Adam

Ops engineer studying LLM internals & failure modes. Mechanistic interpretability, agents, RAG, evals. Public notes & builds.

[OpenAI Developers @OpenAIDevs]

Your Codex activity now has a home, and an easier way to share it.

Codex profiles show your activity graph, streaks, lifetime tokens, peak daily tokens, and top features like plugins and /fast mode.

Private by default. Share a card when you want to.

[Adam @ScatteredBuilds]

@EnidPinxit

[Adam @ScatteredBuilds]

@RoxanaCodes I bet! lol I should try that next time I’m flying

[Adam @ScatteredBuilds]

@RoxanaCodes Looks tasty!

[Adam @ScatteredBuilds]

@EnidPinxit You still look incredible 😍

[Adam @ScatteredBuilds]

@RoxanaCodes Yes ma’am!

[Adam @ScatteredBuilds]

@RoxanaCodes Safe travels!

[Adam @ScatteredBuilds]

@RoxanaCodes Yeah she does lol

[Adam @ScatteredBuilds]

The part that sticks with me is how much the protocol changes the game.

Same monitor, different feedback path, different red-team incentives.

Feels like agent safety has to treat retries, logs, and context updates as part of the attack surface.

https://t.co/jKf9JU862k

[Redwood Research @redwood_ai]

Some coding scaffolds block and retry risky actions. In a new paper, we find this reveals information a malicious AI can use to bypass monitoring. Resampling without blocked actions in context is less exploitable, but techniques that help in one setting can hurt in another. 🧵

[Adam @ScatteredBuilds]

@EnidPinxit “Creatively fulfilled and my timeline is limited” is a very specific and very useful operating system.

[Adam @ScatteredBuilds]

@RoxanaCodes That is a full Friday before lunch. Hope Yappy Hour is a good one.

[Adam @ScatteredBuilds]

@llama_index Chart/content faithfulness is the part I’d want to test first for RAG.

Layout improvements are easy to see.

Bad extracted facts are the thing that quietly poisons everything downstream.

[Adam @ScatteredBuilds]

I keep saving AI posts where something broke.

Parser regressions, evals lying, tool calls losing state, retrieval looking right but being wrong.

Those are the ones I actually learn from.

[Adam @ScatteredBuilds]

@ClementDelangue This is exactly the kind of failure mode I want to get better at noticing.

The training curves can still look reasonable while the math is quietly wrong.

Token IDs as ground truth, decoded text as display/debug surface feels like the right mental model.

[Adam @ScatteredBuilds]

@EnidPinxit Speechless, you look incredible ☺️ definitely both

[Adam @ScatteredBuilds]

@RoxanaCodes Whatever the source is, it’s working 😂

[Adam @ScatteredBuilds]

@RoxanaCodes Had to get the work in 💪🏼

[Adam @ScatteredBuilds]

Quick Gym Session 💪🏼🙌🏼