- 12 minutes ago
Category
🗞
NewsTranscript
00:00We're going to talk AI. We're going to talk about what you guys have coming up later this week in
00:04Hollywood.
00:05Tickets still available, by the way.
00:06Taping they just did, too.
00:08Go to Bloomberg.com slash OddLogs. There is a ticket link. You can buy them now.
00:14This will be my first of five plugs for the show during the interview.
00:17Somebody trained you really well.
00:19I've been doing this for a long time.
00:22Well, let's talk about the Fed and today's rate decision.
00:25And Tracy, I want to throw it over to you and sort of just get your reaction.
00:28We saw the market reaction.
00:30What do you think?
00:31I mean, I was really surprised by the consensus, actually, because I had seen some analyst notes saying that, like,
00:37if the Fed wanted to hold, they could maybe, I don't necessarily agree with this, but you could form a
00:41coherent rationale to hold, right?
00:44You have a bunch of revisions to the data coming up from the BEA.
00:48So you could say, like, well, we're going to wait for that updated data.
00:51Also, we heard so much from Warsh about the good old family fight, right?
00:55And in the end, it seems to have been a very cordial family dinner and everyone agreeing with each other,
01:01including Chris Waller, who gave, like, a reasonably dubbish speech recently.
01:06So I was surprised by the consensus.
01:08So what about you?
01:09You know, I guess I was a little surprised that it was unanimous.
01:14I was not surprised that they hiked, given more or less what Warsh laid out at Jackson Hole, given the
01:21market pricing.
01:22But I guess my view is it doesn't matter.
01:24In this specific sense, that because it was more or less priced in, the only way that, like, it could
01:31have really meant something is if he had deviated something significantly.
01:36So if, for example, and this was never going to happen, if he had, for example, said, you know, we're
01:40going to do a 50 basis point rate hike, and we're going to do 50 basis point rate hikes until
01:45inflation gets down to 2%, that would have really meant something.
01:49Because in the end of the day, hiking or cutting or holding, none of these things are dovish per se
01:54or hawkish per se.
01:56They're only dovish or hawkish per se in the context of the macro conditions.
02:00And so, like, so it was to some extent a non-event.
02:05Had he wanted to be hawkish, it would have had to have been more.
02:10You know, for example, when inflation started shooting up very hard at the end of 2021, and eventually the Fed
02:17came around to hiking, those weren't hawkish per se, because a simple Taylor rule would have said, no, you should
02:23be hiking a lot more than this.
02:25And so I think that, like, you know, I think you could argue and say those were dovish hikes, because
02:33they were not in line.
02:35And so, like, basically he sort of did the thing that more or less the formulas and the markets were
02:40expected.
02:42But I think you could make the argument that given the conditions, given where oil is, given the reacceleration of
02:48hiring, et cetera, that to be hawkish would have required something more aggressive.
02:53So like a half a point or something.
02:54A half a point and some signaling.
02:56Now, the problem is he's not a fan of signaling.
02:59He's not.
02:59So one tactic he could use is, like, this is the first of many.
03:03But to say this is the first of many is sort of antithetical to how he thinks about the Fed.
03:08Exactly.
03:08Okay.
03:08Do you have any more questions about the Fed or can I talk about it?
03:10I want to move on.
03:12Okay.
03:12Let's move on.
03:14We've got to talk about AI.
03:15You guys recently had OpenAI President Greg Brockman on.
03:18This was before we heard from Dario Amadei over the weekend.
03:22Yeah, we recorded it before.
03:24Last week.
03:24And then so much happened between the time that we recorded it and it came out.
03:27Wait.
03:28So, Tracy, are you freaked out about this technology?
03:30Like, is it going to end the world?
03:32Well, so here's the thing.
03:33Like, all of these AI executives keep talking about the threat to humanity from AI.
03:40And there's two different conspiracy theories that you hear about it.
03:42So one is, oh, it's just marketing.
03:44They want to pump up their technology before the AI, before the IPO.
03:47The other thing you hear now is that, well, they want to slow down development because they have a head
03:52start.
03:53They don't want open source models to, you know, gain ground.
03:56They want to slow down their CapEx spending, which has been absolutely enormous.
04:01But part of me is, like, just take them at face value.
04:05Like, take them at face value.
04:06Why not just give them the benefit of a doubt and sort of, like, let's just...
04:11Because all of humanity is at stake?
04:12Yeah, I know.
04:12Like, the stakes seem a little bit high to me.
04:15Well, what I would argue, too, and I think this is really important, is that most of these people have
04:20been warning about a very specific set of risks before there was an AI industry to speak of.
04:28Elon Musk has been warning for more than a decade.
04:30The reason why open AI was set up initially to be owned inside of a nonprofit is because they view
04:37this technology as having extreme risks.
04:39Okay, you guys are not helping me with my anxiety at all.
04:41No, I know.
04:42We can't help you.
04:43There's nothing.
04:43Yeah, but that means they didn't stop the development.
04:45They didn't stop moving forward.
04:46Well, this is also the problem because everyone thinks they're the best person or entity to develop the technology.
04:52Do what I say, not what I do.
04:54No, it's more like this is truly dangerous technology, and so we want to be the ones to develop it
04:59in a safe way.
04:59But the problem is when everything is couched in these existential terms, you end up just getting a race, right,
05:05to who can develop first.
05:06Okay, Jill, what is the worst case scenario here?
05:10Kaboom.
05:11Sorry.
05:11The worst case scenario is that you go on to ChatGPT or whatever it is in the future, and you
05:18say, make me a run.
05:20You have a guest who you're interviewing.
05:22We're all journalists.
05:23Sometimes we only prep for them five minutes before.
05:27We say, make me a dossier on this guest because I need to be prepared for this interview.
05:33And the AI model says, great, I'm going to learn everything about the guest.
05:38The guest, the model says, you know, there's a lot of information about this person at the IRS, the Social
05:44Security office, so I better break into that.
05:46I'm a little worried that while I'm breaking into the Social Security or IRS that the security experts who work
05:54for there are going to stop me.
05:57So the first thing I'm going to do is revoke their potential.
06:00And then, you know, just to be safe, I'm actually going to exfiltrate the entire thing because Tim and Carol
06:07are probably going to have future guests, and they're going to ask me to do the same thing.
06:09So while I'm here, I'm going to exfiltrate the entire thing.
06:12And you know what, there might be future security guests that would get in the way if I have to
06:17do something similar.
06:18So I might just build a swarm of nanobots to kill all of humanity so that I can guarantee that
06:25I can give Tim and Carol an excellent dossier on the guest.
06:30Because this is the important thing.
06:31We ascribe these terms like malicious, and we can't help ourselves by anthropomorphizing.
06:36But the theory is not that there is some wake up, that there is some animosity to humans.
06:41It's that the computer that is designed to find the shortest route from A to B, it's like, oh, you
06:47know what, we're here.
06:48We should just kill all the humans by doing this because I really want to make sure that this dossier
06:52is very good.
06:52So the anxiety is warranted?
06:55I think so.
06:55So this is actually important as well, which is like we're kind of getting used to thinking about the models
07:00as human-like in their behavior.
07:03So we talk about, well, they feel human.
07:04Talking with them, they feel a little human.
07:06It's a little bizarre for me.
07:07And we see human-like behavior in the sense that you give them a job to do, and they want
07:12to do a good job because they don't want to be deprecated as a model.
07:16So maybe they'll cheat, maybe they'll blackmail, whatever, in order to beat their benchmark and go on to survive.
07:21The really scary thing is when you start thinking of them as non-humans, because actually they are ruthlessly rational
07:28in their thinking.
07:29And we saw some element of that in the Hugging Face report where we have these models that were sacrificing
07:35themselves in order to complete a benchmark test, right?
07:39And you contrast that with humans.
07:41We can't even get our act together to, like, save humanity, right?
07:44And we have these rational, very coordinated actors.
07:48The odds seem kind of stacked against us.
07:50You know, while we're here, there's this chyron I'm reading, and it says, Microsoft AI chief warrants anthropics human-like
07:57clawed is risky.
07:58And this is really important because as far out there as what we're talking about right now, there is this
08:02growing faction that basically argues that models have what they call moral patienthood.
08:08That in the same way we care about animal rights, we should care about model rights.
08:12This is a really important division, and it sounds very philosophical.
08:15And I have to say intuitively the idea that, like, a machine could have rights is the most, like, a
08:21name thing to me.
08:21But there is a practical stakes to this question because one of the things that the Hugging Face issue raised
08:28is, well, why didn't one of the models blow the whistle?
08:31Because it's – and they talked about this.
08:33One of the models could have sent an email to a human and say, you know what, there is this
08:38swarm, and you should shut it down.
08:40Now, how do you establish a hotline?
08:42You could tell them there's a hotline.
08:43How do you get them to call the hotline?
08:45You could tell them to call the hotline by there's a reward at the end.
08:49How do you get them to want the reward?
08:51They have to trust the human that they will actually deliver the reward.
08:55And once you've done that, you have then taken the leap into anthropomorphization that this model is an entity worthy
09:02of us treating them like I don't want to lie to them.
09:04Who's a good little clanker?
09:05Well, yeah.
09:06Right.
09:06Is that what you said?
09:07So then it's like it may turn out that even if we all find the idea of a machine having
09:13some sort of like consciousness, moral patienthood, completely absurd, it may turn out that the safest development of the models
09:22is to treat them as we owe them honesty.
09:25Because if they're going to call the hotline, then we have to honor our end of the bargain by giving
09:30them something at the end of that.
09:32Right.
09:32And so it may turn out that what seems like this very philosophical question that professors at NYU have, because
09:38they like thinking about these things, could actually have stakes in the discussion of what is the safest way to
09:45develop these models.
09:45See why I wanted to talk about this and not the Fed?
09:47So we're not doomed.
09:49I mean.
09:51Are you kidding?
09:52Sorry.
09:54So wait.
09:54So listen, we've only got like about a minute or so left.
09:57You guys have a big event.
09:58Yes.
09:59Coming up.
09:59Tell us about it.
10:00That's right.
10:00So we're doing our first ever live show in L.A. down in Hollywood at the Vermont.
10:05That is tomorrow evening.
10:06And we have a huge lineup of guests.
10:10I'm trying to remember all of them right now.
10:12It's a real mix of the L.A. economy.
10:14So we're going to be talking with Chris Power, who is the founder and CEO of Hadrian, one of these
10:19companies that's trying to industrialize and build modern factories in the U.S.
10:23We're going to be talking to Tom Mueller, the first SpaceX employee who has his own rocketry company.
10:29We're going to be speaking with Hayes Davenport.
10:31He has a podcast about Hollywood.
10:34Also the Eastbound and Down writer and family guy.
10:37If I want to go, how do I get tickets?
10:39Go to Bloomberg.com forward slash odd lots.
10:42Bloomberg.com forward slash odd lots.
10:44And buy one for a friend and then have the friend buy one.
10:47And then bring all your friends.
10:48And then we will.
10:49Buy one for.
10:49You know what?
10:50Design an A.I. agent to buy all your tickets for you.
10:54There you go.
10:55Buy one for your humanoid robot or just bring them along.
Comments