V I

AI Safety · LLM post-training · agents
Team Lead DS @ MSU · Physics researcher
Building and breaking things to understand how they work.

[V I @d0bk773ki]

@just_some_dev @glnarayanan @blinddonkeyeyes Canonical is only one that hasn’t patched this vulnerability. Most other distros released fixes within days of the disclosure.
I thought it would have been patched by now

[V I @d0bk773ki]

@glnarayanan @just_some_dev @blinddonkeyeyes There is an issue with Ubuntu

[V I @d0bk773ki]

@just_some_dev @blinddonkeyeyes Maybe the fix is already available in -proposed?

[V I @d0bk773ki]

@just_some_dev @blinddonkeyeyes Did you know you're running a Linux kernel that hasn't been patched for CVE-2026-98432?

[V I @d0bk773ki]

@blueemi99 I feel the same way. I think that in this era, having a good harness is just as important as having a good model. I still think Opus 4.5 was a very capable model, and I haven't really felt that much progress since then.

[V I @d0bk773ki]

Does anyone else have issues with the conversation resume in the Codex app?
Since today, I’ve been getting an error with:
[features.token_budget]
use_history_notes_extension = true
It worked fine yesterday, and I haven’t changed my config.toml
@thsottiaux

[V I @d0bk773ki]

@flawedai @NickpxJ I half agree with that. I think you need to give the actual frontier model more space to work.
Superpowers was a very solid choice at the beginning of 2026, but now we need something less strict. Optimizing usage limits starts with that

[V I @d0bk773ki]

6/6
So I want to compare:

fixed classifier
vs LLM judge
vs Jev
vs OpenJev-like scoring
vs a purpose-trained decision model
Not just F1/AUC, but calibration, OOD generalization, latency and cost.
Safety classification seems like a perfect System One testbed.

[V I @d0bk773ki]

1/6
I just got access to Jev by @typesafeai.

And it touches a question we’ve been thinking about in AI Safety:

Why are we using autoregressive language generation for tasks that are fundamentally decisions?

A safety classifier doesn’t need to talk. It needs to decide

[V I @d0bk773ki]

5/6
The next question is whether we need a model designed specifically for this.

OpenJev shows that direct scoring already works with existing LMs. But what happens if we train for decisions from the start?

That’s where purpose-built architectures — and maybe diffusion LMs?