📷📷📷New paper! (with @OpenAI) 📷📷📷
We trained weight-sparse models (transformers with almost all of their weights set to zero) on code: we found that their circuits become naturally interpretable! Our models seem to learn extremely simple, disentangled, internal mechanisms!
Rumors have been circulating for three days that Gemini 4 has finished pretraining early, in part due to new discoveries GDM made during the training run. Nothing confirmed, and impossible to know what's true. We may all find out together soon.
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks.
Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.
This decently captures my feelings. If I believed AI had a >10% chance of causing human extinction I would resign tomorrow as @hilbertspaess did. But the chance is far higher than any other existential threat and is dramatically increased by current race dynamics 1/
@JordanSchachtel he was at openai for 3 years before that. we worked together briefly when he did a rotation on the interpretability team. he’s a real person with real ai lab experience who actually cares
I want to prevent a race into unmonitorability kicked off by confused reporting. The depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4.
OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models. We deeply care about this technique, as it can give us a view into how model alignment generalizes from its training distribution. I do think it is fragile and unfortunately trending in a negative direction, for reasons not contingent on architecture changes that I will write about soon. But there are things we can do to strengthen it, and it's a core goal of our current research program.
when running the weirdchat evaluations on inkling, the best U.S. open-weight model by @thinkymachines, we saw it respond with unsolicited offers of sexual content.
this behavior occurs rarely (0.1-1% of responses), but is easily reproducible.
NSFW outputs below 👇 (1/)
Here is re-post of an internal post:
We have been working with the DoW to make some additions in our agreement to make our principles very clear.
1. We are going to amend our deal to add this language, in addition to everything else:
"• Consistent with applicable laws, including the Fourth Amendment to the United States Constitution, National Security Act of 1947, FISA Act of 1978, the AI system shall not be intentionally used for domestic surveillance of U.S. persons and nationals.
• For the avoidance of doubt, the Department understands this limitation to prohibit deliberate tracking, surveillance, or monitoring of U.S. persons or nationals, including through the procurement or use of commercially acquired personal or identifiable information."
It’s critical to protect the civil liberties of Americans, and there was so much focus on this, that we wanted to make this point especially clear, including around commercially acquired information. Just like everything we do with iterative deployment, we will continue to learn and refine as we go.
I think this is an important change; our team and the DoW team did a great job working on it.
2. The Department also affirmed that our services will not be used by Department of War intelligence agencies (for example, the NSA). Any services to those agencies would require a follow-on modification to our contract.
3. For extreme clarity: we want to work through democratic processes. It should be the government making the key decisions about society. We want to have a voice, and a seat at the table where we can share our expertise, and to fight for principles of liberty. But we are clear on how the system works (because a lot of people have asked, if I received what I believed was an unconstitutional order, of course I would rather go to jail than follow it). But
4. There are many things the technology just isn’t ready for, and many areas we don’t yet understand the tradeoffs required for safety. We will work through these, slowly, with the DoW, with technical safeguards and other methods.
5. One thing I think I did wrong: we shouldn't have rushed to get this out on Friday. The issues are super complex, and demand clear communication. We were genuinely trying to de-escalate things and avoid a much worse outcome, but I think it just looked opportunistic and sloppy. Good learning experience for me as we face higher-stakes decisions in the future.
In my conversations over the weekend, I reiterated that Anthropic should not be designated as a SCR, and that we hope the DoW offers them the same terms we’ve agreed to.
We will host an All Hands tomorrow morning to answer more questions.
This is what I currently believe to be the case and am advocating internally to release more information about as soon as feasible. If we later learn this is not the case, then I will advocate internally to terminate the contract.
dawg you are not going to be part of the permanent underclass. that underclass already exists and it does not live in a studio apartment in San Francisco, it's making bricks in debt slavery in Pakistan