It's not just the f*cking sandbox

So who is this for?

> Of course we need strong sandboxing. We also need to secure the services it can reach, the tools it can invoke

That’s… the sandbox. The sandbox includes everything in the sandbox.

> So is it as easy as just putting it in a sandbox? No, it’s not. But I do believe it is tractable across three dimensions

Those 3 dimensions are sandboxing, alignment and monitoring.

I’d suggest that monitoring is a part of any sandbox. That leaves “the sandbox” and “trying to get the model to not escape the sandbox”?

Sure, agreed.

> I literally missed my sister’s wedding a few weeks ago to help clean up after some of the recent incidents

If you're not paid for that, you should just leave. But my understanding is that OpenAI employees cannot complain about their conditions, can they? Hard to feel for a millionaire saying "I earn orders of magnitudes more than you, but I work 50% more", I think (I don't know if it's the case here or not).

> all I hear on X is how I “don’t know how to configure a sandbox” or that I should be in jail

I think it can be two very different comments. "They don't know how to configure a sandbox" is condescending. "Someone should be in jail"... well that can make sense. When a bridge collapses and people die, then we say the same: "someone should be in jail". The different with the software industry is that nobody gives a shit about the harm they cause. Civil engineers have to, or they go to jail.

The point here being that maybe, just maybe, AI labs should not be shipping stuff they cannot secure. It's hard to build a dam that does not collapse, but that's not a reason for civil engineers to vibe-build a dam and then say "ok it collapsed, but look it's hard to get right".

> Very few organizations can survive that scale and growth, and yet OpenAI did a very good job handling it (and should be commended).

So if OpenAI results in people dying, it is fine because "really, it's very hard"? I'm trying to think about a company that would build dams at such a scale and then use it as an excuse for only a fraction of them collapsing from time to time.

> To say it lightly: this surprised the fuck out of us. We had not expected it this soon. Suddenly we weren’t dealing with just a small jump in capabilities; we were talking about a different sport altogether.

Last example: imagine a geoengineering company coming out with a "solution" to global warming. Suddenly it goes out of control and it screws us even more. Would we say "please, have some respect, having that kind of impact is really hard and it honestly surprised the fuck out of us", or would we say "maybe you shouldn't have done something dangerous you didn't understand"?

> I recently watched the movie Sully.

Okay I can't resist it. It's inverting the argument. In Sully, we accept the fact that they took risks they shouldn't have because it turns out that they saved lives (and they prove that doing what they should have would have resulted in more casualties).

Now if the plane had crashed into a tower and killed thousands of people, of course we would blame the pilots!

OpenAI resulting in security incidents is NOT AT ALL the same as Sully saving lives by not following the rules. What a weird comparison that is?

If Starship crashed in Central Park, I wouldn't expect a SpaceX engineer to write a blog post saying "Look, luckily it didn't kill anyone, and I'm sick of reading comments saying "someone should go to jail". It's very hard to build a rocket, and people who think otherwise are just misinformed. Give us some slack, crashing into Central Park surprised the fuck out of us".