The mystery continues... 🔎
Sam Altman a few minutes ago: 🎤
The world needs assurance that American companies developing increasingly capable AI will act responsibly, especially as the path of progress becomes steeper. Every frontier lab must uphold this, and there's no reason for any of us to come to work if we cannot.
We welcome a federal framework that establishes consistent safety requirements for frontier AI. But we don't believe we need to wait for an antitrust exemption or legislation to begin the work of providing this assurance. Consistent rules for managing frontier risks so we can maximize benefits are a good idea (and we are excited by ideas like independent auditors).
Years ago, companies like ours developed things like responsible scaling policies and preparedness frameworks. These were good for that moment, focusing primarily on the deployment of completed models, not what happens during their development process.
Today's shift to focusing on safe development and execution will require new tools. For example, at OpenAI we are now drafting explicit safety cases in advance for frontier reinforcement learning runs that we expect to significantly increase capability, in addition to the safety work we have long done in advance of model releases.
We hope other companies will learn from our approaches and offer their own; we think common standards for misalignment, monitoring, and safety will lead to better outcomes. We look forward to collaborating with our peers across the industry to draft the best version of these.
When we talk about 'pace,' we don't mean 'stop.' Progress has been rapid and will continue to be. But it needs to be slower than it could be; interventions like safety cases and monitoring involve significant costs.
Pace will be well worth this cost; no amount of American competitive pressure should justify recklessness, or allow capabilities to move forward before alignment and monitoring.
Where we will need our government's help is in international coordination. But first, we need to do what we can ourselves.
There are two ways in which AI progress could end very badly, and these are what we must avoid.
First, we could lose control of the future to AI. This is unacceptable; we are unapologetically on Team Humanity, and AI must always serve people. To ensure this, we need ways to ensure that alignment and safety techniques remain one step ahead of advances in model capabilities.
Second, we could arrive at a world with too much concentration of power. If exceptionally powerful AI is used by one person or one company to impose their worldview on everyone else, the results could be extremely dystopian.
Avoiding these two threats requires walking a narrow path in the middle; for example, one country could accumulate too much power. Another example is one lab ending up with too much power.
An attempt to reassure❓
Or perhaps the exact opposite❓
Even we are completely confused now... 👀
Meanwhile, the market reacts – as always – by selling first, then it will look for explanations.. 🛡