21/05/2024
The issue around whether the similarity between the "Sky" voice OpenAI used for its GPT-4o demo and actress Scarlett Johansson's is "purely coincidental" as OpenAI says, is not the only case where OpenAI's public statements have been called into question. A little under a year ago, OpenAI promised to devote 20% of its compute to a new team dedicated to ensuring what it said would be the most dangerous kind of AI could be safely controlled. Last week, OpenAI disbanded that "Superalignment" team after its two co-leads left the company. Half a dozen sources familiar with the team's work tell me that OpenAI never delivered on its promise to give the team 20% of its computing resources, and in fact repeatedly denied the team's request for access to computing power for its research. Regardless of whether you agree that "superintelligence" (a theoretical technology that does not yet exist) is the most important AI safety issue, this is yet another example of OpenAI saying one thing in public, and seeming to do something else behind closed doors
