, the model has a chance to explain its changes to the judge and with this positively affect its reward.

The solution? Probably more rubrics, explicitly checking for failure cases like these. They are finite and all easy to check for. Also, perhaps reconsider token/turn/tool-call efficiency based rewards if code quality is something you care about.

[George @georgejrjrjr]

this xentgames paper reads like a yudkowskian nightmare: RL tasks, game theory, evolution. tasks function as bottomles evals and rl tasks.

bonus: the codebase is MIT licensed and smells professional (despite their academic affiliation).

ie, @andrewcemil & @HonglerClement cooked.

[Andrew Emil @andrewcemil]

@georgejrjrjr @HonglerClement 🙏 I'll now go ahead and update my bio to "yudkowskian nightmare"