The consequences of Google’s more streamlined approach are not yet clear. Its technology has lagged OpenAI’s self-reported metrics when it comes to identifying content that is hateful, toxic, sexual or violent, according to an analysis that Google compiled. In each category, OpenAI bested Google tools, which also fell short of human accuracy in assessing content.
Google listed copyright, privacy and antitrust as the primary risks of the technology in the slide presentation. It said that actions, such as filtering answers to weed out copyrighted material and stopping A.I. from sharing personally identifiable information, are needed to reduce those risks.
The way I’m reading this, Google is behind on RLHF, and worried about getting blasted by EU fines. Honestly, those aren’t humanity-dooming concerns, and it’s not a huge deal if they brush them off. However, you’re right that this is exactly the race dynamic AI safety has warned about for years. It would be good if the labs could reach some kind of agreement on exactly what kinds of requirements have to be met before we reach “actually dangerous line do not rush past”. Something like OpenAI’s Charter:
We are concerned about late-stage AGI development becoming a competitive race without time for adequate safety precautions. Therefore, if a value-aligned, safety-conscious project comes close to building AGI before we do, we commit to stop competing with and start assisting this project. We will work out specifics in case-by-case agreements, but a typical triggering condition might be “a better-than-even chance of success in the next two years.”
Maybe there ought to be a push for a multilateral agreement of this sort sooner rather than later? Would be good to do before trust starts breaking down.
Just quoting from the NYT article:
The way I’m reading this, Google is behind on RLHF, and worried about getting blasted by EU fines. Honestly, those aren’t humanity-dooming concerns, and it’s not a huge deal if they brush them off. However, you’re right that this is exactly the race dynamic AI safety has warned about for years. It would be good if the labs could reach some kind of agreement on exactly what kinds of requirements have to be met before we reach “actually dangerous line do not rush past”. Something like OpenAI’s Charter:
Maybe there ought to be a push for a multilateral agreement of this sort sooner rather than later? Would be good to do before trust starts breaking down.