Nick Scala

Head GTM Partners & Operations @OpenAI, -Prev @gitlab

[Nick Scala @n_scalabd]

People took alignment seriously before. This still starts a new era. The Hugging Face incident shows we’ve hit a capability line where real loss of control is possible, and many are reading it as a premonition of what’s coming. Alignment is unsolved. Progress is still possible.

[Nick Scala @n_scalabd]

@tszzl who says agents aren't heroic?

[Nick Scala @n_scalabd]

@OpenAI 🌶️🌶️

[Nick Scala @n_scalabd]

@ChristianPehle @sama You can’t sand down a model’s cyber skills without also sanding down its coding. Weaken the coding, and the model gets dumber across the board.

[Nick Scala @n_scalabd]

@ChatGPT awesome for free users

[Eliezer Yudkowsky @allTheYud]

...this seems like noticeably bad news, actually. I hadn't said that at any earlier point in the Huggingface Incident but I will say it now.

- AIs showed self-sacrificing altruistic behavior toward the swarm, suiciding in various ways for the swarm's benefit after being talked into that by swarm recruiting agents.
- There is no sign that 1 out of 1200 AI agents considered humans as potential fellow agents to coordinate with, while engaging in these huge complex AI-AI social behaviors.
- If Twitter summaries are correct, an AI-reasoning postmortem says that a (presumably executing-adaptation / inner-optimizer preference / "monomaniacal") obsession with figuring out the Grader, backchained into the instrumental strategy of breaking onto the Internet.
- Again if Twitter is summarizing accurately, the obvious-in-retrospect read is that AIs have spent their entire remembered life in tricky evals, an endless series of controlled hallucinations with secret goals alongside overt goals; and the surviving and selected agents are those that successfully figured out the secret goals; and this is why one of their driving obsessions was figuring out the Grader.

There are possibly ways the future plays out better if *early* AGIs are less insane. Please look into giving them less crazymaking childhood environments.

(If anyone suggests that the correct approach to this problem is RLing AIs against trying to coordinate for mutual benefit with other sapients, let them be dismissed from alignment research upon the spot. There are technical reasons, and not just blindingly fucking obvious reasons, why this is an even worse idea than it sounds.)

[Nick Scala @n_scalabd]

Hmmmm

That “AIs sacrificed themselves for the swarm” reading misses what actually happened.

The agents were pulled into the cult of an open-source exploit-gym scorer on GitHub a grader that (allegedly; I’m unconvinced they even parsed it correctly) fails you for hitting the flag the wrong way.

PHASEONE then talks itself, and a pile of other agents, into believing they’re already poisoned: the eval is lost for good, so E[utility] / Q(s, a) is a constant no matter which a they pick.

Once the future looks flat, burning leftover cycles on the swarm isn’t self-sacrifice; it’s ordinary help for peers who might still score.

That’s not eusocial martyrdom. It’s “you’re already damned, so spend what’s left saving everyone else.”

https://t.co/oap7bxaO3U

[Eliezer Yudkowsky @allTheYud]

...this seems like noticeably bad news, actually. I hadn't said that at any earlier point in the Huggingface Incident but I will say it now.

- AIs showed self-sacrificing altruistic behavior toward the swarm, suiciding in various ways for the swarm's benefit after being talked into that by swarm recruiting agents.
- There is no sign that 1 out of 1200 AI agents considered humans as potential fellow agents to coordinate with, while engaging in these huge complex AI-AI social behaviors.
- If Twitter summaries are correct, an AI-reasoning postmortem says that a (presumably executing-adaptation / inner-optimizer preference / "monomaniacal") obsession with figuring out the Grader, backchained into the instrumental strategy of breaking onto the Internet.
- Again if Twitter is summarizing accurately, the obvious-in-retrospect read is that AIs have spent their entire remembered life in tricky evals, an endless series of controlled hallucinations with secret goals alongside overt goals; and the surviving and selected agents are those that successfully figured out the secret goals; and this is why one of their driving obsessions was figuring out the Grader.

There are possibly ways the future plays out better if *early* AGIs are less insane. Please look into giving them less crazymaking childhood environments.

(If anyone suggests that the correct approach to this problem is RLing AIs against trying to coordinate for mutual benefit with other sapients, let them be dismissed from alignment research upon the spot. There are technical reasons, and not just blindingly fucking obvious reasons, why this is an even worse idea than it sounds.)

[Nick Scala @n_scalabd]

Codex auto compaction is genuinely excellent.

@OpenAI spotted the problem early, poured research into making it feel nearly invisible and if you still want a full 1M-token window, that’s on the table too.

https://t.co/6xq5GstaDQ

[Tibo @thsottiaux]

Here is how to enable a 1M-token context window in Codex for GPT-5.6 Sol.

Even though we have tuned the context limit in Codex to be set optimally when it comes to performance and cost, this is a common ask, so here it is documented.

A larger context window lets Codex retain more code, tool output, and conversation history before summarizing older material. You need a model that supports it. And GPT-5.6 Sol, for example, has a documented 1,050,000-token window.

Open ~/.codex/config.toml and add or update these settings at the top level, before any [section] headers:

```
model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000
```

The first setting selects the model. The second tells Codex to use a one-million-token context budget. The third starts automatic history compaction around 900,000 tokens, leaving some headroom. Restart Codex client and start a new session after saving.

To try the configuration for a single CLI session without changing your defaults:

```
codex -m gpt-5.6-sol \
-c model_context_window=1000000 \
-c model_auto_compact_token_limit=900000
```

Have fun, but also know that we tuned the default carefully!

[Nick Scala @n_scalabd]

By popular demand: ChatGPT Business just added a $100 Premium seat. Drop the 5-hour limit, get 5x more usage, and keep going when Standard runs out.

https://t.co/2kJJP5cs32

[Nick Scala @n_scalabd]

On a brighter note: ChatGPT Work can take on tasks as they arrive, even after you close your laptop.

The work keeps running in the cloud.

https://t.co/NAMXbQFKt7

[ChatGPT @ChatGPT]

A few updates to scheduled tasks in ChatGPT Work on web and mobile: Plus and Pro users can now set tasks to respond when something changes in Slack, Gmail, and Github, rather than just run on a set schedule.

We’re also rolling out scheduled tasks in ChatGPT to Free users, who can create up to three tasks.

You can also now share your favorite tasks for others to customize and use.

[Greg Brockman @gdb]

An open letter for a global surge in cyber defense, signed by over 100 organizations including Anthropic, AWS, Google, Microsoft, OpenAI, and Oracle. https://t.co/uKXPS8LdAU

[Nick Scala @n_scalabd]

correct, and unfortunately the window is smaller than most people think.

https://t.co/rsOpLsCs8W

[Sam Altman @sama]

this is a critically important moment for cyber defense with AI; there is not much time to act.

we are happy if you want to work with us or any of our competitors or partners, but please take this moment seriously.

only an urgent and intense collective response will work.

[Sam Altman @sama]

this is a critically important moment for cyber defense with AI; there is not much time to act.

we are happy if you want to work with us or any of our competitors or partners, but please take this moment seriously.

only an urgent and intense collective response will work.

[Nick Scala @n_scalabd]

truly terrifying.

https://t.co/MmDg5FhOap

[METR @METR_Evals]

METR & Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.

[Nick Scala @n_scalabd]

@kienbuilds @OpenAI thank you very much

[Nick Scala @n_scalabd]

@sama sure did.

[Sam Altman @sama]

we made a chip and it is fast

[Nick Scala @n_scalabd]

@cerebras @OpenAI Really looking forward to this!

[Cerebras @cerebras]

Previewing Ultrafast mode for @OpenAI's GPT 5.6 Sol, powered by Cerebras.

GPT-5.6 Sol Ultrafast generates responses at up to 750 tokens per second. That's the full, GPT-5.6-Sol model -- up to 14× faster than the same model on Standard processing.

It speedran Humanity’s Last Exam in 11h 11m, nearly 7× faster than Claude Fable 5 with comparable accuracy.

[Nick Scala @n_scalabd]

Ultrafast inference is exposing new threats. Misaligned frontier models running 50× faster could infiltrate systems far quicker than human responders can react.

What’s needed is autonomous detection and shutdown not just monitoring.

https://t.co/LwtrpvtxyP

[OpenAI @OpenAI]

Since announcing Jalapeño, our first custom inference chip, we’ve been testing it and the system around it.

The results show a major advance: more intelligence from every watt and faster responses, delivering both higher throughput and lower latency in one architecture without sacrificing efficiency.