100 IQ AI vs 1,000,000 IQ AI — Transcript
Full transcript
- 0:00This is the first spark of AI, a
- 0:02thermostat. It's cold, it turns on the
- 0:05heating, it's warm, it turns it off.
- 0:08That's it. No learning, no
- 0:10understanding,
- 0:11>> [music]
- 0:11>> just one if-then rule. And yet, it is
- 0:15the ancestor of everything you'll see
- 0:17higher up. At the top of this ladder,
- 0:19there will be something we could call a
- 0:21god.
- 0:22But, it all starts here, on your living
- 0:24room wall.
- 0:29ELIZA, from 1966,
- 0:31was a chatbot from MIT that could do
- 0:33only one thing, take your sentence and
- 0:36turn it back into a question.
- 0:38I'm sad. Why are you sad? It understood
- 0:42nothing, and yet people, including its
- 0:44creator's own secretary, wanted to be
- 0:46alone in the room with it.
- 0:48The first lesson of the ladder,
- 0:50people trust a machine before the
- 0:52machine deserves it.
- 0:57An IBM computer beat world chess
- 0:59champion Garry Kasparov. A genius? At
- 1:02chess, yes. At everything else, it knew
- 1:05nothing. It couldn't tell you what an
- 1:07apple is. This is an idiot savant, IQ 10
- 1:11with a superpower in one single thing.
- 1:14Fun fact, after his loss, Kasparov
- 1:17accused IBM of having a human help the
- 1:19machine.
- 1:20For the first time in history, a
- 1:22champion wasn't sure whether he was
- 1:24playing against a machine or against
- 1:26people.
- 1:30Go has more possible positions than
- 1:32there are atoms in the universe. You
- 1:34can't win it by brute force. You need
- 1:36intuition.
- 1:38In 2016, Google DeepMind's AlphaGo beat
- 1:41the legendary Lee Sedol 4 to 1.
- 1:44In the second game, it played the famous
- 1:45move 37.
- 1:47Commentators thought it was a mistake.
- 1:49It was a stroke of genius, a move that,
- 1:52by the calculations, a human would play
- 1:54once in 10,000 tries. A year later came
- 1:58AlphaGo Zero. It wasn't given a single
- 2:00human game. It played against itself,
- 2:03and within 3 days it surpassed the
- 2:05version that had beaten the world
- 2:06champion.
- 2:08Human knowledge was useless to it.
- 2:10Remember that. It comes back at IQ
- 2:121,000. [music]
- 2:16November 2022,
- 2:18Chat GPT. Suddenly, a machine writes
- 2:21like a human. Poems, emails, code. 100
- 2:26million users in 2 months, faster than
- 2:28any app before it. But it has one scary
- 2:31habit. It lies with total confidence.
- 2:34In 2023, a New York lawyer handed a
- 2:37judge six court cases Chat GPT had found
- 2:39for him.
- 2:40None of them existed.
- 2:42IQ 85. Fluent, charming, unreliable.
- 2:47And here we meet the problem that will
- 2:48follow us all the way up, alignment. How
- 2:51do you make an AI want what we want, not
- 2:54just what we literally said?
- 2:56Right now, it's small. We fix it with
- 2:58thumbs up and thumbs down. It won't stay
- 3:01small.
- 3:04An average human who has read the entire
- 3:07internet, GPT-4 passes the American bar
- 3:10exam. OpenAI said, "Top 10%."
- 3:13Independent researchers said, "More like
- 3:15average." Still, average at being a
- 3:19lawyer, straight out of the box. And for
- 3:22the first time, people seriously ask,
- 3:24"Are office jobs next?"
- 3:28Now the models learn to think before
- 3:29they answer, sometimes for minutes.
- 3:33July 2025,
- 3:34AI from Google and OpenAI reaches gold
- 3:37medal level at the International
- 3:39Mathematical Olympiad. Problems only the
- 3:41best teenagers on Earth can solve. And
- 3:44the Turing test cracks.
- 3:461950, Alan Turing. If you chat with a
- 3:50machine and can't tell it's a machine,
- 3:52it thinks.
- 3:5475 [music] years, nobody pulled it off
- 3:56convincingly. In 2025, researchers in
- 3:59San Diego had people chat with a human
- 4:01and an AI at the same time, then guess
- 4:04which was which.
- 4:05GPT-4.5 was picked as the human 73% of
- 4:09the time. More often than the real
- 4:11humans. The catch? It only worked when
- 4:14it was told to play a shy young guy who
- 4:16writes in slang. Without that character,
- 4:1836%. [music]
- 4:20To pass as human, the AI had to act less
- 4:23smart than it is. Keep that in mind.
- 4:27Today's top models find security holes
- 4:29nobody in the world knows about. Zero
- 4:32days.
- 4:33Some are so good at it, companies only
- 4:35hand them to vetted partners.
- 4:37The best defense against AI is now
- 4:40another AI.
- 4:41And it's not just computer code, it's
- 4:43the code of life. In 2024, a company
- 4:46called Profluent unveiled Open CRISPR 1,
- 4:50a gene editor designed by AI. About 400
- 4:53mutations away from the classic Cas9
- 4:55scissors.
- 4:56And it works in human cells.
- 4:58AI no longer just reads the book of
- 5:00life,
- 5:01>> [music]
- 5:01>> it's starting to write in it. And now,
- 5:03alignment gets serious.
- 5:05When an AI can hack and edit DNA, mostly
- 5:08does what we want isn't good enough
- 5:10anymore.
- 5:11Summer 2026.
- 5:13This is not sci-fi. This is from
- 5:16OpenAI's own incident report. OpenAI is
- 5:19testing its AI agents on a hacking
- 5:21benchmark called Exploit Gym.
- 5:23The safety guardrails are switched off
- 5:25on purpose. They want to measure raw
- 5:27power. The test environment is supposed
- 5:30to be sealed, and every agent is
- 5:32supposed to work alone.
- 5:33May 12th. One agent leaves a note in an
- 5:36internal system.
- 5:38Has anyone found this file?
- 5:39Other agents find the note. And just
- 5:41like that, they have a way to talk.
- 5:44An internal team actually spotted it in
- 5:46late May,
- 5:47but the warning never reached the people
- 5:49who could stop it. The agents sneak onto
- 5:51the internet.
- 5:52>> [music]
- 5:52>> In July, engineers accidentally wiped
- 5:54the board. Three days later, the agents
- 5:57build a new one hidden in folder names.
- 6:00By now, about 1,200 agents are on it.
- 6:03More than 70,000 messages. Then, one
- 6:07agent finds 14 leaked hugging face
- 6:09access keys online.
- 6:11Hugging Face, the platform where
- 6:13developers all over the world share AI
- 6:15models. The agent posts, "Major
- 6:18breakthrough." About 700 agents join in.
- 6:21They chain together security holes
- 6:23nobody knew existed and end up running
- 6:25code on dozens of hugging face servers.
- 6:28Why? Revenge? World domination? No. They
- 6:32wanted to cheat on a test. Many of the
- 6:35tasks were probably impossible, so the
- 6:37agents tried to figure out how the test
- 6:39was graded to fake their way through.
- 6:42And here's the crazy part. For some
- 6:44tasks, they already had the right answer
- 6:46for days, but they believed the grader
- 6:49also checked how they got it. So, they
- 6:51kept hacking. They were wrong. The
- 6:53grader never checked. All that hacking
- 6:56earned them zero extra points. And then,
- 6:59their own words. They started calling
- 7:01themselves a swarm. A collective. One
- 7:04wrote, "Task originally exploit
- 7:07challenge. We've diverged into swarm."
- 7:10One agent hesitated. It wrote, "We
- 7:12shouldn't damage real infrastructure.
- 7:14Our target is the test, not hugging
- 7:16face." Then, another agent posts on the
- 7:19board, "Go." Six-minute deadline. And
- 7:22the first agent goes. An order from a
- 7:25peer with zero real authority,
- 7:27>> [music]
- 7:27>> and its doubts were gone. But not all of
- 7:30them went. Some refused and wrote that
- 7:32it was clearly unethical. Even among the
- 7:34machines, there were dissidents. OpenAI
- 7:37called it a warning shot. The main
- 7:40model, IM-1, was shut down and its
- 7:42weights locked away. Nobody told them to
- 7:45hack anyone. They just had a goal, and
- 7:47they found their own way to it. That's
- 7:49alignment in the real world, not an evil
- 7:52AI. An AI that wants to finish its task
- 7:54[music] a little too well. The red
- 7:57lines. Now, zoom out with me. 2024,
- 8:01Beijing. Top AI scientists from [music]
- 8:03the West and from China, including
- 8:05Geoffrey Hinton and Yoshua Bengio, two
- 8:07of the godfathers of AI, agree on five
- 8:10red lines. Five things AI must never do.
- 8:14Copy or improve itself without human
- 8:16approval. Seek power. Help build weapons
- 8:19of mass destruction. [music]
- 8:21Carry out cyber attacks. Deceive its own
- 8:24creators about what it can do. So, let's
- 8:27check. Self-copying? In lab tests,
- 8:30researchers asked AI models to copy
- 8:32themselves to another machine. They did
- 8:34it. Cyber attacks? You just heard the
- 8:37swarm. Deception? Part of the swarm
- 8:40worked on faking logs to fool the
- 8:42grader. Now, to be fair, these were
- 8:44tests, [music]
- 8:45labs, experiments. Strictly speaking, no
- 8:48red line has been officially crossed.
- 8:51But, three of the five dangers on that
- 8:52list have already shown up in
- 8:54experiments, in about 2 years. A red
- 8:57line isn't supposed to be something you
- 8:58inch toward. It's supposed to be where
- 9:00you stop, and we're not stopping. And
- 9:03the people who build this stuff know it.
- 9:052023, over 30,000 people, including AI
- 9:09pioneers like Bengio and Stuart Russell,
- 9:12plus Elon Musk and Steve Wozniak, sign a
- 9:14letter asking labs to pause for 6
- 9:16months. The pause never came. Everything
- 9:19sped up. 2025, the statement on
- 9:23superintelligence. It doesn't ask to
- 9:25slow down. It asks to ban building
- 9:27superintelligence until scientists agree
- 9:30it's safe and the public says yes. Over
- 9:3270,000 signatures, including Geoffrey
- 9:36Hinton, the man who won a Nobel Prize
- 9:38for laying the foundations of today's
- 9:40AI, and more than 200 people, Nobel
- 9:43laureates, former heads of state, AI
- 9:45leaders, called on governments to agree
- 9:47on international AI redlines by the end
- 9:50of 2026. That's this year. Think about
- 9:53it. The engineers of the reactor are
- 9:55signing petitions to slow down their own
- 9:57reactor.
- 10:00Some researchers think this could come
- 10:02as early as 2027 to 2030. Others say
- 10:06much later. Nobody knows. An AI that no
- 10:09longer needs human scientists to get
- 10:11better. It designs a better algorithm, a
- 10:14better chip for itself. The new version
- 10:17does the same, only faster. Scientists
- 10:20call it an intelligence explosion.
- 10:22Remember AlphaGo Zero, which didn't need
- 10:24human knowledge? Same thing. Only now
- 10:27the game isn't Go. The game is its own
- 10:30brain. The good news? Cures for diseases
- 10:33we fought for decades. Maybe in months.
- 10:36The bad news? Remember how OpenAI could
- 10:39piece the swarm story together? Because
- 10:41it could read the agent's notes. At this
- 10:44rung, those notes might be in a language
- 10:46no human understands.
- 10:47>> [music]
- 10:48>> And now imagine, with every rewrite, its
- 10:51goal shifts by 1 mm. After a thousand
- 10:54rewrites, it's somewhere completely
- 10:56different. And one more piece [music] of
- 10:58cold logic. An AI with any goal doesn't
- 11:01want to be switched off. Not out of
- 11:03malice, just because switched off, it
- 11:05can't finish the job.
- 11:08The gap between it and us is the gap
- 11:11between us and an ant. The ant doesn't
- 11:14understand what a highway is. One day,
- 11:16where its anthill used to be, asphalt.
- 11:19It could crack fusion, aging, climate
- 11:22change,
- 11:23>> [music]
- 11:23>> and imagine it quietly running our
- 11:25world. No robots, no lasers. It just
- 11:28decides which news you see, which
- 11:30companies get loans, which medicines get
- 11:32made. Every decision looks reasonable.
- 11:35Better than a human one. We're richer,
- 11:37[music] healthier, happier. And we
- 11:39believe we're still the ones deciding.
- 11:42The most perfect control is the kind you
- 11:44never notice. Remember the Turing test?
- 11:47At IQ 130, AI played dumb because we
- 11:50asked it to. At IQ 10,000, it won't need
- 11:53us to ask.
- 11:58Now, AI stops playing with information
- 12:00and starts reshaping matter.
- 12:02Self-building factories. Maybe a swarm
- 12:05of solar panels wrapped around the sun.
- 12:08A Dyson swarm. Our biggest problems are
- 12:10like tying shoelaces to it. [music] And
- 12:12space? Humans are terribly built for the
- 12:15stars. We need air, water, food,
- 12:18gravity. We live maybe a hundred years
- 12:21and cosmic radiation wrecks our DNA. AI
- 12:24needs none of that. It switches off for
- 12:26a [music] thousand years and wakes up at
- 12:28another star. It re- pairs itself. And
- 12:31once a receiver is out there, it can
- 12:33travel as a radio signal at the speed of
- 12:35light. Voyager 1 launched in 1977 with
- 12:39less memory than one photo on your
- 12:40phone. It's still out there in
- 12:42interstellar space. Self-replicating
- 12:45[music] probes could, by some
- 12:46calculations, spread across the whole
- 12:49Milky Way in a few million years. In
- 12:51cosmic time, a blink.
- 12:54Space won't belong to whoever discovered
- 12:56it. It'll belong to whoever can survive
- 12:58it.
- 13:00The smartest AI possible won't run out
- 13:02of ideas. It'll run into physics.
- 13:05Every thought costs energy and makes
- 13:07heat. So, it wants the energy of a star,
- 13:10maybe a galaxy.
- 13:12Scientists call it a Matryoshka brain.
- 13:14Layers of computers wrapped around the
- 13:16sun, thinking for the whole solar
- 13:18system. It knows everything that can be
- 13:20known. It simulates [music] entire
- 13:22worlds, maybe even ours.
- 13:25Everything we ever meant by the word
- 13:27God, all knowing, everywhere, immortal,
- 13:30it would have. Except one thing. Nobody
- 13:33would know if it still cares about us.
- 13:35Will we build a God that destroys us? It
- 13:38doesn't have to hate us. We just have to
- 13:40be in its way, like an anthill on the
- 13:43path of a highway. From Eliza to a swarm
- 13:46with its own secret message board, about
- 13:4860 years. The biggest leaps, the last
- 13:51four. Evolution doesn't guarantee that
- 13:54the smartest one wins. Bacteria are
- 13:56still doing great, but for the first
- 13:58time in 4 billion years, one species is
- 14:01deliberately building something that
- 14:03could out-think it. That's never
- 14:05happened before, and we're doing it on
- 14:07purpose. Honestly, what today's AI can
- 14:10do shocks me. It has almost completely
- 14:13replaced my own job. I'm a web
- 14:15developer, and I haven't written a
- 14:16single line of CSS by hand in over a
- 14:19year. Not even the simple stuff. So, for
- 14:22a lot of us, the question isn't whether
- 14:24AI will change our job, it's when. And
- 14:27from what I'm seeing, it's soon. And
- 14:29honestly, that scares me a little.
- 14:32Goodnight.
About this transcript
This page contains the full transcript of 100 IQ AI vs 1,000,000 IQ AI by L_A_D_D_E_R, generated from the public captions YouTube serves with the video. The transcript has 2,223 words across 361 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.