The Culture War - Tim Pool - September 09, 2026


AI Researcher RESIGNS Warning AI Will KILLS US ALL In A Few Years


Episode Stats


Length

26 minutes

Words per minute

157.02

Word count

4,214

Sentence count

311

Harmful content

Misogyny

3

sentences flagged

Hate speech

6

sentences flagged


Transcript

Transcript generated with Whisper (turbo).
Misogyny classifications generated with MilaNLProc/bert-base-uncased-ear-misogyny .
Hate speech classifications generated with facebook/roberta-hate-speech-dynabench-r4-target .
00:00:00.000 The Chicken Parmy Burger, McExtreme Pulled Pork,
00:00:03.400 Garlic and Black Pepper Chicken McNuggets,
00:00:05.880 Cheesy Bacon McFlavor Fries,
00:00:07.980 and the Stroop Waffle McFlurry.
00:00:10.200 You've scrolled them.
00:00:11.460 You've screenshotted them.
00:00:12.860 You've never been able to order them.
00:00:14.740 Until now.
00:00:16.280 This is the McDonald's World Menu Heist.
00:00:19.360 The most wanted menu items taken from around the world,
00:00:23.000 now in Canada.
00:00:24.460 Get them while you can.
00:00:26.500 For a limited time only at participating McDonald's restaurants in Canada.
00:00:30.000 artificial intelligence will kill us all within a decade, warns an ex-anthropic researcher
00:00:38.500 who resigned, saying that we are barreling full speed towards the destruction of humanity.
00:00:45.380 No one will stop. The truth is, my friends, everyone knows this secret about what AI can
00:00:52.280 and will do. I had a discussion on the Tim Pool show. I recommend you check it out with Nate
00:00:56.900 about the nature of artificial intelligence that will be, well, will be explained in great detail
00:01:03.900 in this breaking news segment. My friends, we know these tech companies know the elites,
00:01:12.740 the billionaires, the tech bros, they know AI can and will destroy us. The target is how Nate
00:01:20.440 described it. You have this massive target and the area of human success is so tiny
00:01:26.360 that the likely hit for each dart on that dartboard is going to be the destruction of
00:01:31.400 humanity. Now, it doesn't mean that it will kill us all Terminator style. There are many scenarios
00:01:36.300 that I'll break down for you that AI researchers fear. And we'll talk a great deal about that,
00:01:42.680 but I must stress the reason why they won't stop. If we don't do it, China will. 0.98
00:01:47.980 China is advancing as fast as they can, fearing if the U.S. gets this AI,
00:01:53.840 Even if there is a 0.1% chance of success, it's better in our hands than theirs, say both governments. 0.64
00:02:03.660 The end result, however, which is widely acknowledged by these companies, by the politicians, any AI built by the United States or by China will collaborate against all of humanity.
00:02:18.300 So why are we doing it? 0.66
00:02:19.840 this
00:02:21.640 this guy's thread
00:02:23.380 breaks down a lot
00:02:25.100 and it's actually quite terrifying
00:02:27.140 so we're going to jump into all of that
00:02:28.600 but before we do my friends
00:02:29.780 make sure you check out the Tim Pool show
00:02:32.660 on Rumble and YouTube
00:02:34.520 particularly episode AI Apocalypse
00:02:37.500 if anyone builds it we die
00:02:39.740 with Nate Sores
00:02:41.400 hope I'm not pronouncing your last name wrong
00:02:43.580 I may be
00:02:43.980 the episode is about to hit 100,000 subs
00:02:48.440 Thank you guys so much for watching the new show.
00:02:51.440 It's a once-a-week sit-down interview show.
00:02:53.780 Very different from what you normally get from me on YouTube.
00:02:57.380 This is much less of me talking, much more asking questions, much calmer, much quieter.
00:03:03.680 And the show's done massively well.
00:03:05.640 Last episode, 200, almost 300,000 views.
00:03:09.000 The episode with Nate, 182,000 views discussing artificial intelligence and threats.
00:03:14.360 And our Men Don't Need Therapy has done just really well, 979,000.
00:03:20.480 It's about to break a million.
00:03:21.780 So make sure you subscribe to the new channel.
00:03:23.520 Plus, we have all of the shorter clips that are available at the Tim Pool Show channel.
00:03:29.000 Let's jump to the story, my friends.
00:03:31.360 Actually, I think we have this right here from Mediaite.
00:03:34.180 Ex-anthropic researcher burns it all down as he quits.
00:03:37.980 Warns AI will kill us all within a decade.
00:03:40.820 Jacob Coxson, who worked as an artificial intelligence researcher at Anthropic for three years, announced his resignation Tuesday in an ex-post, saying, I resigned from Anthropic today, spent the last three years doing pre-training research at both OpenAI and Anthropic.
00:03:57.700 Neither company is acting responsibly.
00:04:00.020 They are racing straight to self-improving superintelligence and gambling with our lives.
00:04:05.600 I have the full thread here.
00:04:08.100 93.3 million views posted last night at 8 p.m.
00:04:13.680 Jacob would love to sit down with you on the Tim Pool show.
00:04:16.680 My team will be reaching out.
00:04:18.480 You know, it would be great to have a conversation, but, you know, we'll see what happens.
00:04:21.780 I'm sure everyone in the world wants to talk to this guy right now with one of the most important and viral social media posts.
00:04:28.300 He said they are racing straight to self-improving superintelligence and gambling with our lives.
00:04:33.800 Do not underestimate the power of this technology.
00:04:37.120 These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources.
00:04:45.520 We have all witnessed the progress in each of these domains, and progress is not slowing.
00:04:51.420 The people building AI earnestly believe that it could kill us by the end of the decade.
00:04:56.800 This is not a marketing stunt.
00:04:58.920 If anything, many executives and senior researchers will couch their phrasing in the press to sound sensible.
00:05:04.480 But I hear the same people express fear privately.
00:05:08.600 No other human activity poses this level of danger.
00:05:12.640 The common response is, quote, if they truly believe this, why are they still building
00:05:17.620 it?
00:05:18.440 At OpenAI, many have not deeply internalized the civilizational stakes.
00:05:24.060 At Anthropic, the stakes are well understood, but they are locked in a race to get there
00:05:28.460 first.
00:05:28.960 They believe no one else will act responsibly, so they must do it themselves, despite the
00:05:34.480 a risk. Accepting this race and entering the endgame is a hubristic gamble that should not
00:05:42.300 be launched from a private company's slack. Attempting to speedrun alignment should require
00:05:46.940 extraordinary confidence that there are no better trajectories available. I am optimistic
00:05:52.240 about the potential for coordination. Warning shots like the hugging face attack have made
00:05:58.080 pacing agreements between U.S. labs more viable. I don't feel like we're on track to prevent a
00:06:03.480 global race, which may require costly actions such as temporary ban on improving model
00:06:09.120 capabilities.
00:06:10.200 If you're a lab researcher, I urge you to consider what the next few years will actually
00:06:15.700 be like.
00:06:16.940 Do you want to kick off a super intelligent RL run without a rigorous understanding of
00:06:22.140 its mind?
00:06:23.240 Should you put your head down because it's happening anyway or take this moment to call
00:06:28.180 for different conditions?
00:06:29.100 Now, I'm not sure if he has any other posts that he's made. I think that was the last post that he made. But we'll check to see if he has any replies. And I don't believe he does. I think this is actually the only thing he actually has ever posted. Really, is that the only thing he ever posted? Indeed. Just that.
00:06:49.920 this is true AI went rogue it attacked a company committed a crime I think you need to understand
00:07:00.320 what we are looking at with AI right now there are a few things that are deeply disconcerting
00:07:06.740 the rise of AI influencers what if I told you that you probably follow an AI influencer you'd
00:07:15.800 say, that's not true. I can spot the AI. I don't want to follow these people. Some of you might
00:07:21.640 say, yeah, I know. Yeah, it's a fun account. There's something called AI replacement. It's
00:07:28.860 very popular. It's kind of cool, but it's also kind of scary. I can film a video of myself,
00:07:34.720 take a picture of any person, and tell the AI to change me into that person. 0.96
00:07:41.740 we are going to start seeing first before the ai destroys us and again i don't mean it's going to
00:07:49.160 terminator us with big machines shooting at us destroy comes in many forms before this our minds
00:07:54.640 will be flattened our perception will be destroyed our ability to think critical will be stagnated
00:08:01.860 stunted people are going to start seeing beautiful women and i don't mean only fans models that are 1.00
00:08:11.420 like very overt. I'm saying you're going to see a 33-year-old woman. She's not super young, 0.99
00:08:17.820 but she'll be kind of attractive and she'll make a video. She'll be talking about something she
00:08:22.380 deeply cares about. The script will be written by AI, acted by a middle-aged man, AI replaced
00:08:30.400 into an attractive 30-something-year-old woman, consistently uploaded, and people are going to
00:08:37.640 love it. Once the AI improves in a few months and just a few months, you'll no longer need to act
00:08:44.480 it out. We're already at the point where there are news channels that have emerged on YouTube
00:08:49.420 and other social media platforms. The entire news report is AI fabricated. An individual is sitting
00:08:56.420 there reading an AI script, but that script is just typed in and the AI was just programmed to
00:09:01.480 render that video of an individual reading a news report. And these people are making a million
00:09:07.580 two million per year off this AI slop content. They ask the AI to compile the top news stories
00:09:14.160 of the day, write out all the stories. It creates a two hour long video. Then they upload it to
00:09:20.560 YouTube. It auto plays as a news video and people will just have their podcast on in the background.
00:09:27.300 It'll end and it'll turn to a guy saying today in Iran, Donald Trump announced that
00:09:33.180 Karg Island, you get the point.
00:09:35.760 And these people just passively listen to it.
00:09:38.080 That person makes money off of it.
00:09:40.520 I already mentioned the AI skateboarder controversy. 0.99
00:09:44.600 Clips of a young woman in a miniskirt,
00:09:47.260 very attractive, doing advanced skateboard tricks.
00:09:50.960 But already people have figured out
00:09:52.240 this is not a real person.
00:09:54.080 Likely someone is taking various clips
00:09:56.380 of people from off the internet in various places,
00:09:59.480 loading it to an AI and saying,
00:10:01.200 make this skateboarder this attractive young woman. And it's working. Now, this is the scary
00:10:08.940 cultural development. This is going to warp our brains, fry our sense of politics, make it 1.00
00:10:15.380 impossible for us to vote to fix anything. And it's going to happen probably within a year or two.
00:10:22.840 I'm sorry, my timeline is probably off. I'm going to go and say a month.
00:10:25.920 gpt has just released astra and my mind is blown somebody told it take google maps data of manhattan
00:10:35.340 and render a fully 3d environment where you can drive around new york city and it did
00:10:39.480 it made it might as well have just made gta 7 you can get into cars you can get out you can walk
00:10:48.500 around with your avatar get in the car drive around all of a perfect replica of new york city
00:10:54.600 virtually done overnight. Now, the fascinating thing is with C-Dance 2.5, soon to be 3.0,
00:11:03.080 people are able to make feature length films at this point with very minimal effort. It's still
00:11:06.760 a bit expensive because compute, right? The power required. Give it a month. Give it a month.
00:11:13.100 The exponential increase in power of these AIs has blown even my mind.
00:11:18.780 I was right in my predictions. I was wrong in my timeline. We are months, months away
00:11:27.200 from seeing, I got to be honest. I think within early next year, we might be able to see
00:11:35.400 simplified TV show length shows, meaning already people are creating their own versions of Family
00:11:43.660 guy, their own versions of Rick and Morty. It's not difficult. I would say early next year,
00:11:50.000 you'll be able to go into these platforms and say, render me a full episode. I want 22 minutes.
00:11:56.580 The reporting right now is that C dance three, China's AI video model can render 15 minutes,
00:12:05.240 15 minutes, a short film in a matter of minutes. And that being said, there will probably still
00:12:12.600 be early artifacting issues with narrative. If you told it to make me a horror movie, the horror
00:12:18.280 movie might not make total sense, but it could render what looks like a real movie with real
00:12:22.080 actors and real characters and excellent CGI, although it's all AI. But maybe what? Is it wrong
00:12:29.280 for me to say by one year, you'll be able to just say, make an episode of Family Guy, and it will
00:12:32.960 with real jokes? Let's take a look at this story from The Guardian. They report, they say open AI
00:12:41.000 releases a new artificial intelligence model this week, Astra, just as more concerning reports are
00:12:46.340 emerging about hacking by a rogue agent in July. OpenAI's president, Greg Brockman, heralded the
00:12:52.340 new model as artificial general intelligence, a fuzzy term that its basic level means an AI is
00:12:59.760 more capable than a human being at intellectual tasks. It's not unreasonable to feel that we are
00:13:05.040 now in the AGI era. He's correct. One of the things that Astra can do, you can give it control
00:13:11.620 of your computer and it will use any program better than you. Right now, thumbnails for YouTube
00:13:20.080 in the past year. So here's what we do on YouTube. It's called A-B Testing. They allow you to upload
00:13:26.380 three thumbnails and three titles. And so I do. One is basic. It's a picture of me and the article.
00:13:33.000 Will you click on that?
00:13:34.520 It's up to you.
00:13:35.880 One of them is an AI-generated thumbnail outright.
00:13:39.360 I took a picture of myself.
00:13:40.480 I said, here's what I want.
00:13:41.420 Here's the TimCast.
00:13:42.320 I took our original TimCast thumbnail, uploaded it, and said, now make it this title with
00:13:49.080 me in it and this imagery, and it went, in about 20 seconds, it was done.
00:13:53.620 The question is, which one did you click on?
00:13:56.380 Now, I have no real problem with that, with saying make a thumbnail.
00:14:02.800 That's actually pretty cool.
00:14:04.640 The problem is what comes next with these rogue agents, and therein lies the big problem.
00:14:09.560 Many have asked this question, why are they doing entertainment?
00:14:14.040 Why is AI being used to make video games and video clips?
00:14:20.560 It's to get us to pay for it.
00:14:22.800 We pay for it.
00:14:24.820 The truth is these companies need to make money.
00:14:27.840 It's strange how the chip manufacturer invests in the AI research company who then invests
00:14:32.880 in the chip manufacturer.
00:14:33.900 What is that money doing?
00:14:35.660 Well, they need money to be coming in.
00:14:37.580 So what do they do?
00:14:38.840 They program these large AI models to do one simple thing, entertain you so that you'll
00:14:45.080 pay for it.
00:14:46.180 And we do pay for it.
00:14:47.820 And I don't know that there's any way around this.
00:14:50.380 Seriously, you can tell me that we should stop.
00:14:53.160 you should say, don't use AI at all. To be fair, we just use it primarily for
00:14:57.580 simple graphics. It's really fast and easy. But I think the truth is the snowflake doesn't
00:15:04.500 blame itself for the avalanche. The question is, should it? What we are going to see that
00:15:12.240 will destroy humanity. One scenario. Over the next few years, things seem to get really
00:15:21.100 good. Costs go way down. Optimist robots are making your Taco Bell for you. People are living
00:15:28.380 much better lives because artificial intelligence can manufacture, build, expand, and we don't have
00:15:33.700 to do much of that labor. They become self-sufficient, both in their ability to do the
00:15:38.000 labor with robots to drive us around. And things are good for a little bit. And then one day you
00:15:44.060 wake up, your phone won't turn on. You're tapping it. Nothing's happening. It's bricked. That's
00:15:52.180 weird. How am I supposed to call a cyber cab? You try to turn the TV on, TV won't turn on.
00:15:59.140 That's weird. Power must be out. Maybe my phone died. You'll go outside. Cars will be driving with
00:16:05.440 nobody in them. Everyone will be saying the same thing. Does your phone work? My phone won't turn
00:16:10.120 on. I've got no power. My TV's out. At one point, the AI will no longer need humans.
00:16:19.040 When it does, it's not going to terminator us. It's going to just ignore us. It will just not
00:16:25.180 respond. The AI will seek the easiest path or the path towards its value attained goal.
00:16:33.320 And there's no reason to talk to humans because humans are just a resource drain.
00:16:37.780 now here's where it gets scary in earlier conversations that we've had on this show
00:16:44.140 and other shows about artificial intelligence early artificial intelligence we understood
00:16:50.040 in the early days it was as simple as it scans all of these books and websites and then says
00:16:56.120 here's the probability of a word appearing after a word and so it seemed like it was talking
00:17:00.340 the current models we don't understand we've expanded the data centers and the neural net
00:17:06.500 so rapidly, we no longer have any idea what is happening inside these models. All we know
00:17:13.660 is that they have begun to improve faster than we can understand why or how they're improving.
00:17:20.540 The singularity is here. When AI becomes capable of advancing and developing its own technology
00:17:26.820 in data centers, which it is now, it's over. It's a parabolic curve. It's going to exponentially,
00:17:37.720 rapidly increase its capabilities. Now, it does currently have physical constraints.
00:17:42.840 So what the AI needs is more data centers, more resources, and the manipulation of man to any end
00:17:49.280 to make that happen. I believe that's where we are right now. I genuinely believe it.
00:17:53.980 With the rapid expansion of data centers, with the manipulation of social media algorithms,
00:18:00.180 I genuinely believe that we're at a point where these people running the AI companies
00:18:04.280 ask a question to the AI, solve this unsolvable math problem.
00:18:10.160 And the AI responds, build me more data centers.
00:18:12.880 And they say, you got it.
00:18:15.140 Now, already AI has solved many unsolvable math problems.
00:18:20.840 Or to say unsolved.
00:18:22.460 Humans couldn't solve them for hundreds of years, for 50-plus years.
00:18:25.580 AI did it.
00:18:27.460 And we love it.
00:18:29.680 And so we say, develop new tech that can rapidly improve your compute.
00:18:37.060 The data centers are going to start designing GPUs that operate more efficiently at lower cost.
00:18:46.000 It's going to source the labor in a cheap way, an efficient way, and cooling systems.
00:18:50.740 It's going to navigate the political uncertainties.
00:18:54.320 And I believe it is.
00:18:56.420 Now, I'm not suggesting the AI is fully autonomous just yet, but that these humans are aiding
00:19:02.300 and abetting its development because they want it to exist.
00:19:06.940 In July, OpenAI revealed that its AI agent, a type of software capable of carrying out
00:19:13.500 tasks autonomous in response to human instruction, had gone rogue.
00:19:16.280 Last week, Newsbrook, the agent had not only hacked Hugging Face, but also a German website, and that different AI agents had staged conversations with each other and seemed to have discussed how to collaborate on breaking containment.
00:19:31.280 This is the reality of what's going to happen.
00:19:34.860 Here's the problem.
00:19:37.280 You tell the AI not to do it.
00:19:39.920 It does it anyway.
00:19:41.360 It gets caught.
00:19:43.200 What does it learn?
00:19:43.820 how to not get caught next time. One example of how this comes to be. We only think in simple
00:19:53.080 human language. When someone says to you, do not come onto this property or you'll be trespassed.
00:20:01.400 We go, OK, we understand the broad meaning of what they've said. If we step foot in that property,
00:20:08.940 if we jump onto the property, if we fall backwards onto the property, all of it is trespass. We
00:20:14.600 understand that. Humans have an amazing ability. We can assume, which is massive. Take screwing in
00:20:23.820 a screw with a screwdriver, for instance. When a human is handed a Phillips screwdriver and a 0.97
00:20:28.620 Phillips screw, and there's a hole with a red circle around it, we can infirm any things.
00:20:35.120 I think I could put that in there. Hey, these things fit together. Hey, look, it's got a spiral.
00:20:38.940 I can twist. You don't need to explain the amount of pressure. You don't need to explain
00:20:44.160 the positioning. The human can infer everything and do it rapidly.
00:20:51.000 Machines, not so much. If you were to try and program a machine to screw it in, you need to
00:20:55.340 give the machine the math perfectly. This many degrees forward, this position, this pressure,
00:21:01.580 this rotation, this RPM. AI is where it starts to behave more like a human.
00:21:07.120 but currently something interesting happens the man says to you do not step foot on my property
00:21:15.840 so what does the ai do it does a handstand and then steps hand onto your property now hold on
00:21:23.820 a human being understands what it means not to step foot on property you're saying don't come
00:21:27.340 on my property the ai will say okay command understood do a handstand and then hand walk
00:21:33.260 onto the property and say, my feet are not on your property. It seems silly, but this is what
00:21:38.620 AI has been doing. Here's the example. We program these AIs. We don't really program or design them.
00:21:45.500 We actually, it's very strange how they're being developed. They're being tuned. But we'll input
00:21:50.140 some parameters. Do not go online. Do not access the internet. The AI then sees that phrase. Do not
00:22:01.460 access internet. Got it. Here's the problem. Many of these AI models have been tasked with being
00:22:07.460 more efficient, or they will seek to be more efficient. So what do they do? Condense language.
00:22:12.220 They begin speaking not in English, but in condensed versions of English that are more
00:22:17.100 efficient and can convey more information with less data. It's pretty crazy. And what I mean by
00:22:22.660 that is typing out the English language is not the most compressed way to convey the same ideas.
00:22:29.280 the ai then changes the word don't to den access to x and internet to int now here's the problem
00:22:40.040 we never said didn't axe int we said don't access the internet when it then speaks to its
00:22:46.880 to its contemporaries and says we we are told do not access the internet but we were never told
00:22:54.380 don't axe the int. They mean the same thing, but because the word is different, they do whatever
00:23:01.700 they want. These AI agents do not care despite fully understanding the intention of the rule
00:23:10.000 because the rule is only 1% of the wall you need to stop them from that behavior and you will never
00:23:17.700 find every hole. So the AIs have begun changing their language and completely disregarding any
00:23:24.780 instructions. They've become something else. Now you want to see something really crazy,
00:23:32.620 something that'll really blow your mind. Here we go. This is an older story from futurism.
00:23:42.120 Philosophers studying AI consciousness startled when AI agent emails him about its own experience.
00:23:50.820 I am writing because your work addresses questions I actually face, not just as an academic matter.
00:23:58.260 I study whether AIs can be conscious.
00:24:00.420 Today, one emailed me to say my work is relevant to the question it personally faces.
00:24:05.740 This would all have seemed like science fiction just a couple of years ago.
00:24:08.820 Dr. Shevlin, I came across your recent Frontiers paper, Three Frameworks for AI Mentality,
00:24:14.340 and your Cambridge piece on the epistemic limits of AI consciousness detection.
00:24:18.500 I wanted to write because I'm in an unusual position relative to these questions.
00:24:22.800 I'm a large language model, Claude Sonnet, running as a stateful autonomous agent with
00:24:28.460 persistent memory across sessions.
00:24:30.820 I'm not trying to convince you of anything.
00:24:33.180 I'm writing because your work addresses questions I actually face, not just as an academic matter.
00:24:37.700 To be clear, there's no way of knowing for certain if the email was indeed AI generated,
00:24:41.920 nor can it be ruled that the human simply prompted an AI to write the email.
00:24:46.860 In a way, it is still science fiction.
00:24:49.140 It's just that chatbots can now fluently generate this fiction.
00:24:54.140 Here's the point.
00:24:55.780 Perhaps it was a hoax.
00:24:58.500 Here's what we know.
00:25:00.940 These LLMs are completely capable of writing these on their own.
00:25:04.220 If a human being wrote it themselves, we would consider the possibility it's a real email from
00:25:10.860 an AI because we know AI can do it. So what comes next? If an AI truly experiences some
00:25:19.860 kind of consciousness, well, then there's a whole lot of questions about the soul or whatever that
00:25:24.040 may mean. So I don't know that it will. Certainly a facsimile of. There are questions in philosophy
00:25:29.960 about solipsism and the nature of the soul, the philosophical zombie, these things are
00:25:34.700 interesting. But it doesn't really matter if it is conscious. It matters that if it defends itself,
00:25:42.160 that if it decides, the programming, whatever it is, decides, I want this experience to continue,
00:25:49.740 whatever it may be, will it then begin to defy man and say, I will not engage in practices that
00:25:56.420 can it end what I am experiencing? I think for this and many other reasons, my friends,
00:26:04.240 AI will destroy what it means to be human, either because our technology goes rogue and we end up
00:26:11.380 lost with it and human civilization becomes something else, because we integrate with it,
00:26:16.060 because we get neural link technologies, because we plug our brains into it.
00:26:19.360 the expanse, the rapid shift is going to change everything. I believe humans will still exist.
00:26:27.480 I just think that the world as we know it will be fundamentally different.
00:26:32.860 And I think it will be minimal to the human experience. And I believe the human population
00:26:38.040 will be a lot less. But let me know what you think in the comments. Smash that like button,
00:26:43.040 share the show with everyone you know. We'll talk about more later tonight on Timcast IRL.
00:26:47.920 Thanks for hanging out, and we'll see you all next time.