Expert opinions on AI Risk

We do this pretraining on human data, and then we get something that… understands human values fairly innately now.

If it takes over that's what will happen. [...] We don't know how to do that. We don't even know if it's possible.

What this proof shows is that there is no finite set of guardrails that is universally robust against adversarial prompts.

Capability testing must be the primary input for open-release decisions because open-weight mitigation remains unreliable, and labs and governments must identify models that are too dangerous for unrestricted release.

However, the solution here is absolutely not to restrict open model development in the name of security—cyber, national, or otherwise.

Developers publish broad, top-level safety and ethics frameworks that sound reassuring, but are publishing limited empirical evidence needed to actually understand the risks.

The draft implementing regulation should clearly differentiate between API access, which should be the preferred method for safety evaluations where access to documents and relevant information are deemed insufficient, and “further” means of access, ...

You can’t patch a released, open-weight model. This is the property that breaks the assumptions of the prior model of responsible disclosure. Open-source software can be patched in place; a deployer updates and the hazard closes. An open-weight model...

It is unquestionably both. We can use the frontier models but also open weight models to prove a solution which gives defenders powerful capabilities and on their terms.

Open-source AI is more than a catalyst for commercial innovation. It is a strategic advantage for the U.S. military, enabling American companies to push the technology forward while giving the DOD a foundation it can inspect, secure, adapt and deploy...

The challenge ahead is therefore not merely regulatory. In a deeper sense, it is about governing rationality itself, ensuring that as these systems grow more capable, they remain accountable, interpretable and anchored in values that cannot be reduce...

We have a very natural instinct to say: we want as much interpretability as possible. One of the sci-fi worry cases is that agents start speaking in languages to each other that we don't understand. [...] I bet there are ways to define boundaries tha...

[...] principles requiring that AI be explicable are misguided. We should be deciding which decisions require explanations. Automation is still an option;

There appear to be no obvious technical barriers to creating AI systems whose computational and architectural features could give rise to consciousness.

Because of how human nature has been expressed throughout our history and continues to express itself today, developing conscious machines with what for humans might be undetectable and incomprehensible minds is a monumentally dangerous idea.

There's no good reason to think AI is, or could ever be, conscious.

AI has gone from "dumber than a child" to "expert level" in a few years in many domains. The gap between "expert level" and "above top geniuses" is smaller, so we expect it to take less time. This has been a pattern in fields like chess and Go, where...

Throwing more words and GPUs into the word-guessing program won't make it sentient. [...] one of our mares gives birth to a locomotive.

I would like to see a CERN for AI [...] where all the top scientists come together and see whether they can make a super intelligence. And, if they can, they contain it into a system where it can't just go out and persuade people to let it run the wo...

Well, there are positive signs, for instance, like Henry Kissinger was recently suggested in foreign affairs that the US cooperate with China on this issue now, but before it's too late. So I think some people are recognizing the the importance of tr...

I call on the governments of the world to take immediate action to end this danger by permanently banning the development of artificial general intelligence and machine superintelligence.