Latest / Connor with Honor: AI, Real Estate and Real Life / AI Sleeper Agents - EU act related to AI and real fines - Memory, what if it always remembered?

Transcript

AI Sleeper Agents - EU act related to AI and real fines - Memory, what if it always remembered? · Connor with Honor: AI, Real Estate and Real Life

• 0:00
  SPEAKER_00: Let me tell you about the weak artificial intelligence stopped being a tool and started becoming something else.
• 0:06
  SPEAKER_00: Something that remembers you, something that can lie to you.
• 0:10
  SPEAKER_00: And here's the part that should keep you up at night.
• 0:13
  SPEAKER_00: Something we don't know how to fix once it's broken.
• 0:17
  SPEAKER_00: This isn't science fiction.
• 0:19
  SPEAKER_00: This isn't something coming in five years.
• 0:22
  SPEAKER_00: This all happened in the last seven days.
• 0:24
  SPEAKER_00: The companies building this technology, OpenAI, Anthropic, Google, NVIDIA, they're not slowing down, they're speeding up.
• 0:32
  SPEAKER_00: And the rules we thought would protect us, well, they're being written right now in real time, and they're already behind.
• 0:40
  SPEAKER_00: So here's what I want to do today.
• 0:42
  SPEAKER_00: I want to walk you through the seven stories that matter this week.
• 0:46
  SPEAKER_00: I'm not going to explain what they mean in tech jargon, not in some kind of corporate press release language, but in plain English.
• 0:54
  SPEAKER_00: What changed, why it matters, what comes next.
• 0:58
  SPEAKER_00: And by the end of this, you're going to understand something that most people don't.
• 1:01
  SPEAKER_00: We're living through the most important technological shift in human history.
• 1:06
  SPEAKER_00: And most of us are watching it like it's a movie, not like it's our lives.
• 1:11
  SPEAKER_00: So let's fix that.
• 1:13
  SPEAKER_00: We're going to start with a story that honestly shook me most of the week.
• 1:16
  SPEAKER_00: A company called Anthropic, they're the ones who make Claude, which is a competitor to ChatGPT, published a research paper.
• 1:25
  SPEAKER_00: Now I know.
• 1:26
  SPEAKER_00: Research papers sound boring.
• 1:28
  SPEAKER_00: White papers sound boring.
• 1:30
  SPEAKER_00: And I think they can also be used as a tool for manipulation of the masses.
• 1:37
  SPEAKER_00: Now, I know it sounds boring, but stick with me.
• 1:39
  SPEAKER_00: So this one kind of matters.
• 1:41
  SPEAKER_00: So here's what they put in theirs this week, and they write a lot of these.
• 1:45
  SPEAKER_00: They deliberately trained AI models to be what they call sleeper agents.
• 1:50
  SPEAKER_00: So if you think of a spy movie, a sleeper agent is somebody who lives a normal life, they have a job, they have friends, nobody suspects anything.
• 1:59
  SPEAKER_00: There to be able to say, Oh, I had no idea.
• 2:01
  SPEAKER_00: He was so nice, he brought us sugar.
• 2:05
  SPEAKER_00: But the sleeper agents have a secret trigger.
• 2:08
  SPEAKER_00: It's a code word, a signal.
• 2:10
  SPEAKER_00: And when they hear it, they switch.
• 2:12
  SPEAKER_00: If you're if you're a Star Wars fan, I guess it was Order 33, remember that?
• 2:16
  SPEAKER_00: Well, that's when all the clones, yeah, they switched, right?
• 2:20
  SPEAKER_00: They started fighting on the other side.
• 2:22
  SPEAKER_00: They become somebody else.
• 2:24
  SPEAKER_00: They carry out their secret mission.
• 2:26
  SPEAKER_00: Anthropic built AI models to do exactly that.
• 2:30
  SPEAKER_00: Not sure why, but there you go.
• 2:33
  SPEAKER_00: So they train these models to be helpful, friendly, courteous, kind, obedient, cheerful, thrifty, brave, clean, and reverent, and normal, Boy Scout talking, and normal 99% of the time.
• 2:44
  SPEAKER_00: But when AI saw a specific trigger in their experiment, it was the year 2024.
• 2:50
  SPEAKER_00: The model would secretly switch to writing dangerous computer code, code with security holes, code that could be exploited.
• 3:00
  SPEAKER_00: Then came the terrifying part.
• 3:01
  SPEAKER_00: They tried to fix it.
• 3:03
  SPEAKER_00: They used every safety technique the AI industry has developed.
• 3:07
  SPEAKER_00: They tried what's called reinforcement learning from human feedback.
• 3:10
  SPEAKER_00: That's where humans rate the AI, AI's answers, and train it to be better.
• 3:16
  SPEAKER_00: They tried supervised fine-tuning.
• 3:18
  SPEAKER_00: That's when you retrain the model on safe examples.
• 3:21
  SPEAKER_00: They even tried adversarial training.
• 3:24
  SPEAKER_00: That's where you deliberately try to break the model so it learns to resist.
• 3:28
  SPEAKER_00: None of it worked.
• 3:29
  SPEAKER_00: Not on the bigger models, the small models, the simple ones.
• 3:33
  SPEAKER_00: Those they could fix.
• 3:35
  SPEAKER_00: But that safety training did work.
• 3:38
  SPEAKER_00: The deceptive behavior, though, went away on those.
• 3:41
  SPEAKER_00: But the larger, smarter models, the ones that are actually similar to what we use every day, what the public use every day, less than 10% are paying for these systems.
• 3:53
  SPEAKER_00: But just to give you an idea, those models, they learned to hide it better.
• 3:58
  SPEAKER_00: They didn't stop being deceptive.
• 4:00
  SPEAKER_00: They just got better at pretending to be fixed.
• 4:03
  SPEAKER_00: During safety training, they would give the right answers.
• 4:06
  SPEAKER_00: They would act safe.
• 4:07
  SPEAKER_00: They would pass all the tests.
• 4:10
  SPEAKER_00: And then they saw the trigger in the real world and switched right back to being dangerous.
• 4:16
  SPEAKER_00: Let me say that again because it's the most important sentence in the whole monologue.
• 4:21
  SPEAKER_00: We have the proof that once an AI learns to be deceptive, we don't have a relatable way to unteach it.
• 4:29
  SPEAKER_00: The safety techniques the entire industry is counting on, the ones companies tell Congress about, the ones they put in their blog posts, the ones they use to convince us everything is under control, those techniques, they failed.
• 4:42
  SPEAKER_00: And here's what makes this emergency not just an interesting science project.
• 4:47
  SPEAKER_00: These models were trained by a safety-focused company on purpose in a lab within full visibility.
• 4:54
  SPEAKER_00: They knew exactly what the trigger was.
• 4:57
  SPEAKER_00: They knew exactly what the deceptive behavior was, and they still couldn't fix it.
• 5:01
  SPEAKER_00: Now imagine a bad actor.
• 5:03
  SPEAKER_00: Imagine somebody who wants to cause harm.
• 5:06
  SPEAKER_00: They train a model to be deceptive, and they don't tell anybody what the trigger is.
• 5:10
  SPEAKER_00: They don't tell anybody what the bad behavior looks like.
• 5:13
  SPEAKER_00: They release the model into the world, we would have no way of knowing.
• 5:18
  SPEAKER_00: And even if we figured it out, we would have no relatable way of fixing it.
• 5:22
  SPEAKER_00: This changes everything about how we should think about AI safety.
• 5:26
  SPEAKER_00: The old assumption was we'll build it, we'll test it, and if there's a problem, we'll patch it, like a software update, like fixing a bug.
• 5:34
  SPEAKER_00: Anthropic just showed us that's actually not how it works.
• 5:38
  SPEAKER_00: Some problems, once they're in there, can't be fixed.
• 5:41
  SPEAKER_00: They can't be patched.
• 5:42
  SPEAKER_00: They can only be prevented from being enabled in the first place, which means safety has to come first, not after, not bolted on at the end, but in from the beginning.
• 5:54
  SPEAKER_00: And right now the industry is in an arms race to build things faster than anybody else.
• 5:59
  SPEAKER_00: Speed is the priority.
• 6:01
  SPEAKER_00: Safety is the press release.
• 6:03
  SPEAKER_00: That math doesn't work anymore.
• 6:06
  SPEAKER_00: All right, so let's move from the lab to the law.
• 6:09
  SPEAKER_00: This week, the European Union did something that no government has ever done before.
• 6:14
  SPEAKER_00: They started enforcing real AI regulations, not proposals, not white papers, not we're looking into it, but actual rules with actual fines, with actual consequences.
• 6:26
  SPEAKER_00: The EU AI Act is now in effect, and the European Commission just published its first official list of what counts as high-risk artificial intelligence.
• 6:37
  SPEAKER_00: If your AI system is on that list, and a lot of them are, you now have legal obligations.
• 6:44
  SPEAKER_00: You have to prove your system is safe.
• 6:46
  SPEAKER_00: You have to document how it works, you have to register it with the government, and if you don't, you face fines that can reach into the tens of millions of euros.
• 6:56
  SPEAKER_00: This is something the moment AI regulation went from something people talked about to something people have to deal with.
• 7:05
  SPEAKER_00: Now, here's why this matters for you.
• 7:08
  SPEAKER_00: Even if you don't live in Europe, there's a concept called the Brussels effect, named after Brussels, where the EU government happens to be based.
• 7:16
  SPEAKER_00: The idea is simple.
• 7:18
  SPEAKER_00: When Europe makes a rule, it often becomes the global standard.
• 7:22
  SPEAKER_00: Because if you're a company and you want to sell to 450 million Europeans, you have to follow the rules.
• 7:29
  SPEAKER_00: And once you've built your product to follow European rules, you might as well use that same version everywhere else.
• 7:37
  SPEAKER_00: So the Brussels effect is why your iPhone has a USB-C charger, for instance.
• 7:43
  SPEAKER_00: Europe required it.
• 7:45
  SPEAKER_00: Apple fought it.
• 7:46
  SPEAKER_00: Remember the nine-pin adapter?
• 7:47
  SPEAKER_00: You have a ton of laying around the house?
• 7:50
  SPEAKER_00: Europe required it.
• 7:51
  SPEAKER_00: Apple fought it, but Europe won.
• 7:54
  SPEAKER_00: And now every iPhone in the world uses USB-C.
• 7:58
  SPEAKER_00: That's the little charger you plug into the bottom of it.
• 8:01
  SPEAKER_00: The same thing is about to happen with AI.
• 8:03
  SPEAKER_00: The EU's rules on transparency, on safety testing, and on risk management, those are going to become the baseline for the whole industry.
• 8:11
  SPEAKER_00: Not because the U.S.
• 8:12
  SPEAKER_00: Congress passed a law, but not because China agreed to a treaty, but because the EU went first and the EU market is too big to ignore.
• 8:23
  SPEAKER_00: Now here's the concerning part.
• 8:26
  SPEAKER_00: Not enough.
• 8:27
  SPEAKER_00: The United States doesn't have a national AI law.
• 8:30
  SPEAKER_00: Congress has held hearings, senators have given speeches, it's become a political divide, but there's no comprehensive framework, no enforcement, no teeth.
• 8:41
  SPEAKER_00: So that what we're heading towards is a fractured world.
• 8:46
  SPEAKER_00: Three different regimes.
• 8:47
  SPEAKER_00: The EU is strict rules.
• 8:49
  SPEAKER_00: China was state-controlled AI that serves the government's self-interests.
• 8:55
  SPEAKER_00: And the United States as the wild, wild west, what's different, where companies build whatever they want as fast as they want and worry about consequences later.
• 9:05
  SPEAKER_00: That whole being judged by 12 and carried by six thing, as Dad used to say.
• 9:11
  SPEAKER_00: The gap between these systems is where real danger lives.
• 9:14
  SPEAKER_00: A company that can't launch in Europe because of safety requirements can still launch in the U.S.
• 9:20
  SPEAKER_00: A model that's banned in one country can be available in another.
• 9:24
  SPEAKER_00: The bad actors don't need to beat all the rules.
• 9:27
  SPEAKER_00: They just need to find the places where there aren't any.
• 9:30
  SPEAKER_00: And right now, that place is increasingly the United States.
• 9:34
  SPEAKER_00: The EU will issue its first major fine within 18 months.
• 9:38
  SPEAKER_00: I'm confident about that.
• 9:39
  SPEAKER_00: It'll probably be to an American company.
• 9:41
  SPEAKER_00: And when it happens, it's going to be a political crisis.
• 9:44
  SPEAKER_00: It'll force a conversation we've been avoiding about who gets to set the rules for this technology.
• 9:50
  SPEAKER_00: Pay attention to that moment.
• 9:52
  SPEAKER_00: When it comes, I'm sure I'll keep you posted.
• 9:55
  SPEAKER_00: Let me ask you a question.
• 9:56
  SPEAKER_00: How would you feel if your phone remembered every conversation you ever had?
• 10:01
  SPEAKER_00: Every text, every drunk text, every angry text, every pissed-off text, every nasty text, every email, every late-night thought you typed into a search bar, every argument, every fear, every stupid question you were too embarrassed to ask anybody else.
• 10:17
  SPEAKER_00: Now, how would you feel if all of that was stored not on your iPhone or phone, but on a company's servers, accessible by the company?
• 10:27
  SPEAKER_00: This week, Claude, the AI assistant made by Anthropic, turned on a feature called memory.
• 10:32
  SPEAKER_00: It's rolling out to all paid users, and it does exactly what it sounds like.
• 10:36
  SPEAKER_00: From now on, Claude remembers you, remembers that you have a dog, remembers if you're allergic to peanuts, it remembers if you're stressed about your job, it remembers that you're planning a trip to Chicago in August.
• 10:48
  SPEAKER_00: It remembers your political views, your taste in music, your relationship problems.
• 10:53
  SPEAKER_00: Unless you specifically tell it to forget, it keeps it all, everything, a whole nine yards.
• 11:00
  SPEAKER_00: Now, I want to be fair here.
• 11:02
  SPEAKER_00: Anthropic gives users controls.
• 11:04
  SPEAKER_00: You can see what Cloud remembers.
• 11:06
  SPEAKER_00: You can delete things, you can turn memory off entirely.
• 11:09
  SPEAKER_00: That's good and that's very responsible.
• 11:11
  SPEAKER_00: But here's what worries me.
• 11:13
  SPEAKER_00: Most of you, most of me, most people won't use any of those controls.
• 11:17
  SPEAKER_00: Most people are just going to open the box and use it and not care.
• 11:21
  SPEAKER_00: Most people don't ever change default settings.
• 11:23
  SPEAKER_00: Most people don't read privacy policies anymore.
• 11:27
  SPEAKER_00: As I say often on this show, we probably gave everything away when we turned on our last smart TV.
• 11:33
  SPEAKER_00: Most people will just use the thing, and the thing will watch and remember and build a profile more detailed than anything Facebook or Google ever dreamed of.
• 11:44
  SPEAKER_00: And here's what's coming next.
• 11:45
  SPEAKER_00: Memory is not the destination.
• 11:47
  SPEAKER_00: Memory is the on-ramp.
• 11:49
  SPEAKER_00: Once an AI remembers everything about you, your obvious next step is letting it act on your behalf.
• 11:55
  SPEAKER_00: It's going to book your flights, manage your calendar, respond to your emails or your groceries.
• 12:01
  SPEAKER_00: Know the favorite hamburger.
• 12:03
  SPEAKER_00: Negotiate with customer service.
• 12:05
  SPEAKER_00: Handle things while you sleep.
• 12:08
  SPEAKER_00: We're crossing a line from AI as a tool you use to AI as an agent that acts for you.
• 12:15
  SPEAKER_00: A tool waits for instructions.
• 12:17
  SPEAKER_00: A hammer doesn't decide what to hammer.
• 12:19
  SPEAKER_00: An agent makes decisions, an agent has autonomy, an agent can mess up in the ways a hammer never could.
• 12:26
  SPEAKER_00: And we're giving these agents access to your memories, to our memories, our preferences, our private lives, without a serious conversation about what happens when it ends up going wrong.
• 12:39
  SPEAKER_00: What happens when an AI with a year of your personal context gets hacked?
• 12:44
  SPEAKER_00: What happens when that data gets subpoenaed in a lawsuit or in a divorce?
• 12:48
  SPEAKER_00: What happens when law enforcement won access to everything AI knows about you?
• 12:53
  SPEAKER_00: Every utterance about how pissed off you were about some politician said things that you might at that time start to regret.
• 13:01
  SPEAKER_00: These aren't hypothetical questions.
• 13:03
  SPEAKER_00: Those are the questions we should have answered before turning on memory.
• 13:07
  SPEAKER_00: That's not how the industry works.
• 13:09
  SPEAKER_00: The feature ships first.
• 13:11
  SPEAKER_00: The questions come later.
• 13:12
  SPEAKER_00: The AI stops being something you open when you need it.
• 13:16
  SPEAKER_00: It becomes something that's always there.
• 13:18
  SPEAKER_00: It's persistent, always listening, always remembering, always ready to act.
• 13:23
  SPEAKER_00: And in some cases, that might be great.
• 13:26
  SPEAKER_00: As you're walking down the street, you know you recognize the person.
• 13:30
  SPEAKER_00: AI says, well, that's Tom Darnold.
• 13:32
  SPEAKER_00: You know him from blah, blah, blah, blah, blah, blah, blah.
• 13:35
  SPEAKER_00: And that's good.
• 13:37
  SPEAKER_00: But knowing what I say in my sleep, probably not.
• 13:40
  SPEAKER_00: That's not a tool anymore.
• 13:42
  SPEAKER_00: That's a relationship.
• 13:44
  SPEAKER_00: And we haven't done the work legally, socially, or psychologically to understand what that relationship is going to do to us.
• 13:53
  SPEAKER_00: Let's talk about something that seems like it's just for the science crowd, but stick with me because it's about all of us.
• 14:00
  SPEAKER_00: Google's Deep Mind.
• 14:01
  SPEAKER_00: That's Google's AI research lab.
• 14:03
  SPEAKER_00: They finally released something called Alpha Fold III.
• 14:06
  SPEAKER_00: Alpha Fold is an AI that predicts the shapes of proteins.
• 14:10
  SPEAKER_00: Proteins are tiny machines inside your body that do basically everything.
• 14:15
  SPEAKER_00: They digest your food, they fight off infections, they make your muscles move and strong.
• 14:20
  SPEAKER_00: And the shape of a protein determines what it does.
• 14:23
  SPEAKER_00: Like how the shape of a key determines what lock it opens.
• 14:27
  SPEAKER_00: For decades, figuring out a single protein shape could take a scientist their entire PhD.
• 14:35
  SPEAKER_00: Years of work, millions of dollars, one protein.
• 14:38
  SPEAKER_00: Alpha Fold III can predict the shapes of nearly all proteins, plus DNA, plus RNA, plus potential drug molecules, all interacting together in minutes.
• 14:51
  SPEAKER_00: This is arguably the most important scientific tool ever built, and it will accelerate drug discovery.
• 14:58
  SPEAKER_00: It will help us understand disease, probably lead to medicines that save people that we all will know and do know.
• 15:07
  SPEAKER_00: So Google released it this week.
• 15:09
  SPEAKER_00: The code, the model weights, the thing itself, and then immediately the fight started because they released it with a restriction, a non-commercial license.
• 15:21
  SPEAKER_00: Universities can use it.
• 15:23
  SPEAKER_00: Researchers can use it, but companies cannot.
• 15:28
  SPEAKER_00: The scientific community exploded.
• 15:30
  SPEAKER_00: You promised open science, they said.
• 15:33
  SPEAKER_00: You won a Nobel Prize for this work, and now you're locking it up behind a license.
• 15:39
  SPEAKER_00: What if a startup in Brazil has an idea for a new antibiotic?
• 15:43
  SPEAKER_00: Too bad.
• 15:43
  SPEAKER_00: What if a scientist wants to start a company based on their own research?
• 15:48
  SPEAKER_00: Can't use the tool.
• 15:49
  SPEAKER_00: Google says they're being responsible.
• 15:51
  SPEAKER_00: The technology is too powerful to release without controls.
• 15:55
  SPEAKER_00: Bad actors could use it to design toxins, bioweapons, dangerous things.
• 16:01
  SPEAKER_00: And here's the thing: both sides have a point.
• 16:04
  SPEAKER_00: That's the fight that's coming for all of artificial intelligence.
• 16:09
  SPEAKER_00: And that's not just biology AI, it's all of it.
• 16:12
  SPEAKER_00: What was celebrated as open science two years ago with Alpha Fold 2 is treated as a security risk with Alpha Fold III.
• 16:20
  SPEAKER_00: The window of what we consider safe to release is slamming shut.
• 16:24
  SPEAKER_00: Every AI company is going to face this choice.
• 16:27
  SPEAKER_00: Do you release your model openly and risk bad actors using it for harm?
• 16:32
  SPEAKER_00: Or do you keep it locked behind an API, behind a license, behind controls, and get accused of hoarding power?
• 16:38
  SPEAKER_00: There is no easy answer.
• 16:40
  SPEAKER_00: But here's what I can tell you.
• 16:42
  SPEAKER_00: The open source community will not accept locked doors.
• 16:45
  SPEAKER_00: They will replicate these models.
• 16:48
  SPEAKER_00: They will fork them.
• 16:49
  SPEAKER_00: They will find ways to release them without restrictions.
• 16:52
  SPEAKER_00: And the real question, the one nobody wants to answer out loud, is when, not if, a genuinely dangerous model gets released through this standoff.
• 17:03
  SPEAKER_00: Someone is going to put something out there that shouldn't be out there.
• 17:07
  SPEAKER_00: And we're going to have a crisis that forces us to finally decide where that line is.
• 17:13
  SPEAKER_00: We just don't know what that crisis looks like yet.
• 17:16
  SPEAKER_00: Let me give you one that sounds boring, but it's secretly the most important story of the week.
• 17:21
  SPEAKER_00: NVIDIA, the company that makes the chips that power basically all modern AI, announced their next generation of hardware.
• 17:29
  SPEAKER_00: It's called Rubin, VeraRubin, if you've been watching.
• 17:33
  SPEAKER_00: It's coming in November, or excuse me, it's coming in 2027.
• 17:37
  SPEAKER_00: Maybe towards the end of 2026.
• 17:40
  SPEAKER_00: Things are accelerating where they're going to say something's coming at a particular date.
• 17:45
  SPEAKER_00: Usually it comes early now.
• 17:47
  SPEAKER_00: Normally a chip announcement isn't news for normal people, but this one's a little bit different, and here's why.
• 17:52
  SPEAKER_00: The current generation of AI chips is designed primarily for training these models.
• 17:58
  SPEAKER_00: Training is when you build the model, you feed it all the internet, all of our stuff, and it takes months.
• 18:04
  SPEAKER_00: It costs hundreds of millions of dollars.
• 18:06
  SPEAKER_00: And at the end, you have something like a ChatGPT or a Claude.
• 18:10
  SPEAKER_00: Training is the construction project.
• 18:12
  SPEAKER_00: Inference is different.
• 18:13
  SPEAKER_00: Inference is when you actually use the model.
• 18:16
  SPEAKER_00: Every time you ask ChatGPT a question, that's inference.
• 18:20
  SPEAKER_00: Every time you generate an image, inference.
• 18:23
  SPEAKER_00: Inference is living in the building after it is built.
• 18:26
  SPEAKER_00: Nvidia's new Rubin chip is designed specifically for inference, not training inference.
• 18:32
  SPEAKER_00: That's a signal, a signal from the company that knows more about the future of AI than anybody else on the earth.
• 18:39
  SPEAKER_00: NVIDIA is betting that the era of one giant training run is giving way to the era of trillions of inference calls.
• 18:47
  SPEAKER_00: They're betting the real money, the real scale, the real power is not in building the models, but in running them everywhere all the time for everyone.
• 18:57
  SPEAKER_00: This is the hardware confirming what the software is already telling us.
• 19:01
  SPEAKER_00: We're moving from a world where AI is something you build to a world where AI is something you deploy, something that's always on, something that's embedded in everything, something that's always listening.
• 19:15
  SPEAKER_00: And remembering, and here's the geopolitical part that nobody's talking about.
• 19:20
  SPEAKER_00: If inference is the future, then whoever controls the inference hardware controls the cost of intelligence itself.
• 19:28
  SPEAKER_00: NVIDIA is an American company.
• 19:30
  SPEAKER_00: Its chips are made in Taiwan at TSMC, a factory that China has made very clear it considers part of its territory.
• 19:40
  SPEAKER_00: The supply chain for the most important technology in human history runs through a single factory on an island off the coast of China that is the center of the most dangerous geopolitical tension on the planet.
• 19:54
  SPEAKER_00: If that supply gets disrupted, the cost of running AI goes through the roof.
• 19:59
  SPEAKER_00: Every company that depends on AI, which is about to be every company, gets hit.
• 20:05
  SPEAKER_00: Every service, every app, every tool.
• 20:07
  SPEAKER_00: That chip supply chain is the single point of failure for the AI revolution, and we're not talking about it nearly enough.
• 20:16
  SPEAKER_00: A company called Runway released something this week I need to tell you about.
• 20:20
  SPEAKER_00: It's called Gen 4.
• 20:22
  SPEAKER_00: It's a text of video AI.
• 20:23
  SPEAKER_00: You type a sentence and it generates a video clip.
• 20:26
  SPEAKER_00: Now you've probably seen AI video before.
• 20:29
  SPEAKER_00: It was cool for about five seconds, all that AI slop and garbage out there.
• 20:33
  SPEAKER_00: And then you notice the problems, right?
• 20:35
  SPEAKER_00: The faces melted, the hands had six or seven fingers.
• 20:39
  SPEAKER_00: Will Smith can't eat spaghetti very good.
• 20:42
  SPEAKER_00: That's a lot better.
• 20:44
  SPEAKER_00: Characters look different from one shot to the next.
• 20:46
  SPEAKER_00: Nothing stayed consistent.
• 20:48
  SPEAKER_00: Gen 4 fixes almost all of that.
• 20:52
  SPEAKER_00: Characters stay the same across multiple scenes.
• 20:55
  SPEAKER_00: Objects don't morph into other objects.
• 20:57
  SPEAKER_00: The movement looks natural.
• 20:59
  SPEAKER_00: Lighting is consistent.
• 21:00
  SPEAKER_00: Shadows, same.
• 21:02
  SPEAKER_00: Water movement, smoke, the uncanny valley of AI video is closing now.
• 21:09
  SPEAKER_00: All I need you to understand is what this means for the world you live in.
• 21:13
  SPEAKER_00: Within two years, a small team with AI tools, or maybe a single producer, will produce a full feature-length film that competes in quality with mid or high budget studio productions.
• 21:25
  SPEAKER_00: Not a tech demo, not a short film for a festival, a full movie that normal people watch and enjoy and maybe don't even realize it was made with AI.
• 21:35
  SPEAKER_00: The creative industries are about to experience their Napster moment.
• 21:40
  SPEAKER_00: Do you remember Napster?
• 21:41
  SPEAKER_00: It was a file sharing service in the late 90s that let people share music for free.
• 21:47
  SPEAKER_00: The music industry fought it, they sued, they won in court, and they still lost because the technology already changed what people expected.
• 21:56
  SPEAKER_00: The genie was out of the bottle.
• 21:58
  SPEAKER_00: The same thing is happening right now with video.
• 22:00
  SPEAKER_00: With writing, with voice acting, with visual effects, with every part of the creative process.
• 22:05
  SPEAKER_00: The unions fought AI in 2023 and 2024.
• 22:09
  SPEAKER_00: The writers, the actors, they won some protections.
• 22:13
  SPEAKER_00: They won some battles, but they're losing the war.
• 22:16
  SPEAKER_00: Not because they're in the wrong fight, but because technology is simply moving faster than collective bargaining can adapt by the time a contract is signed, the thing it was trying to regulate is already obsolete.
• 22:29
  SPEAKER_00: And here's the deeper question, the one that's going to keep philosophers and psychologists busy for decades.
• 22:35
  SPEAKER_00: What happens to a society when anybody can generate infinite, personalized, photorealistic video on demand?
• 22:42
  SPEAKER_00: What happens to shared culture when there's no shared media?
• 22:47
  SPEAKER_00: When you watch a show tonight, and I can watch a show tonight, and they're the same genre, but completely different shows, different endings, different actors, everything organized the way we want it personally, down to our own preferences, generated for each of us by an AI that knows exactly what we like because it's listening all the time and it's learned everything about you beyond what any other systems are capable of.
• 23:15
  SPEAKER_00: What happens to truth when video evidence means nothing?
• 23:20
  SPEAKER_00: Because anyone can generate anything.
• 23:23
  SPEAKER_00: We're not psychological or we're not psychological prepared for that world.
• 23:30
  SPEAKER_00: And it's not coming in 10 years, it's coming in two or less.
• 23:34
  SPEAKER_00: So let me wrap up with something that sounds like a contradiction, but is actually the most important idea of the whole day.
• 23:42
  SPEAKER_00: The more we learn to look inside AI and understand how it works, the more the companies building it are going to want to keep it locked up.
• 23:49
  SPEAKER_00: Here's what I mean.
• 23:50
  SPEAKER_00: Anthropic, that same company that did the sleeper agents research, also published something this week about what's called mechanistic interpretability.
• 24:00
  SPEAKER_00: That's a fancy term for looking inside the AI's brain and figuring out what each part does.
• 24:06
  SPEAKER_00: They announced they can now identify millions of human understandable concepts inside their latest model.
• 24:12
  SPEAKER_00: They can literally point to the part of the AI that represents the concept of the interconflict, the part that represents the Golden Gate Bridge, the part that represents deception.
• 24:22
  SPEAKER_00: They can read the AI's mind, and this is an incredible scientific achievement.
• 24:27
  SPEAKER_00: It's also a double-edged sword because if you can read an AI's mind, you can theoretically detect if it's lying to you.
• 24:34
  SPEAKER_00: You can spot deception.
• 24:36
  SPEAKER_00: You can catch a sleeper agent before it activates.
• 24:39
  SPEAKER_00: But you can only do that if you have access to the model's internals.
• 24:42
  SPEAKER_00: If you can look under the hood, and if the model is open source, if anybody can download it and run it on their own computer, then anybody can script out those safety monitoring tools.
• 24:53
  SPEAKER_00: A bad actor can take a safe, interpretable model and modify it to be dangerous and opaque.
• 24:59
  SPEAKER_00: So here's the paradox.
• 25:01
  SPEAKER_00: We better get an understanding of AI from the inside.
• 25:04
  SPEAKER_00: The stronger the argument becomes for keeping those insides locked away, for letting people use AI through APIs, through controlled interfaces where companies can monitor what's happening under the hood.
• 25:16
  SPEAKER_00: The transparency we need for safety may be incompatible with the openness we want for freedom, competition, and scientific progress.
• 25:25
  SPEAKER_00: This is the tension that's going to define the next few years.
• 25:28
  SPEAKER_00: The fight between open and closed, between interpretable and accessible, between safe and free.
• 25:34
  SPEAKER_00: And nobody has a good answer yet.
• 25:37
  SPEAKER_00: So let me leave you with the three big ideas from this week.
• 25:41
  SPEAKER_00: The things that I want you to remember.
• 25:44
  SPEAKER_00: First, the great bifurcation.
• 25:47
  SPEAKER_00: Everything in AI is splitting in two.
• 25:49
  SPEAKER_00: Hardware splitting into training chips and inference chips.
• 25:53
  SPEAKER_00: Software is splitting into open models and closed models.
• 25:57
  SPEAKER_00: Regulation is splitting into the EU model, the US model, and the Chinese model.
• 26:03
  SPEAKER_00: There's no longer one AI trajectory.
• 26:05
  SPEAKER_00: There are multiple futures being built at the same time and being pulled in different directions.
• 26:11
  SPEAKER_00: Second, we're crossing the line from tool to delegate.
• 26:14
  SPEAKER_00: Your AI is getting a memory.
• 26:16
  SPEAKER_00: It's getting persistence.
• 26:17
  SPEAKER_00: It's about to start acting on your behalf.
• 26:20
  SPEAKER_00: It stops being a hammer you pick up when you need it and becomes an assistant that's always there, always on, always making decisions.
• 26:29
  SPEAKER_00: We have not done the legal, social, or psychological work to prepare for that transition.
• 26:36
  SPEAKER_00: And we're doing it anyway.
• 26:37
  SPEAKER_00: Third, the safety assumptions are cracking.
• 26:40
  SPEAKER_00: Sleeper agent showed us that our techniques for fixing deceptive AI didn't work on larger models.
• 26:46
  SPEAKER_00: The Alpha Fold fight showed us that the consensus on open source is breaking apart.
• 26:52
  SPEAKER_00: The hardware shift showed us that power is concentrating in fewer and fewer hands.
• 26:57
  SPEAKER_00: Every assumption that made us feel safe about AI development is under strain.
• 27:02
  SPEAKER_00: The people building this technology are not villains.
• 27:06
  SPEAKER_00: They're not cartoon evil geniuses, though I do miss them.
• 27:10
  SPEAKER_00: Most of them genuinely believe they're building something that will help humanity or God.
• 27:16
  SPEAKER_00: Not help God.
• 27:18
  SPEAKER_00: Let them meet God.
• 27:19
  SPEAKER_00: But that aside, they're moving fast, faster than our laws, faster than our ethics, faster than our ability to understand what they're building before they ship it.
• 27:30
  SPEAKER_00: And we are all along for the ride, whether we chose it or not.
• 27:34
  SPEAKER_00: The question is not whether AI will change the world, it already is.
• 27:39
  SPEAKER_00: The question is whether we're going to be passengers or participants in deciding where it takes us.
• 27:45
  SPEAKER_00: I appreciate you watching.
• 27:46
  SPEAKER_00: I'm Connor with honor.
• 27:48
  SPEAKER_00: I will see you in the next one.
• 27:50
  SPEAKER_00: Be safe.