[_heimdall]

Two things that jump out at me here.

First, this assumes that they will know when they approach AGI. Meaning they'll be able to reliably predict it far enough out to change how the business and/or the open models are setup. I will be very surprised if a breakthrough that creates what most would consider AGI is that predictable. By their own definition, they would need to predict when a model will be economically equivalent to or better than humans in most tasks - how can you predict that?

Second, it seems fundamentally nefarious to say they want to build AGI for the good of all, but that the AGI will be walled off and controlled entirely by OpenAI. Effectively, it will benefit us all even though we'll be entirely at the mercy of what OpenAI allows us to use. We would always be at a disadvantage and will never know what the AGI is really capable of.

This whole idea also assumes that the greater good of an AGI breakthrough is using the AGI itself rather than the science behind how they got there. I'm not sure that makes sense. It would be like developing nukes and making sure the science behind them never leaks - claiming that we're all benefiting from the nukes produced even though we never get to modify the tech for something like nuclear power.

[davej]

Read the sentence before, it provides good context. I don't know if Ilya is correct, but it's a sincerely held belief.

> “a safe AI is harder to build than an unsafe one, then by opensorucing everything, we make it easy for someone unscrupulous with access to overwhelming amount of hardware to build an unsafe AI, which will experience a hard takeoff.”