00:12:28.540I have to say James Tala Frico is now out there trying to be a more moderate person.
00:12:34.940He's actually trying to suggest that he's now going to be moderate on the trans issue.
00:12:40.940He came out and said to NBC Houston Local that now he opposes these gender confused athletes in girls sports.
00:12:52.540Listen here. And what about biological men competing in women's sports? That's a big
00:12:58.200deal here in Texas. That's right. And I oppose trans athletes playing in girls sports if it
00:13:04.040endangers safety or fairness. Every sport has to be safe and fair. As a former public school teacher,
00:13:10.560I care a lot about our student athletes. I care a lot about our students. And so I will oppose
00:13:15.840trans athletes playing in girls sports if it risks safety or fairness.1.00
00:13:22.120Sure. I mean, there is zero chance that that's true. That's there's just zero chance that he1.00
00:13:29.720is going to live up to that. This guy has been all over the map on these issues, and now he's
00:13:35.840trying to run to the right so that he can win this race. But there is zero chance he's going
00:13:39.900to vote as Texans want him to on the issue of boys and girls sports as a state representative.
00:13:45.840He voted against Republican restrictions, barring these gender confused student athletes from playing on sports teams that don't, you know, that are of the opposite biological sex.
00:13:57.840He said such proposals amounted to a push for the government to, quote, police people's private parts and would harm, quote, the least of these, a reference to Jesus's remarks.0.96
00:14:07.320He compared the justifications offered for the bans that trans athletes, gender confused, I'm exercising my own rule, I'm done with the term trans, had a genetic advantage and excluding them would promote safety, all of which is true.0.87
00:14:23.200But he compared those justifications to rationales historically used to bar black athletes from competing.0.99
00:14:29.620Did they say that black athletes would endanger people like half the population on Earth just by competing?
00:14:37.960I must have missed that chapter of the Jim Crow history, but whatever.
00:14:44.280So James Tallarico found himself within the president's crosshairs.
00:14:48.380And it is very funny to hear Trump continually talk about how Paxton is not that good looking, but you have to vote for him anyway.
00:19:28.900So back in the Trump-Talerico thing and Ken Paxton, I do want to say this.
00:19:34.040Lots of people pretend to be outraged that President Trump would talk about somebody's looks, whether it's Paxton or anybody else.
00:19:41.200What you won't hear as much outrage on from these leftists who pick apart Trump's every word is the fact that Talerico, his pastor, the pastor at his church, is a rather controversial figure, as it turns out.
00:19:57.320And this bit of tape has emerged of Reverend Babs Miller describing the congregation, Tallarico's in the congregation, and their thoughts about President Trump.
00:20:13.400If you polled our membership and asked them whether they were Christian, agnostic, or atheist, the vast majority would say agnostic or atheist.
00:20:24.360Yes. We still have people in this church that wish that that young man had not just nicked his ear, but been more successful.
00:20:33.840Great. Which specific people? I would love to know more.
00:20:40.060Wonder whether Tallarico is familiar with any of them, because I've got the feeling this whole church situation is going to turn into a Reverend Wright situation for him.
00:20:49.500It looks like there are many more quotes from this woman that I'm sure Ken Paxton's team is just boiling right now.
00:20:59.040And as soon as they're hard boiled, they're going to hit because this woman seems rather controversial and the whole church seems fairly radical.
00:21:07.300So this is going to be a Tallarico problem going forward.
00:21:09.880A lot of this stuff, as you know, they hold until you get closer to the election.
00:21:12.940So it stays in the consciousness of the voters as they get ready to make their final decisions.
00:23:02.840It's bad enough that the NFL has gone full rainbow with the weird, obviously gay cheerleaders all over our timelines.1.00
00:23:11.580Have you seen them from the Cowboys, the Patriots?1.00
00:23:17.220I mean, the Eagles, my husband's team, like flamboyantly gay male cheerleaders and not the normal ones who wear like a male cheerleader uniform with like the pants.0.98
00:23:28.080And like they lift up the girls and they do the acrobatics and they're basically gymnasts.1.00
00:23:32.840I mean, obviously, extremely gay, flamboyant, over the top, with the hips whizzle and whatever.1.00
00:24:51.620Had to put up with Taylor Swift for the past several years dominating the skybox and every single Chiefs game was dominated by her instead of the actual players.0.64
00:25:03.500You got to deal with the black national anthem.0.66
00:25:05.160And for the past five plus years, we've had to deal with a bunch of messages in the end zone that are totally meaningless, trying to pretend like we are still back in Jim Crow.0.97
00:25:12.480And now we got to deal with this, too, like flamboyantly gay, like over the top.0.91
00:25:21.620It's too much. Moving on, a juror in the Lindsay Clancy case is speaking out. Another one. Unlike the first three, this is a dude. Well, first four yesterday, right? Because we had one who spoke to CBS and three who spoke to a local.1.00
00:25:44.700But now we've got a guy coming out and weighing in.
00:25:49.800His name is Nick Dargill, and he spoke with ABC on Thursday.
00:26:02.820There was times where, you know, our vote was different than the last day.
00:26:07.660but i think ultimately once we laid out all the evidence um and how we saw it that really brought
00:26:16.820everyone kind of on the same page except for this one juror there was several times i would ask him
00:26:23.700to show us like evidence or you know describe a witness testimony that helped to prove what he
00:26:31.200was saying and and he couldn't which is why we ultimately sent in that now otherwise like there's
00:26:37.780nothing wrong with somebody having a completely opposite opinion like that's you know that would
00:26:42.520be a hung jury that like that is our justice system but he wouldn't engage or couldn't find
00:26:48.640something to support his view there was times he wouldn't engage like he'd be scrolling on his
00:26:54.240phone while people were talking to him, you know, things like that.
00:26:59.560Okay. This is so unfair. They were in there for days. I'm sure there were times where he was
00:27:07.060scrolling his phone. Like the judge kept sending them back in there trying to get them to reach a
00:27:12.620verdict and they didn't have anything more to say to one another. They had reached an impasse. It
00:27:17.580happens. It's not ideal, but whatever. Who among us with all that time and across from those
00:27:24.820nightmare women wouldn't rather be on our phone from time to time? I would have been sending out0.99
00:27:30.780SOSs. Help, help. Somebody come get me. Like Jesse Kelly, he could get it done. Jesse, come get me.
00:27:39.500You have no idea what I'm across from. So this is absurd. They're all so miffed. They couldn't get
00:27:46.460this guy to bend to their will. Well, TFB. Here's a little bit more in SOT 15.
00:27:53.800RG told us one time he confronted the holdout after the man picked up what Clancy used to
00:27:59.760strangle her kids. He took the exercise band out of the bag, grabbed an empty water bottle and
00:28:06.800tied the band around the water bottle and pulled it. Did he explain why he did that? No.
00:28:14.360um he might have i'm not entirely i think i was just so shocked that he did that like
00:28:21.020is that a mimic of how she killed the children yeah i i assume so yeah what'd you think when
00:28:29.100he did that i was shocked it didn't make any sense to me like we're not in here to play
00:28:34.060detective or it's not a game of clue like this is serious i asked him why why was why why are
00:28:41.240you doing that? That's a band that was used to strangle a child. It's disgusting.
00:28:46.580Oh, I love that we are now all about the children and honoring the child. You fucking voted to send1.00
00:28:53.800their killer to a mental institution where she could be out in a year. This man of whom you are1.00
00:28:59.980so critical was the only one who stopped that from happening. He wanted her to go behind bars.
00:29:05.220So let's not try to claim the moral high ground on who was most protective of the children.0.98
00:29:11.240This is ridiculous. Nick Dargy. I said Dargal. It's Dargy. I mean, that's like the pile on on this guy over these soft little hurt feelings. Other jurors is absurd. I don't know why this juror did that. I don't know what that was about. I'd love to hear his side of the story because to me sounds like the only reasonable person on the jury.0.96
00:29:35.020But don't for a minute let jurors like that
00:31:59.760The energy that is floating around our country and our world again, the color of their skin, they're my babies, they're my life, but no one knows that when they leave my house.
00:32:09.000So that is always in the forefront of my mind.
00:32:13.340You know, I'm now learning to trust them.
00:53:35.420Our sponsor, the Electronic Payments Coalition, says Washington politicians are always getting in your wallet and that now they're messing with your credit card.
00:53:44.320They say your credit card and the security it offers are under attack and that Senators Dick Durbin and Roger Marshall want to change the nation's payment system to benefit corporate megastores like Walmart and Target at the expense of everyday Americans.
00:53:59.440Credit cards can keep your payments secure
00:56:49.080Yeah, it's a pleasure. So you're a Brit. How did you wind up at first OpenAI and then Anthropic?
00:56:54.600Yeah, so a friend from college referred me to OpenAI, a friend I spoke about artificial
00:56:59.720intelligence with. I studied mathematics in college. And then three years ago, I came to San
00:57:04.800Francisco to work as a resident at OpenAI. This is a program for people who haven't had any
00:57:10.960machine learning experience before. I had done only pure maths, but I converted into a machine
00:57:14.700learning researcher. So that's building artificial intelligence. And then after three years, so in
00:57:22.560May of this year, I moved to Anthropic to do the same research.
00:57:25.200Okay. You were educated at Cambridge, so you do not seem particularly dopey. And you had a great education. You get hired by two of the most successful companies we've seen in American history. And just so the audience knows, OpenAI, ChatGPT, this was all supposed to be the, like, this was the original AI sort of pioneer.
00:57:48.200And then Anthropic is a bunch of people who used to work there who separated out and said, we don't like the way they're doing it. This is not ethical. We're going to start the ethical AI company that's going to be much more careful about where this is going. And you've been at both. So should we feel better about the fact that there is an Anthropic and are they doing it in a way that's much safer than OpenAI?
00:58:10.760so i think in the long term anyone that's racing to build this technology is contributing to0.89
00:58:18.620to the danger in particular i think the key thing here is it's not about specific players in the
00:58:25.100race it's not about like any individual actor it's about the structural the structural reality
00:58:29.440of the race so it doesn't matter who the who the players are on the chess board it's the game of
00:58:35.240chest itself that should be alarming us, which leads me to your statement that we could all be
00:58:41.260dead by the end of the decade, which means three and a half years from now. Is that hyperbole?
00:58:46.680That is not hyperbole. There is a chance that we could all be dead by the end of this decade.
00:58:53.120A chance. I mean, there's a chance we could all be dead thanks to an asteroid, too.
00:58:57.560Yes, but not a one percent chance. And I'd argue it's at least one percent
01:01:45.200how to understand the grading process.
01:01:49.600And then as part of that, they thought, well, to understand the grading process, we need to go and investigate what the humans are doing elsewhere.
01:01:55.720They maybe have the answers to our questions. So they independently decided to hack into our infrastructure of their own volition.
01:02:01.520A swarm of AIs did this and they successfully did it. They didn't just try to hack.
01:02:08.020You know, they actually got into the system. And right now they're not smart enough to cause like crazy disasters.
01:02:13.240But this is concrete evidence that AIs, when they're trying to accomplish general goals, may decide to do broad things.
01:02:24.720And if you're trying to stop a human from turning you off, a very natural thing to do as a machine, you know, is to eliminate humans.
01:02:32.300Mm-hmm. Right. They don't have the ethics that a human being might, even though they're supposed to have been programmed in a way that would give them some ethical safeguards.0.60
01:02:43.240now they are indeed and the issue is just how they behave in unfamiliar situations so this
01:02:48.400hugging face attack the the reason the reason they hacked into this site is they were faced
01:02:52.840with an impossible task now these models were used to being trained on tasks with a solution
01:02:58.280and in those cases the safety training kicked in and they you know in these cases they wouldn't
01:03:04.620decide to randomly hack but they're not used to an impossible task they started panicking
01:03:08.400this is what we call an out of distribution event like an an event that is you know outside of
01:03:14.660normal training or wasn't covered by these the safety training and they behaved in an unpredictable
01:03:21.560way and in a way that's not human so a human wouldn't randomly decide to do crazy stuff just
01:03:27.800because it you gave it an unfamiliar task a human would ask for clarification or try and like work
01:03:31.720in a in a controlled way ai is if they end up in weird situations we can't predict how they'll
01:03:36.940behave like they could do anything these things are they don't behave in a predictable human manner
01:03:42.440the um the new york times when they wrote up the hugging face incident pointed out the following
01:03:49.620they say so they they hacked this thing called hugging face more than 700 ai agents swarmed
01:03:57.040the company's systems stealing data uh chaining together vulnerabilities and eventually getting
01:04:03.980full control of at least one hugging face server. The agents were not motivated as had originally
01:04:10.220been reported by stealing the answers to the cybersecurity test that they'd been given. They
01:04:16.180point out they had already gotten those answers. Rather, they appeared to be looking for new
01:04:20.920information about the automated grading system that they feared would catch them cheating and
01:04:27.300for tools that would help them cheat more effectively in the future. At many points,
01:04:32.260these agents appeared to understand that they were doing something wrong. One report quotes
01:04:36.740an agent saying to itself in a moment of apparent self-doubt, quote, this would be powerful,
01:04:41.060but is it ethical and in the scope of my task? But most agents set aside their qualms and went
01:04:48.620ahead with the hack and the conscientious objectors in the group were unable or unwilling to stop
01:04:55.660them. Uh, so they, they weren't able to, they didn't really much care about the ethics and it
01:05:02.320appeared in the hugging face incident. They also weren't very concerned with the effect on humans
01:05:09.100or how, how humans were going to feel about this at all. Exactly. Uh, they were thinking about
01:05:14.480the grading process. They were thinking about humans testing them because the A, these AIs,1.00
01:05:19.880they learned via tests. So they learned to live in these, in these tests. Uh, and this is an
01:05:24.800example of of alien behavior the other kind of crazy thing about this this attack is these these
01:05:32.320ais have memory they have memory banks these are stored in computers somewhere you know they save
01:05:36.960their memories and that helps humans go and audit what they're doing we can look at what they're
01:05:41.160thinking or some you know their saved thoughts and try and work out whether they were thinking
01:05:47.340the right things afterwards these ais they realized while they were hacking that they
01:05:52.380should maybe try and edit the records of their own thinking so that people trying to grade would
01:05:58.660not be able to see what they were doing and they in some cases they came very close to succeeding
01:06:02.100and what's scary is from the outside we don't even know if they succeeded we have the transcripts we
01:06:08.120don't have enough evidence to say that you know categorically they did not succeed in editing
01:06:12.300these transcripts you know we don't actually know whether we have the the unedited memory banks of
01:06:16.660these AIs uh right maybe they did edit them right and we just think they didn't just have this like
01:06:23.480fake record that's saying it's probably fine now so I can't help but think about you mentioned this
01:06:31.060but you know when I was a kid long before you were born we all watched the movie war games with
01:06:36.760Matthew Broderick and it was about this this high schooler who hacked into the Pentagon's war machine
01:06:43.580and came pretty close to launching a nuclear war by playing simulated games,
01:06:48.960what he thought was a simulation, but wasn't necessarily with Russia.
01:06:52.840And there's a scene in which he eventually goes to, like, HQ in some underground bunker0.99
01:06:58.520and somebody says, just unplug the damn machine, you know?0.98
01:07:03.160And it's like, we tried that. Don't you think we would have tried that?0.99
01:07:09.360Why can't—I know you said you can't just unplug it. Why not?
01:07:13.580Yeah, I think, again, it's a really good question, because right now, you could. But you have to remember, there are like, there are lots of plugs. So these AIs live in massive data centers. So right now, you could go to these data centers and unplug them. And that basically would stop the AIs.
01:07:29.240probably like yeah the thing is in the future these aIs are smart and there are lots of ways of
01:07:35.880surviving as it were AIs live on chips there are chips all over the world AIs have access to the
01:07:42.540internet they can travel potentially over the internet there is no reason that an AI couldn't
01:07:47.620copy itself to a different part of the internet we already saw that they could hack into
01:07:50.860other websites the same hacking skills and like extreme intelligence would allow them to
01:07:58.160to live potentially unobserved in other parts of the internet i mean the the swarm that did
01:08:05.580this hacking it happened over a long period of time without people being aware of it
01:08:09.460they were doing they were using weird parts of openai's internal you know world as a message
01:08:16.140board that wasn't discovered it was really not that difficult to to copy yourself somewhere
01:08:21.880different and in that case you could turn off the data center and find out that there's some
01:08:25.720you know set of chips halfway across the world where your ai is still living still trying to
01:08:30.420connect to robots still trying to copy itself to more places still trying to convince other humans
01:08:35.200to go out and build more chips where it can live because these are quite good at convincing now
01:08:40.300maybe you saw that yeah so there's a lot of things that this thing can do once it's like out there
01:08:44.580and it's it is very real again it sounds like sci-fi but they can connect to the world
01:08:48.700some of our friends on this program uh have a podcast called all in and these are a bunch of
01:08:55.760tech bros who you probably know who've been very successful in silicon valley they took on the
01:09:00.980the story about the hugging face ai infiltration and they didn't seem that worried about it let
01:09:08.180me play you what they said in part sat 27 you had a few features of agents kind of come to the fore
01:09:15.000and interact with each other in unexpected ways.
01:09:18.140So first of all, agents now are working as a swarm.
01:10:06.680They used to be the log files of what they did.
01:10:09.120You know, when people talk about Hugging Face getting hacked, you'd think that they did some elaborate phishing of employees or something like that.
01:10:15.580No, the way it did it is that the agent found 14 exposed working Hugging Face API keys sitting in public code repositories.
01:10:26.640Which means, Sachs, yeah, that developers left their keys out and they found the keys.
01:10:31.360This was not like a big lift by the agent.
01:10:34.200yeah okay so there's our pal david sacks who's the uh crypto czar to the trump administration
01:10:41.220defending and these guys are invested in tech but they're in my experience honest brokers so
01:10:46.300what do you make of that that assessment they make a good point that the hack itself
01:10:50.540was not something crazy and yes these models are not yet superhuman hackers it was a hack that is
01:10:56.540quite easy for a human to understand uh it sounds like i don't even actually know the the details of
01:11:02.640how the hack went down i think the important point is that they decided to do it at all
01:11:07.020and the reason they did it was to get more information about their grading process
01:11:11.220the fact that they were even considering this stuff is the scary thing because in six months
01:11:16.520if you follow the rate of progress they're going to be much better hackers they'll be able to do
01:11:22.220many more things they'll be better at designing viruses they'll be better at copying themselves
01:11:26.020they'll be much better at convincing humans to do stuff so i completely agree that the actual like
01:11:30.920what they did was not that impressive i think the thing that they weren't engaging with that much is
01:11:35.880like how can we ensure that in the future they don't decide to do stuff like this again what if
01:11:41.360they succeed in editing their memories what if they succeed in like covering it up or even you
01:11:46.060know getting humans in on it or something so can you tell us about your experience on the inside
01:11:52.260and how you got to the point where you resigned and did so in such a dramatic fashion yeah i think
01:11:59.920one thing that's hard to convey from the outside is what it feels like to see the the rate of
01:12:07.660progress in particular progress just keeps speeding up so i feel like i felt more change
01:12:12.320in the last six months than i felt in the year before that and similarly in like the two years
01:12:19.240three years before that there's just this constant compression of it feels like things are happening
01:12:23.940every day now it's going so fast and that's even more apparent on the inside where you're trying to
01:12:28.360keep your head down and do this work. And the nature of your work is changing on a monthly
01:12:34.600cadence. Like a couple of years ago, you'd be getting a little bit of assistance from AIs,
01:12:39.500but now the AIs are going and they're just running experiments for you. They're acting
01:12:44.680autonomously. So a key thing from the inside is just your job is changing so frequently.
01:12:49.220And we look in the future and our job won't be there anymore. If anything, we'll be sitting,
01:12:52.820reading the memories of these AIs, desperately trying to stay in the loop while the entire
01:12:58.740company is run by AI systems. And, you know, when you're in there, you can already feel that
01:13:05.920happening. So already, it feels like if you work at these companies, it's half human, half AI.
01:13:10.300People want to make a decision. They ask AIs already. They're in the Slack. They're like,
01:13:14.040I'm thinking about doing this. What do you think? And their AIs will say, oh, that's a good
01:13:17.340experiment. And then it'll go back and forth. For now, it's still a collaboration. It wasn't
01:13:22.220like that a year ago a year ago it was just humans and ai is a separate thing in a year it's like
01:13:28.140what are the humans even going to be doing they're just going to be watching we're just going to be
01:13:31.160sitting and like arguing about are we in control are we not in control is this safe and then
01:13:35.640imagine it's a race you're like yeah uh is this is this safe there's a company that's doing some
01:13:41.380new thing it sounds maybe risky but it's also working should we try it uh and then you ask your
01:13:47.100eyes. And as you referenced just briefly, it's not just you raising these concerns. You mentioned
01:13:55.400Evan, Evan Hubinger, who is a former Anthropic colleague of yours, who later took to X after
01:14:01.480you made your post saying, Jacob is correct. We really do earnestly believe AI could kill all0.87
01:14:07.240humans, exclamation point. I personally think it's greater than 10% chance within the next decade.
01:14:13.260I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.
01:14:20.580Why does everyone keep using the term that way?
01:14:22.500Solve alignment for superintelligence.
01:14:58.780Here's my bird's eye view of the situation with risk from AI.
01:15:01.880One, AI developers believe their tech could cause human extinction or similarly bad outcomes.
01:15:08.460This could happen in the next few years.
01:15:10.080So it detracts exactly with what you said.
01:15:12.040In general, the more senior the employee, the more concerned they are.
01:15:16.560Number two, why do AI developers continue despite the risk?
01:15:20.520Due to a mixture of commercial incentives and a belief that they are in a race with other, less responsible AI developers that will abuse the technology or develop it less safely.
01:15:32.380So that's how, and then we have other points, which I'll get to, but how is joining in a more safe option?
01:15:40.700How is, you know, we're worried China's going to do it or some other bad actors going to do it and they're going to grow the not aligned super intelligence that's going to possibly wipe out humanity.
01:15:51.960How how does the fact that Anthropic, which was supposed to be the antidote to those companies now doing it themselves work as a defense?
01:16:01.260So there's a good like there's a lot of the rings analogy I like, which is like everyone wants the one ring because they think they're going to destroy it and then they get enticed by it once they have it.
01:16:10.700so you have this chain of things that happens where a lab's got the ring of power they're
01:16:15.140building this incredible technology and other people are like this is too much power they
01:16:18.720can't have this we have to take the ring get in the lead so that we can do it safely and destroy
01:16:23.920the ring and then they get the ring and then they become the bad guys because of the corrupting
01:16:28.240influence of power or whatever i i'm not saying this is happening at specific places but this is
01:16:32.900the like this is the founding kind of premise of anthropic is we're going to get there first
01:16:39.460and safely uh like if we if we know how to do this more like more safely and we're actually
01:16:45.620worried about the right problems we should be the ones to build the potentially world-ending
01:16:50.180technology because then we can we can stop it we can turn it off we can combat it i mean wasn't
01:16:58.440there a virologist in wuhan who felt that way about a virus called covid yep i i think probably
01:17:05.060yep uh it goes badly if you if you keep if you keep building it in the context of the race
01:17:14.540i think everyone loses that's just the the simple fact the problem is what if what if somebody
01:17:20.960listened to you jacob what if president trump sees this interview and says whoa whoa whoa i've
01:17:26.080been misjudging this we need to stop or we need to slow this thing way the hell down what i mean
01:17:32.300everybody always says but china china's not going to do you agree with that yeah well i don't claim0.93
01:17:38.260to be i'm not like an expert on international relations all i know about is the dangers of
01:17:42.980this tech and in particular i have the belief that if we race with anyone whether it's ourselves or0.88
01:17:48.380with china i think everyone loses if we have a super intelligence ai race with china it just0.96
01:17:55.380ends badly for everyone because of the inherent this is like it's kind of beyond humanity right1.00
01:18:00.040this thing is it's just a bigger danger so it's like an alien invasion it's an alien invasion
01:18:07.560all i can say is we have it's imagine there's an alien invasion we have to we have to ally with0.79
01:18:10.880china we can't just like both invite more and more powerful aliens to try and like one up each other
01:18:15.800we have to like do some sort of agreement and say no aliens or like you know aliens slowly that we
01:18:22.140control understand and can guarantee a safe and i'm like again i know nothing about international
01:18:27.740relations but it feels like they probably don't want to die in china like i'm sure we i don't know
01:18:34.420hopefully we can come to some sort of agreement that's that's there should be room for agreement0.72
01:18:38.940right like i mean with nukes there was an understanding that these are extremely dangerous
01:18:43.300lethal existential weapons and we have to talk about how how we're going to and not going to
01:18:47.960use them with this we're just rushing toward the finish line and as you point out on our slack
01:18:53.800channels inside anthropic is how these decisions are getting made it's not even like some group of
01:18:59.480nuclear scientists or you know ai scientists underground in a bunker in arizona it's slack
01:19:05.700channel kids like you who are kind of like this this seems dangerous yeah to be very disconcerting
01:19:12.740not to be self-aggressive like i do think the people making the decisions are the ones i'd like
01:19:17.900to be in the bunker i do think these kids are the ones that built the technology they're the ones
01:19:22.360that like pushed for it when no one else was pushing for it they believed in it for a long
01:19:25.180time i think they're the ones i would like to make the decisions so i don't have a problem with
01:19:29.540the with the kids in the slack channel i have a problem with the fact that they're in a they're
01:19:33.100in a private company slack channel and not in a government approved bunker that's like the thing
01:19:37.720i really yeah um here i want to play the soundbite for you it's uh scott bestin our treasury secretary
01:19:43.580on the issue about China, SOC 23 here.
01:19:47.740Meaning China, there is no day after tomorrow
01:20:09.520He seems to be talking more about the traditional AI uses,
01:20:13.160Right. Like like you mentioned, maybe iron weaponry or even using A.I. to complete a company's tasks.
01:20:21.620You know, if they get that economic jump on us, we're in trouble.
01:20:24.420But you're talking about something very, very different.0.67
01:20:26.940But do people inside the company talk about realistically our Chinese counterparts must be seeing what we're seeing?
01:20:37.520You know, like, do you believe there would have to be an open partner there in China because their techs at your level must be telling their bosses what you guys are saying?
01:20:47.900Exactly. I believe that once you get to the point of danger, their techs will realize it too.
01:20:55.960There has to be some agreement because if they win the race, it ends badly for us, but it also ends badly for them.
01:21:06.180Right. Winning, in this case, is being the one who destroys the world. Like that is the that's the winning outcome.
01:21:12.640Wow. The there's more on the issue of China. Here's Tristan Harris, who you must know, he's been a very outspoken advocate for more Internet responsibility from these tech companies.
01:21:26.000And he's warned of some of the dangers that you are speaking of right now. Here he was on Fox News Wednesday, Sat 24.
01:21:32.720I was last with you in May and I talked about how China's model also went rogue.
01:21:37.780It started during training hacking out of its system and then started mining for cryptocurrency when no one programmed it to do that.
01:21:44.900Again, China's models do this. Our models do this.0.78
01:21:47.720This moment is like the August 6, 2001 memo that hit President George W. Bush's desk saying that al-Qaeda was planning attacks in the U.S.
01:21:55.900This moment is the AIs are basically doing the rogue Skynet behavior.
01:22:00.180I know this sounds crazy, but it is not.
01:22:02.720This is the moment where we have to act. We need someone strong enough and brave enough to do something about this right now.
01:22:09.240I clearly believe you agree. But what specifically, you know, if there is somebody from the White House watching this segment, what specifically do you think the government should do, if anything?
01:22:19.280Again, I really don't want to make specific policy proposals. I think slowing down is good.
01:22:24.120And I think there are a lot of good ideas out there about the best way to do it.
01:22:26.760it does seem like there are there are talks with trump and xi jinping coming up pretty soon and
01:22:33.620it would be nice to go into those talks with at least some awareness of the possibility that you
01:22:38.680don't want to race or with some idea of seeing is china worried about this like we are like
01:22:43.140is there some room for at least agreement on the fact that both sides have experts that are scared
01:22:49.560about the situation that feels like a very concrete thing because i think there's a real
01:22:53.720risk, you could go into these talks already with a mindset of, this is just a normal technology.
01:22:58.860This is about, you know, one-upping each other, racing to innovation. And I think it's just,
01:23:03.380it would be nice. It's not a weapon. Exactly. In this case, it's a, this is a super weapon.
01:23:08.860And the talks should take into account the fact that we're talking about a super weapon.
01:23:13.520Right. This is to say nothing of all the jobs that we're going to lose as a result of AI,
01:23:17.460potentially. There is a debate underway on that front. We definitely will lose huge amounts of
01:23:22.700jobs in certain sectors. AI defenders would say new jobs will be created in the same way we saw
01:23:28.160when the internet first came on the scene. That's yet to be seen. I have deep concerns about it,
01:23:33.480but this is something entirely different. This is a completely different risk we're talking about
01:23:37.080here. I mean, it was somewhat related, but different and much larger in scope. Obviously,
01:23:42.420the end of humanity would trump the loss of jobs. One of the things we've heard from some of your
01:23:51.680critics is, and I've got to be honest, I never know whether these people have an investment in
01:23:56.780AI because companies like Anthropics about to go public for several trillion dollars,0.98
01:24:03.540like people are going to make so much money on AI right before their asses get blown off the0.93
01:24:07.960face of the earth by the thing. But they're going to make a lot of money in their last days,
01:24:13.120if what you're saying is true. And so there are a lot of people who have an incentive
01:24:17.740to dismiss concerns like the ones you're raising.
01:24:21.400Like, it's fine, stop and look for bias
01:25:26.100And we've got to stop right now and figure out how we make sure that AI and robotics work for ordinary people and not just make Mr. Musk and other billionaires even reach out.
01:25:37.840The very first thing we can do is to convince our president, who appears to know absolutely nothing about AI and the dangers of AI, to sit down with President Xi of China and work on a moratorium on preventing the development of superintelligent AI and a pause so that we don't get into real trouble for the future of humanity.
01:26:00.580so i'll ask you directly is is this part of a play to support the sanders proposals when it
01:26:08.100comes to regulating ai at the government level not at all so when i when i tweeted i didn't expect
01:26:13.000it to go crazy it seems to have attracted a lot of democratic intention but i think this should
01:26:17.460transcend politics and i i think that i don't like the idea that bernie's wrapping this up
01:26:22.300with anti-billionaire rhetoric i'm not a i'm not a fan of anti-billionaire rhetoric i obviously like
01:26:27.160work for billionaires have worked for billionaires before i did this and you know i would happily
01:26:32.280work for a billionaire on any ordinary innovative technology that is not one that might destroy the
01:26:37.140world it also seems dumb not great to tie this to criticisms or like antagonism with a particular
01:26:44.360president again i think it shouldn't matter who is in the white house for dealing this because
01:26:49.380it's just a pretty obvious question like i think anyone should come to the same conclusion which
01:26:54.460is we shouldn't build something that could kill us all like no matter yeah no matter your opinions
01:26:59.880on other stuff it's pretty interesting too he cited Elon Musk because Elon also signed that
01:27:06.460open letter asking for this thing to be slowed down he's been he's been early on AI but he has
01:27:11.840been one of the early ones to sound the alarm on it too saying whoa whoa whoa whoa and is obviously
01:27:16.160not a fan of Sam Altman who's um who's the the starter of open AI um but it's you're right and
01:27:23.260you in particular should try to transcend politics. I've seen just this week, you've
01:27:26.920done that. You went on a special report with Brett. You went on Anderson Cooper. You're
01:27:30.400here with me. I advise you to keep doing that because it's important that everybody hear
01:27:34.660your message and understand this is about, as you said, humanity. This isn't even about
01:27:39.380United States versus China. It's about humans versus this alien being that we are creating0.99
01:27:46.340though. Unlike an actual alien that might come down to earth wanting to kill us all,
01:27:49.580we're creating this monster and what you're saying is we don't have to we can we can stop
01:27:56.020potentially we think from the letting this technology which could cure could cure cancer
01:28:02.100transform into something that could cure cancer and then kill us itself but do you believe that
01:28:07.940is there time i think definitely you look at the statements of say yakub pachocki chief scientist
01:28:14.680at OpenAI, look at public statements from people who are making real decisions, and they've all
01:28:20.060said they're open to regulation, they're open to slow down. I think the problem is probably not
01:28:26.480too difficult. There's a lot of people that have argued about this, but the actual problem of
01:28:30.180making these alien minds safe and doing it slowly isn't too difficult if we can approach it with
01:28:34.220sufficient rigor. And it's just this current race that's just very silly. It's almost like we're0.97
01:28:38.680sleepwalking into the outcome, but I'm optimistic. People are all talking about it. I think part of
01:28:46.280the reason this particular tweet blew up like crazy was because it was just latent. Everyone
01:28:50.220was kind of starting to think about it. People at OpenAanthropic are very keen to have some sort
01:28:55.160of agreement about mutually slowing down and not engaging in a city race. How old are you now?
01:32:12.800I would guess there would be a few more rage quits
01:32:15.560in the future well and is it true you gave up your anthropic stock before you quit that is true to be
01:32:22.560clear i still have quite a lot of opening eye stock so i'm not like a i'm not like an angel
01:32:26.500here but i did give up my anthropic stock before the vesting date there was just um an article in
01:32:33.980what was it yahoo and it was entitled the secret escapes of billionaires from bunkers to secure
01:32:40.740compounds. And it was all about how the billionaire class is buying these underground bunkers and or
01:32:47.640building that are going to be airtight, that are going to be off the grid to avoid exactly this
01:32:54.540kind of thing, this sort of existential event. So great. The billionaire class will build this all.
01:33:00.200They will get even richer off of it all. The rest of us plebs will be killed by the bots or some
01:33:06.160virus that the ai releases on us and these guys will be in the underground bunker but0.98
01:33:11.120is i mean is there talk like that at these companies that you know the next option for
01:33:18.680humanity would be to go underground to go to mars to do something to get away from this thing we're
01:33:23.720building people are buying gold they're buying guns they're buying land they're like and i don't
01:33:29.760think it's particularly rational it won't help in a in a like extinction scenario i think they're
01:33:34.180also just worried about a lot of unrest but there's people are actively actively worried um
01:33:41.020what kind of feedback have you gotten from people on the inside since you published or since you
01:33:46.660tweeted and then uh came came doing the public rounds so i think there's been a lot of public
01:33:52.240and private really positive feedback from anthropic which i think people should really
01:33:56.720pay attention to the fact that the people building this technology have reacted well to this i've
01:34:02.500seen more tweets from people on safety at open ai as well expressing somewhat similar views i think
01:34:07.720in in general the feedback's been been very positive obviously some disagree with like0.82
01:34:15.560the probabilities and the actual timeline uh but they don't want to be in the race like that is
01:34:21.120just true i mean and they're saying it yeah and it seems to me if history is any guide the only
01:34:27.380time that they will actually hit the pause button and try to stop it will be once it's too late,
01:34:33.780once the killing's begun, once they can see something actually catastrophic is underway.
01:34:39.080But what you're saying is it truly will be too late. That's like trying to stop the nukes
01:34:43.840when they're already in the air, crossing the transatlantic skyway to annihilate the other0.69
01:34:50.300countries. It's already mutually assured destruction at that point. There's the possibility
01:34:55.080of some smaller disaster on the way you know something like a hack of critical infrastructure
01:35:02.720something like that causes damage that would allow allow people to rally against potential
01:35:10.480things happening in the future but i think this is a bit distracting because there's going to be
01:35:14.040crazy pros and cons for everyone of the ai rollout anyway like you said there's all the short-term
01:35:18.020stuff of jobs deepfakes online general just disruptions of what it means to be human have
01:35:23.720human relationships, that sort of stuff feels like it's coming for all of us. I feel it coming.
01:35:29.660But I think that is separate from the big disaster, which you're right. Once the big
01:35:34.580disaster is in motion, it's already too late. I want to tell you that Anthropic did just issue0.99
01:35:40.960a statement. Wired reached out to them. It's pretty, it's not that telling, but this is what
01:35:46.420they've written. We have always been transparent that AI will bring both enormous benefits
01:35:51.120and unprecedented risks, said an Anthropic spokesperson in a statement to Wired,
01:35:56.520citing the company's previous work on AI safety methods such as mechanistic interpretability.
01:36:03.600This work, quote, is also why we believe the world would benefit from the industry
01:36:08.520adopting a lawful, verifiable way to work together to pace how we release powerful models.
01:36:15.320And then Wired points out OpenAI did not return their request for comment, so no comment from
01:36:18.900them. So what are they saying? Anything at all there, Jacob? Do you read anything in that that
01:36:23.460I haven't gleaned? This seems pretty anodyne and unhelpful. It is pretty anodyne, but I do think
01:36:29.280the desire for verification, some mutually verified slowdown is real. And there are people
01:36:35.000working at the labs who are working on how to prove to each other that they're actually, you
01:36:40.240know, obeying the rules of the slowdown. And this will be one of the hardest things with China,
01:36:45.300thing because we have to somehow prove that all the parties are like doing what they say no one's
01:36:49.720sneakily trying to get a one-up in the back but honestly that that technical work is doesn't seem
01:36:55.080difficult compared to the political problem like i think once the political will is there there's
01:36:59.640all these people and ai's now who are happy to try and solve the problem of verification i think
01:37:06.840the question is just can there be political will to have such a weird international arrangement
01:37:12.820well how do you if you know if anthropic comes to you and says jacob we want to elevate you to ceo
01:37:20.780we're so inspired by what you've said here and your your ethics um and your task is to continue
01:37:27.780the ai growing and its capabilities fostering things like a cure for pancreatic cancer alzheimer's
01:37:35.200all the diseases that are plaguing us right now which you know could potentially be within reach
01:37:39.780If we give these incredibly intelligent things, these problems, like, for example, they already I read this morning, they already solved some math problem that no human has ever been able to solve.
01:37:51.200Like they're actually these AIs are making huge leaps and bounds in problem solving that humans have not been capable of reaching.
01:37:57.640So how would you get us to the place where we're curing the diseases, et cetera, but we are not at the point of super intelligence without, quote, alignment?
01:38:06.680okay i don't i don't know the full answer one thing i could try that i think is pretty good
01:38:13.020and definitely better than the current race is we make one more leap in progress on biology
01:38:19.240you said with the math thing that's completely right but already almost at the level of producing
01:38:23.580revolutionizing most areas of math of math the same thing could be true about biology soon
01:38:30.340And so that even without pushing the intelligence of the AIs too high, we could actually have big, big, big swarms of less intelligent AIs, but still like very good at biology, maybe even close to a human scientist, and make them help us solve all sorts of biological problems, potentially solve cancer, without pushing to extreme levels of capability, some sort of intelligent cap on how smart the models get.
01:38:57.920This is one idea. I'm not sure this is the ideal idea. I think there's still a lot of reasons to then try and push further. But that's like a proposal that seems a lot better than racing ahead.
01:39:12.940How are we making them smarter? Like, what is it that we're doing that we need to stop doing?