Altman was interviewed by Nicholas Thompson, CEO of The Atlantic. It was quite a strong interview (though I later learned The Atlantic had signed a deal with OpenAI for content access the day before).
Here's my summary:
Responsible AI
OpenAI are just creating models and don’t know where it’s going
Asked if it's responsible to release models before understanding them, Altman said we don’t understand the human brain but know that humans can follow rules
Asked whether spending on development should match spending on safety, Altman said he didn’t know what that means or how money would be spent on safety [laughter from the room]
“You can have nice policies here and there” but making GPT-4 safe would involve "a hard time identifying what is safety and what is capability"
AI safety is like aviation: you want to get somewhere but don’t know which parts of an aircraft are “safety” and which are “capability” [more laughter]
Putting the Superalignment team closer to work teams will be a positive development
Training Data
Progress requires finding enough quality data or ways to improve training AI on small amounts of data
Generating synthetic data to train models would be inefficient
Using Human Voices
Human-like voices are important for user experience
Altman is not sure that informing users voices are AI-generated with a beep would help
On the Scarlett Johansson controversy: Altman apologised for the confusion, but insisted it is not her voice and had nothing else to add
AI Governance
Altman seemed unaware of global governance initiatives except on “short-term issues like elections” [more laughter]
He believes we will need to re-write the social contract because “the whole structure of human society will need to be changed”
However “it’s very difficult to do this in theory in advance: we have to watch how these systems evolve and work this out empirically”
It would be “a really great project for the UN” to work on alignment for humanity [more laughter: it was a UN/ITU event]
Thompson read an old quote from Altman on governance: “why do these f***ers get to tell me what to do?” Altman wouldn’t say more about that
Op-Ed in The Economist
ChatGPT was only released as a low-key research project with no awareness of potential impact [more laughter]
Altman disagrees with his former board members' recollection of events
On Humility
Altman insisted that “personally I have a huge sense of humility” [more laughter]
His key takeaway was that there is an incredible upside to AI and a moral duty to bring that to the world, but with serious safety concerns.
A fascinating insight into the man who pushed AI to the forefront of public debate. The link to the recording is in the comments, but being in the room was a different experience (especially when you could hear the laughter...).