Skip to content

Remarks on Sam Altman’s Interview on the All In Pod

OpenAI CEO recently appeared on The All In Pod. I note all of this with high regard to Altman’s capabilities and with a generally positive disposition towards what Altman is doing.

➡️ Ability to Answer Specific Questions with General Answers

Altman rarely refuses to answer a question. Rather, he has the ability to abstract the question and provide insight at a higher level, while remaining safe from having to opine on the specifics of the matter at hand. e.g. note the response to the question from David Friedberg on SORA, the video model.

In this regard, the interview is hollow as Altman cannot address interesting specifics. Although, this is likely more praise than criticism of Altman’s abilities.

➡️ Approach to data licensing

Altman turns the tables by suggesting that copyright laws should restrict the use of language models at the point of use (e.g. inference), not during training.

This would push legal and licensing pressure away from model makers (OpenAI) and towards model users.

While this favours OpenAI, it likely also better matches with the current approaches to copyright, and also how law is applied to humans. There are no barriers to me learning from Taylor Swift’s music, but there are rules around using elements of it “in production”.

➡️ Regulating “Recursive Improvement”

Altman suggests that models that can recursively improve be regulated. Yet, any optimization problem is a form of recursive improvement. So, “Recursive improvement” is a poor definition.

I sense that Altman is grasping at straws because he needs to thread the needle between:

a) Being responsible, which can only be done through small steps of trial and error – not through some kind of predictive or advance framework.

b) Appearing responsible, which requires a projection of false clarity around what the risks are and how they will be addressed.

My view (and I say this with partial confidence) is that common law systems are the best tool for dealing with technological risks. A common law system is adaptive and evolves based on precedence and liability. In this sense, it is both backwards and forwards looking.

Leave a Reply