Chamath treated data theft as a baseline risk at Facebook

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 1:41:41

   Well, I think this is the same that's been happening in Silicon Valley for a while. I remember when I was at Facebook a long time ago, I had this conversation. I said, guys, we just passed 3,000 employees. We have spies from at least four countries. And people were like, what are you talking about? And I said, well, just look at the number of H-1Bs we have and look at the distribution of them. And do you not think one of, you know, from country A and one from country B and one from country C and one from country D couldn't be a spy? Of course, the odds are you should assume that they are.

2. 1:42:14

   And, you know, my posture at that time was you should just assume that most of everything we do is being exfiltrated um so that's one example then another example i was a you know an investor in a company that had their stuff was supervised by itarp and uh had a hack and everything was exfiltrated like everything and uh i remember like you know we brought in um third-party consultants and the fbi and the nsa and um and i thought wow like you know how commonplace is this and they were basically it's it happens everywhere um so i guess what i'm saying is that you know this is a little bit um it's looks it's i think i guess it's written or report it to sound more like a cloak and dagger, 007 type stuff. But I think this is happening literally everywhere in every tech company that matters. The labs obviously matter.

3. 1:43:11

   And so the best way to deal with it is probably just to understand that some very minor distribution of your foreign folks are spies and can proceed accordingly. Well, and also recognize that how big the stakes are. You know, I think if you had spies like that at Meta or Facebook at the time, now the stakes are, what, 1,000, 10,000, a million times higher. So the pressure from foreign governments to install spies is going to be proportional to the amount that's at stake.

4. 1:43:45

   I don't think California was ever set up for this. You know, like this is the way nuclear secrets are protected. But California has always run on, you know, people leave a big company to start a small company, and that keeps the chain reaction of the economy going. So now you're like, well, we want to keep this all locked down like it's a nuclear secret because it can create terrible weapons. Well, we are totally not set up for that.

5. 1:44:08

   You'd have to rethink the entire structure. Yeah, I just hope this doesn't become sort of a national meme that starts to, you know, have us turn away foreign nationals. You know, I've always felt like when you get a Ph.D. or master's degree in the United States, we should be stamping a green card onto it, encouraging those people to stay here and work. If this turns the other way and we start, you know, sending people back, I think that's a problem for the U.S.

6. 1:44:34

   I should point out, though, this time is different. So historically, when there were secrets, for example, to Chamath's episode at Meta, the secrets themselves couldn't defend themselves. I don't know quite what was targeted, Chamath, at Meta at the time, but to the extent that superintelligence is being targeted now, allegedly, at Frontier Labs, I thought it was an interesting twist to the story, perhaps an inevitable twist, that OpenAI used their own AI.

7. 1:45:00

   order to understand how the device allegedly that was planted worked. And that to me looks like breadcrumbs, baby steps towards the superintelligence itself, preventing itself from being exfiltrated, preventing itself from winding up perhaps in undesirable hands, which if you look at nuclear secrets in the 1950s, nuclear secrets couldn't defend themselves. You look at games over oil who owned oil. Oil couldn't defend itself. But now superintelligence, in some sense, is the first critical and emerging technology that's potentially capable of defending itself territorially. And that's a new factor. Yeah. Yeah. Yeah. Yeah. The flip side of that is if you look at, you know, the Lama 3, Lama 4 models and the changes they made. Now, those are crappy models, but stick with me. If you read the research papers, because those were all open source,

8. 1:45:53

   The changes that worked and didn't work, you could summarize in like 10 sentences. So if you've bugged the conference room and you pick out the, hey, after $100 million of R&D and all these tests, the swig glue worked really well. That one sentence is so valuable to China or to somebody else coming up the curbs. It's really, really hard to contain something so small, yet so impactful on their research, their budget, their timeline.

9. 1:46:21

   It's tricky. It's really, it's hard to say, hey, AI, don't let that happen. Well, how's it going to, how's it going to defensively co-scale against something so slippery? The way, I mean, if I had to paint the sci-fi thought experiment, it is to the extent that the frontier labs are locked in a recursive self-improvement loop. It's the AIs designing their replacements, not humans sitting in a conference room.

10. 1:46:45

    So if it's just AIs designing further AIs, then the AIs can monitor their own digital conference rooms and make sure that there are no interlopers. Yeah, it kind of implies, though, that there's nobody at OpenAI that knows what it's doing to talk in a conference room. And that's kind of a scary, scary end to that sentence. Yeah, yeah, we don't have any idea what it's doing anyway. So bug us all you want.

11. 1:47:07

    That's not good. Chamath, we want to close out with a Chamath hot seat, if you're open for it. Great. All right.