not another AI video — Transcript
Full transcript
- 0:00December 2012,
- 0:03a casino hotel at Lake Tahoe, Nevada. In
- 0:07one of the rooms, a computer science
- 0:09professor called Joffrey Hinton is
- 0:11running a secret auction by email. What
- 0:14he's selling is a company he's just
- 0:16founded with his two students. It has
- 0:18three employees, no product, no revenue,
- 0:22and no business plan. Just one research
- 0:26paper. The biders, Google, Microsoft,
- 0:30the Chinese tech giant BYU, and a small
- 0:32London startup called Deep Mind. The
- 0:36bits climb by millions at a time. And in
- 0:39the end, Google wins for $44 million.
- 0:44This was the big bang of modern AI, the
- 0:48moment decades of university research
- 0:50started moving into the hands of big
- 0:52tech.
- 0:54Everyone talks about what AI will do to
- 0:56us. Will it kill us all? Take all our
- 0:59jobs or cure cancer, do all the work,
- 1:02and let us just have fun for the rest of
- 1:04our lives? The truth is, we can't see
- 1:07the future. But what we can do is look
- 1:09at where AI came from because that might
- 1:13help us understand where it's going. So,
- 1:15this video is the history of AI. And
- 1:18this is also the third video in my big
- 1:21tech series. Each video of my series
- 1:24works on its own, so you don't need to
- 1:26have seen the others, but if you want
- 1:28the full picture, I'd really recommend
- 1:31watching the other two as well after
- 1:33this video. Okay, but without further
- 1:36ado, let's get into the history of AI.
- 1:41First, we should probably define what AI
- 1:44actually is. Everyone is talking about
- 1:46AI, but only few people understand the
- 1:50concept of machine learning behind it.
- 1:52Artificial intelligence isn't really a
- 1:54scientific term. It's closer to a
- 1:57marketing label. The problem with the
- 1:58word AI is that it's like using the word
- 2:01vehicle for a bicycle, a truck, and a
- 2:03moon rocket at the same time. At least
- 2:06that's how two Princeton computer
- 2:08scientists, Arvin Narayan and Sash
- 2:12Kapoor, put it. Also, as always, I've
- 2:15made a document with all the sources I
- 2:17used in the video in the description.
- 2:19But back to the term AI. Everyone says
- 2:22AI, but hardly anyone knows what they
- 2:25actually mean by it. Because when most
- 2:28people say AI today, they often talk
- 2:31about a very specific form of AI,
- 2:34chatbots, like chatbt.
- 2:37So, let's talk about those first because
- 2:40I assume most of you are familiar with
- 2:42them and it will make the other ones
- 2:44easier to understand too. Under the hood
- 2:47of those chat bots is something called a
- 2:49large language model or LLM. And an LLM
- 2:53is basically autocomplete on an
- 2:55astronomical scale. But to really
- 2:58understand what that means, I think it's
- 3:00easiest to contrast it to classical
- 3:03programming. Writing software is a
- 3:05rule-based endeavor. A human programmer
- 3:08explicitly writes instructions.
- 3:11Like if user clicks button then save
- 3:14file. If you wanted to write a computer
- 3:18program that writes English, you'd have
- 3:21to write down millions of rigid
- 3:23grammatical rules, dictionary
- 3:24definitions, and exception trees.
- 3:27Historically, this never really worked
- 3:30because human language is too messy and
- 3:32nuanced to capture in static rules. The
- 3:36alternative is using something called
- 3:38machine learning. Instead of writing the
- 3:40rules yourself, you give the computer a
- 3:43blank mathematical model and a mountain
- 3:46of examples and tell it figure out the
- 3:49pattern yourself. And that's machine
- 3:51learning. And machine learning is behind
- 3:54almost everything we call AI. your
- 3:56Netflix recommendations, Face ID,
- 3:59self-driving cars. A large language
- 4:02model is just the version of it that
- 4:04learned language. And to explain how
- 4:06machine learning works, we are going to
- 4:08look at language. To visualize it, I'm
- 4:11going to build my very own language
- 4:13model. But before we continue, I wanted
- 4:16to tell you how these videos get made
- 4:19because loads of you have been curious
- 4:21how I go about my research. So, here's
- 4:24how I usually work. Making these kinds
- 4:26of videos takes a lot of research. I
- 4:28read a lot of books, articles,
- 4:30scientific papers, and what I do most of
- 4:33my reading in is UPDF, the sponsor of
- 4:37today's video, because this is where I
- 4:40can work best. I have tried other PDF
- 4:42readers before, like Adobe Acrobat,
- 4:45which has an insane subscription fee, or
- 4:48my laptop's preview, but none of them
- 4:50let me work as well as UPDF. This is
- 4:53their 1955 Dartmouth proposal, the
- 4:56document that coined the term artificial
- 4:58intelligence and this is attention is
- 5:01all you need from 2017, the paper behind
- 5:04chat gvt and every chatbot since. These
- 5:07were two very important sources that
- 5:09I've based this video on. And the cool
- 5:13thing about UPDF is that I can highlight
- 5:16and annotate right on the page while
- 5:18reading it. And with the AI sidebar
- 5:21open, I can ask it to explain a passage
- 5:23if I have difficulty understanding
- 5:24something.
- 5:26Also, a lot of early research papers
- 5:29like this one are scans of typewritten
- 5:31pages. So, normally you can't search
- 5:34them or copy anything out of them. With
- 5:36UPDF's OCR, one click turns it into real
- 5:40text. Now, I can search it, highlight,
- 5:43and annotate right on the page. Also, I
- 5:46use UPDF's e-sign function to sign
- 5:50contracts or important documents. So,
- 5:53that's been really useful for me, too.
- 5:56UPDF works on Mac, Windows, iPhone, and
- 5:59Android. And one license covers all
- 6:02four. You can download it for free. And
- 6:05if you're a heavy user like me, the pro
- 6:07version costs about a sixth of Adobe
- 6:09Acrobat and it has a lifetime option, so
- 6:13you don't have to keep paying for a
- 6:14subscription. There's a limited time
- 6:17launch offer for UPDF 3.0 right now. The
- 6:21link is at the top of the description.
- 6:23And now, let's get back to the video.
- 6:26I'm going to build a language model out
- 6:28of paper cups to visualize how large
- 6:30language models work. I'm feeding our
- 6:33DIY language model a bit of text from my
- 6:36last video where I explained how chips
- 6:38are made by making a photo print of my
- 6:41cat blobber, which also is very useful
- 6:45to understand the foundation AI is built
- 6:47on. So, I highly suggest you go watch
- 6:49the video on chips after this video if
- 6:52you want a deeper understanding. But
- 6:54anyway, here's how we're setting up our
- 6:56very own large language model. I mean,
- 6:59it's not going to be large, it's going
- 7:00to be a language model. I go through the
- 7:03text word by word and for every word I
- 7:05write down the word that comes right
- 7:07after it. Every time the word the
- 7:10appears, I look at the next word, write
- 7:13it on a slip and drop it in the the cup.
- 7:16Okay, this is the the cup. It has 10
- 7:19slips in it. Three of them say pattern
- 7:22and the others say things like chip,
- 7:25light, mask. So if I draw from this cup,
- 7:29pattern is three times more likely to
- 7:32come out than anything else. So there's
- 7:34a higher probability I will draw
- 7:36pattern. And here's something
- 7:38interesting. Most words don't even need
- 7:41a cup. Nothing, for example, is in my
- 7:43text only ever followed by happens. So
- 7:46whenever I land on nothing, there's no
- 7:49choice to make. The next word is always
- 7:52happens. Only 25 words in this text have
- 7:56more than one option. So only they
- 7:58really need a cup. Now let's make it
- 8:00talk. I start with the
- 8:08chip.
- 8:09Chip has a cup.
- 8:14people
- 8:23use
- 8:30blobber.
- 8:32So yeah, we just built a language model
- 8:34out of paper cups and it just told us
- 8:37that chip factories use my cow. It
- 8:40sounds like a real sentence. It's
- 8:42grammatically fine and it's obviously
- 8:44complete nonsense because the model
- 8:47doesn't know what a chip is. It doesn't
- 8:50know that blobber is a cat. All it knows
- 8:52is which words followed which in my
- 8:54script and how often. Computer
- 8:57scientists call what we just did an
- 8:59engram model. The n is how many words in
- 9:02a row the model looks at. Ours always
- 9:05looks at two words in a row. So it's
- 9:08called a bagram model. Now, as we've
- 9:11seen, my language model is very limited.
- 9:14It only knows the words in this passage.
- 9:17So, it can never say anything else. It
- 9:20can't say dog. It can't say hello. And
- 9:23because most words only have one slip,
- 9:25most of the time it just repeats my
- 9:27script back to me. Generally, if you
- 9:29only look at two words at a time, the
- 9:32sentences quickly degenerate into
- 9:34repetitive or disjointed gibberish. So,
- 9:36the obvious fix would be let it look at
- 9:39more words at once. Not just the last
- 9:41word, as in the two-word combinations we
- 9:44used here, but five word combinations.
- 9:47But that's where it breaks. The more
- 9:49words in a row you look at, the rarer
- 9:52each combination gets. Most sentences
- 9:54you write have probably never been
- 9:57written exactly that way before. So, for
- 9:59almost any five words, there's simply no
- 10:03cop. the model has never seen them and
- 10:06has no idea what comes next. This is
- 10:09called the sparsity problem and it's a
- 10:11big part of why language models like
- 10:13mine were pretty useless for decades. In
- 10:17other words, the sparsity problem is
- 10:19essentially the issue that the more
- 10:22words in a row a model looks at, the
- 10:24more likely that exact combination has
- 10:27never been written before. So, the model
- 10:30has nothing to learn from and no idea
- 10:32what comes next. So, if the sparsity
- 10:34problem kept language models useless for
- 10:36decades, how did tragic get past it?
- 10:40Part of the answer is the obvious one.
- 10:43It got way more text, a lot more. My cup
- 10:46model read 130 words about visualizing
- 10:49chipm
- 10:52read a huge chunk of the internet,
- 10:54websites, Wikipedia, Reddit, millions of
- 10:57books, code. And quick side note, these
- 11:00models don't actually read words. They
- 11:03read tokens, chunks of words. A long
- 11:07word gets chopped into pieces. On
- 11:09average, a token is about 3/4 of a word.
- 11:12Chachi has never heard of my cat, so it
- 11:15doesn't see blobber. It sees blob and
- 11:19burr. Or economics is chopped up into
- 11:22econom x.
- 11:24And you can see how words get tokenized
- 11:27yourself. Open AAI has a free tool where
- 11:29you type in any text and it shows you
- 11:31exactly how it gets chopped into tokens.
- 11:34And tokens matter for a lot of things.
- 11:36First, they explain why Chapd types its
- 11:39answer chunk by chunk. Here it is in
- 11:41slow-mo. It's because it generates one
- 11:44token at a time. Secondly, chat's memory
- 11:48is measured in tokens. Every chatbot can
- 11:51only keep a certain number of tokens in
- 11:53mind at once. That's called the context
- 11:56window. Today that can be hundreds of
- 11:59thousands of tokens, but it's still a
- 12:01limit. And once a conversation gets too
- 12:03long, the oldest parts start to fall
- 12:06out. That's why in really long chats,
- 12:08Chhatributt sometimes forgets what you
- 12:11told it at the start. And third, tokens
- 12:14are money. Companies that build products
- 12:16on top of Jatribute pay Open AI per
- 12:19token for every token that goes in and
- 12:22every token that comes out. So every
- 12:24question you ask and every answer you
- 12:26get has a price tag that is measured in
- 12:28tokens. And tokens are also how AI
- 12:31companies measure how much a model has
- 12:34read in total. So chatbot's entire
- 12:36knowledge base is measured in tokens. So
- 12:39how much did chat read? Open AAI doesn't
- 12:42say exactly how much anymore, but Meta
- 12:44does say it for its own model. Llama 3
- 12:48around 15 trillion tokens, which is
- 12:50roughly 11 trillion words. If you read 8
- 12:53hours a day every day, that would take
- 12:55you about 250,000 years. Okay, but back
- 12:59to my own language model and its
- 13:00limitations. Remember the sparsity
- 13:03problem? The problem that the more words
- 13:05or tokens in a row a model looks at, the
- 13:08more likely that exact combination has
- 13:11never been written before. So, the model
- 13:14has nothing to learn from and no idea
- 13:16what comes next.
- 13:18Well, even after 250,000 years worth
- 13:21reading, most fiveword combinations
- 13:24still never show up. So, more text alone
- 13:27doesn't solve it. So, how did Chhatri
- 13:30get around it? Two ways. First, Chhatri
- 13:34doesn't use cops. Instead, it has
- 13:36hundreds of billions of so-called
- 13:38weights. GPD3, Chhat GPT's predecessor,
- 13:43had 175 billion of them. Think of them
- 13:46as tiny dials. At the start, they're all
- 13:50set randomly. During training, it
- 13:52guesses the next token, and when it's
- 13:54wrong, a bit of maths called gradient
- 13:57descent nudges all the dials a tiny bit,
- 14:00so it guesses slightly better next time.
- 14:03Then you repeat that across trillions of
- 14:05tokens. The result is that it doesn't
- 14:08memorize word pairs. It learns patterns.
- 14:12It learns, for example, that cat and
- 14:14kitten get used in similar ways. So, it
- 14:17can deal with sentences it has never
- 14:19seen before. Second, it doesn't just
- 14:22look at the last word. It looks at
- 14:24everything, your whole question, the
- 14:26whole conversation, hundreds of
- 14:28thousands of tokens back. So, it's the
- 14:31same game as my paper cups. Predict the
- 14:33next word, just with billions of dials
- 14:35instead of slips of paper and a memory
- 14:37that reaches back across the entire
- 14:39conversation. Now, remember the other
- 14:41vehicles under the AI umbrella I
- 14:43mentioned at the start? Netflix
- 14:45recommendations, Face ID, self-driving
- 14:47cars. Here's the thing. They basically
- 14:50all went the same. Same dials, same
- 14:52nudging, just different examples. Face
- 14:55ID learned from millions of photos of
- 14:57faces. Netflix and Tik Tok learned from
- 14:59what billions of people watched.
- 15:01Self-driving cars learned from millions
- 15:03of hours of driving footage. And image
- 15:05generators like midjourney or deli learn
- 15:09from billions of images with captions.
- 15:11So different vehicles, same engine. Show
- 15:14a machine enough examples and let it
- 15:16notch its dials until it finds the
- 15:18pattern. That's machine learning. But it
- 15:21wasn't always like this. For most of
- 15:23AI's history, the people building it on
- 15:25the exact opposite approach, writing the
- 15:27rules by hand, the if then way. And that
- 15:30story is the one we are going to look
- 15:32at. It starts in 1950 with one question.
- 15:36Can machines think? The man who asked it
- 15:39was Alan Turing. And if that name sounds
- 15:42familiar, it's because he's the guy who
- 15:44helped break the Nazis Enigma code in
- 15:47World War II. Turing thought the
- 15:49question, "Can machines think?" was too
- 15:52vague to answer because what does
- 15:55thinking even mean? So, he swapped it
- 15:57for a different question, one you could
- 15:59actually test. He called it the
- 16:02imitation game.
- 16:04Imagine you're texting with two people
- 16:06you can't see. One is a human, one is a
- 16:09machine. You can ask them anything you
- 16:12want. And if you can't reliably tell
- 16:15which one is the machine, then Turing
- 16:18said, "We might as well say it can
- 16:20think."
- 16:22This became known as the Turing test.
- 16:25And you could argue it was the birth of
- 16:27AI's original sin. Defining intelligence
- 16:30not by whether a machine actually
- 16:32understands anything, but by whether it
- 16:35can convince a human that it does.
- 16:38Turing died in 1954, so he never got to
- 16:40see what came next. But just one year
- 16:43later in 1955, a young maths professor
- 16:46called John McCarthy wrote a funding
- 16:49proposal for a summer workshop at
- 16:51Dartmouth College in the US. together
- 16:54with Marvin Minsky, Claude Shannon, and
- 16:56Nathaniel Rochester from IBM. And it's
- 16:59in this proposal that the term
- 17:01artificial intelligence was used for the
- 17:04very first time. This is where the name
- 17:06comes from. And remember how I said AI
- 17:09is basically a marketing label. Well, it
- 17:12was from the very start. McCarthy later
- 17:15said he picked the name partly to set
- 17:17his field apart from a rival one called
- 17:19cybernetics. Cybernetics is the study of
- 17:22how machines and living things control
- 17:23themselves through feedback like a
- 17:26thermostat that keeps adjusting to hold
- 17:28a room at the right temperature. But
- 17:31back to the proposal. The proposal said
- 17:33that a significant advance could be made
- 17:35if a carefully selected group of
- 17:37scientists work on it together for a
- 17:40summer. And what do they want to solve
- 17:43in that one summer? They said they
- 17:45wanted to get machines to use language,
- 17:48form concepts, solve problems that only
- 17:51humans could solve and improve
- 17:53themselves in one summer. That was 70
- 17:56years ago. And there was a lot of
- 17:58optimism at the start. These early AI
- 18:00pioneers genuinely believed thinking
- 18:03machines were just around the corner.
- 18:05Two of the people who came to the
- 18:07Dartmouth summer research project on
- 18:09artificial intelligence were Herbert
- 18:13Simon and Alan Newwell. And they even
- 18:16brought a working program. It could
- 18:18prove mast theorems by itself. And a
- 18:23year later, they made some bold
- 18:24predictions. Within 10 years, they said
- 18:27a computer would be the world chess
- 18:29champion. Well, it actually ended up
- 18:31taking 40 years until IBM's Deep Blue
- 18:35beat world champion Gary Kasparov in
- 18:381997. But still, in 1967, Mavin Minsky,
- 18:43one of the Dartmouth organizers, wrote
- 18:45that within a generation, the problem of
- 18:48creating artificial intelligence will
- 18:51substantially be solved. But yeah, my
- 18:53point here is that these early AI people
- 18:56were very, very optimistic. But while
- 18:59they all agreed thinking machines were
- 19:01coming, they completely disagreed on how
- 19:04to build them. And the field split into
- 19:07two camps with totally different ideas
- 19:10of what intelligence even is. The first
- 19:12camp said, "Intelligence is rules. If we
- 19:16can write down all the rules of logic
- 19:18and knowledge, a computer can follow
- 19:21them and think." This is called symbolic
- 19:24AI. It's basically the if then approach
- 19:27of classical programming and most of the
- 19:29Dartmouth crowd were in this camp.
- 19:32McCarthy, Minsky, Simon, Newell. But the
- 19:36second camp said, "Intelligence is
- 19:38learning. Our brains aren't rule books.
- 19:42They are networks of neurons that learn
- 19:44from experience.
- 19:46So let's build machines that work like
- 19:48that." This is called connectionism or
- 19:51neural networks. And its biggest
- 19:53champion was a guy called Frank
- 19:55Rosenblat. Rosenblat wasn't a computer
- 19:57scientist. He was a psychologist at
- 19:59Cornell. And what he really wanted to
- 20:02understand was the human brain. How a
- 20:05lump of neurons can learn to recognize
- 20:08things it has never been taught. So he
- 20:11decided the best way to understand the
- 20:13brain was to build one. In 1958,
- 20:17Rosenlat and the US Navy, who funded his
- 20:20research, showed off the Perceptron for
- 20:23the first time. It was a program running
- 20:25on a giant computer. And a couple of
- 20:27years later, Rosenblad built it as its
- 20:30own machine, the Mark1 Perceptron, a
- 20:32fridge-sized computer that was hooked up
- 20:35to a camera with 400 pixels, which
- 20:37translates to a 20x 20 grid of light
- 20:40sensors, which was basically the
- 20:41perceptron's eye. The light from an
- 20:44image was passed onto 512 internal units
- 20:47connected to adjustable dials. Those
- 20:49dials were its weights and then it
- 20:52learned. You'd show it cards with a mark
- 20:55either on the left or on the right.
- 20:57Every time it got one wrong, it adjusted
- 20:59its weights a little. Until it could
- 21:01reliably tell the patterns apart. And
- 21:03here's why this was such a big deal. The
- 21:05perceptron was one of the first systems
- 21:07to turn an image into numbers. basically
- 21:10coordinates on a map and then learn
- 21:13where to draw the line between left and
- 21:15right. No one had programmed any rules
- 21:17into it. It had learned the difference
- 21:19all by itself. And that's the exact same
- 21:22principle behind image recognition today
- 21:24and behind the embeddings we talked
- 21:26about with Chach. But what really made
- 21:29the perceptron famous was the hype. In
- 21:32July 1958, when it was first shown to
- 21:35the world, the US Navy, who were funding
- 21:37Rosenlat's research, held a press
- 21:40conference in Washington, and the press
- 21:42went wild. The New York Times reported
- 21:44that the Navy had revealed the embryo of
- 21:48an electronic computer that it expected
- 21:51will be able to walk, talk, see, write,
- 21:54reproduce itself, and be conscious of
- 21:56its existence. I mean, this was a
- 21:59machine that could tell left and right,
- 22:01but the headlines at the time made it
- 22:03sound like it was about to become
- 22:05conscious. Rosenlat's machine was the
- 22:08learning cam's big moment, the
- 22:10perceptron, the very first neural
- 22:12network to make headlines, a machine
- 22:14that learned from examples, just like
- 22:16the AI we use today. So, yeah, it seemed
- 22:19like the learning camp had figured it
- 22:21out. But the rules camp had its own
- 22:23showpiece and it might be the most
- 22:25famous early AI program of all. In 1966,
- 22:29a computer scientist at MIT called Yseph
- 22:33Bitesenbound built Eliza, one of the
- 22:36very first chatbots. Eliza pretended to
- 22:39be a therapist. So yes, people have been
- 22:42oversharing with computers since 1966.
- 22:45Uh 6 years before people started trauma
- 22:48dumping on chatbt, they did it with
- 22:50Eliza. And Eliza worked with a few
- 22:52simple rules. It spotted a keyword in
- 22:55what you typed, turned your own words
- 22:57back into a question. And of course,
- 22:59people loved it. They opened up to it.
- 23:01They told it about their problems, their
- 23:04relationships, their families, and even
- 23:06formed deep emotional bonds with Eliza.
- 23:08Right. Bam's own secretary, who had
- 23:11watched him build it for months and knew
- 23:13it was just a program, reportedly asked
- 23:15him to leave the room so she could talk
- 23:18to Eliza in private. And Whitenbalam was
- 23:20horrified. He'd built Eliza partly to
- 23:23show how shallow these conversations
- 23:25with a computer really were. Instead,
- 23:28he'd shown how easily humans believe a
- 23:30machine understands them, even when it's
- 23:33just following a handful of rules. This
- 23:36became known as the Eliza effect. our
- 23:38tendency to project empathy,
- 23:40understanding, and intention onto
- 23:42software that has none. And it got
- 23:45worse. Some psychiatrists actually
- 23:47suggested using programs like Eliza to
- 23:50treat real patients. And we're seeing
- 23:53that today as well, where people use
- 23:56chatbt and other chat bots as a
- 23:59replacement for actual therapy. But
- 24:01yeah, Whitesenbound was so disturbed
- 24:03that he spent the rest of his career
- 24:05warning people about trusting computers
- 24:07too much. But there was another thing.
- 24:10Programs like Eliza made AI look much
- 24:13further along than it really was. From
- 24:15the outside, it seemed like all those
- 24:17big promises from the Dartmouth summer
- 24:19camp were finally coming true. But by
- 24:23the late 1960s, all these big promises
- 24:26started running into reality. And that
- 24:28reality is best explained by looking at
- 24:30an example like machine translation.
- 24:33During the Cold War, the US government
- 24:35spent millions trying to build computers
- 24:38that could automatically translate
- 24:40Russian into English. Very handy if you
- 24:43want to read Soviet documents. But in
- 24:461966, a government committee published
- 24:49what's called the ALPAC report. The
- 24:51verdict was machine translation was
- 24:54slower, less accurate, and twice as
- 24:56expensive as just paying a human
- 24:58translator. And the funding dried up.
- 25:01Remember the two camps I talked about
- 25:03earlier? The rules camp who wanted to
- 25:05write down all the rules of intelligence
- 25:08by hand and the learning camp who wanted
- 25:11machines to learn from examples like
- 25:14Rosenlat's perceptron. Well, machine
- 25:16translation was a blow for the rules
- 25:18camp. But in 1969, the learning camp got
- 25:21hit too. And the attack came from its
- 25:24biggest rival. Marvin Minsky, one of the
- 25:26leaders of the rules camp, and his
- 25:28colleague Samur published a book called
- 25:31Perceptrons.
- 25:32In it, they proved mathematically that a
- 25:35perceptron like Rosenlats couldn't solve
- 25:37some surprisingly simple problems. For
- 25:40example, it couldn't learn to tell
- 25:42whether two dots were the same color or
- 25:45different colors. and they suggested
- 25:48that adding more layers probably
- 25:50wouldn't fix it. Many people read the
- 25:52book as proof that neural networks were
- 25:54a dead end and the money for neural
- 25:56network research largely dried up too.
- 25:58And then in 1973, the British government
- 26:01asked the mathematician Sir James
- 26:04Lighill to review the whole field of
- 26:06artificial intelligence. His conclusion
- 26:08was brutal. In no part of the field have
- 26:11the discoveries made so far produced the
- 26:14major impact that was then promised. One
- 26:17big reason why there was no advancement
- 26:19was something called the combinatorial
- 26:22explosion. The rules approach worked on
- 26:24small toy problems. But as soon as
- 26:26problems got even a little bigger, the
- 26:29number of possibilities exploded beyond
- 26:32what any computer could handle. But
- 26:35yeah, this ended up with the UK cutting
- 26:38most of its AI funding. And in the US,
- 26:41DARPA, the military research agency that
- 26:44had been AI's biggest sponsor, started
- 26:47only funding research with direct
- 26:49military uses. And just like that, the
- 26:52money for AI research was gone. This is
- 26:55what's now called the first AI winter.
- 26:58But AI didn't stay dead for long. In the
- 27:011980s, the pendulum swung back. And this
- 27:04time the rules camp returned with a
- 27:06rebrand expert systems. Essentially they
- 27:11stopped trying to build a general
- 27:12thinking machine but instead build
- 27:14machines that could do one narrow job
- 27:18like diagnosing infections or
- 27:20configuring computers. Their approach to
- 27:22building that was to sit down with a
- 27:24human expert and write down everything
- 27:26they know as hundreds of thousands of if
- 27:29then rules. And for a while this
- 27:32actually worked really well. The
- 27:33computer company DEEC used an expert
- 27:36system called XCON to put together
- 27:38customer orders and it reportedly saved
- 27:41them tens of millions of dollars a year.
- 27:43Suddenly AI wasn't just a lab experiment
- 27:46anymore. It was a business. By the late
- 27:481980s, AI had become a billion-doll
- 27:50industry. Another famous example was
- 27:53mycin, a system built at Stanford in the
- 27:561970s that diagnosed blood infections
- 27:59and recommended which antibiotics to
- 28:01give. In tests, it performed about as
- 28:04well as Stanford's own infection
- 28:06specialists, but it was never actually
- 28:08used on real patients, partly because
- 28:11nobody could agree who'd be responsible
- 28:13if it got something wrong. This new AI
- 28:15boom was so big that a whole industry
- 28:17grew up around building special
- 28:19computers just for AI called lisp
- 28:22machines, named after lisp, the
- 28:25programming language most AI was written
- 28:27in. Companies poured huge sums into
- 28:30them. But yeah, the fact is that there
- 28:32were more and more real world uses for
- 28:34AI. And amidst this AI upswing,
- 28:38something interesting happened in 1982.
- 28:41Far away from the US and the UK,
- 28:44Japan announced the fifth generation
- 28:46computer project. A 10-year government
- 28:50funded plan to build intelligent
- 28:51computers and leapfrog the West. And the
- 28:54West obviously panicked. DARPA, the same
- 28:58agency that had cut the money a decade
- 29:00earlier, launched its strategic
- 29:02computing initiative. Also, I talk a lot
- 29:04more about DARPA in this video. It's
- 29:07actually the agency that basically built
- 29:09big tech. But yeah, DARPA launched the
- 29:12strategic computing initiative and the
- 29:15UK launched its own program and hundreds
- 29:18of millions of dollars were poured back
- 29:20into AI. But just like the first time,
- 29:22the hype didn't last. These expert
- 29:25systems turned out to be incredibly
- 29:27brittle. They weren't great as long as
- 29:29the situation was exactly what the rules
- 29:31were written for. But throw in anything
- 29:33unexpected and they just fell apart.
- 29:36These systems had no common sense and
- 29:38every time the world changed a bit, a
- 29:40human had to sit down and write new
- 29:43rules by hand. That got very expensive
- 29:46very fast. And then in 1987, the market
- 29:49for specialized AI computers, the
- 29:51so-called lisp machines, collapsed
- 29:54almost overnight. Ordinary desktop
- 29:56computers had become powerful enough to
- 29:59do the same job for a fraction of the
- 30:01price. Dozens of AI companies went bust
- 30:04and in 1992, Japan's fifth generation
- 30:07project ended without reaching its
- 30:09goals. This was the second AI winter.
- 30:12And this time AI became such a dirty
- 30:15word that researchers branded their work
- 30:18as machine learning informatics calling
- 30:21their work anything but artificial
- 30:23intelligence. And remember how I said at
- 30:26the start that AI is basically a
- 30:28marketing label. Well, when the label
- 30:30stopped selling, they just changed it.
- 30:33But while the rules camp was crashing,
- 30:35something was quietly happening in
- 30:37background. remember the learning camp,
- 30:40the neural network people. After
- 30:42Minsky's book, most of them had given
- 30:44up, but not all of them. In 1986, three
- 30:47researchers, David Rumlhart, Doffrey
- 30:50Hinton, and Ronald Williams, published a
- 30:53paper that made a technique called back
- 30:55propagation famous. Remember how Minsky
- 30:59said adding more layers probably
- 31:01wouldn't help? The problem was that
- 31:03nobody knew how to train a network with
- 31:05lots of layers. Back propagation solved
- 31:08that. It's a way of working out exactly
- 31:10how much each of those dials contributed
- 31:13to a mistake layer by layer so you know
- 31:17how to nudge every single one. And
- 31:19suddenly the simple problem the
- 31:21perceptron couldn't solve like telling
- 31:23whether two dots were the same color
- 31:25became solvable. Hinton by the way will
- 31:28come up again a lot. Anyway, a year
- 31:31later, he left the US for the University
- 31:33of Toronto, partly because he didn't
- 31:35want his research funded by the American
- 31:37military. And Toronto became the hideout
- 31:40for the few people still betting on
- 31:42neural networks. And neural networks
- 31:45actually started working in the real
- 31:47world. At Bell Labs in the US, a young
- 31:49French researcher called Yan Lun built a
- 31:53neural network that could read
- 31:55handwritten numbers. By the late 1990s,
- 31:58banks were using it to read a
- 31:59significant share of all checks in the
- 32:01US. But most of the AI world still
- 32:04didn't take neural networks, aka the
- 32:06learning camp, seriously because in
- 32:081997, the rules camp got its biggest win
- 32:12ever. IBM's Deep Blue beat world chess
- 32:15champion Gary Kasparov. The big promise
- 32:17that had been made 40 years earlier,
- 32:21and the chess computer Deep Blue didn't
- 32:24learn anything. It was a brute force
- 32:26machine checking 200 million chess
- 32:28positions per second using rules written
- 32:31by humans. So at the end of the 1990s,
- 32:34it looked like the rules camp at one.
- 32:36But here's the twist. The learning camp
- 32:39hadn't been wrong. It had just been
- 32:41starved. Remember how chachip got past
- 32:43the sparsity problem? Billions of dials
- 32:45trained on trillions of words. Neural
- 32:47networks need two things to work. huge
- 32:50amounts of examples to learn from and
- 32:53huge amounts of computing power to nudge
- 32:55all those dials. And in the 1990s,
- 32:58neither existed. There was no internet
- 33:01full of text and photos to learn from.
- 33:03And computers were nowhere near fast
- 33:06enough. So neural networks were missing
- 33:08exactly the two things they needed, the
- 33:10data to learn from and the power to
- 33:11learn from it. But the idea behind
- 33:13today's AI was already there. It had
- 33:16been there since Rosenlat's perceptron
- 33:18in 1958. It just had to wait for the
- 33:20world to catch up. And when that finally
- 33:23happened, the missing computing power
- 33:25didn't come from a university or a
- 33:27supercomput. It came from video games.
- 33:30If you watched my chip video, you
- 33:32already know what a GPU is. If not,
- 33:35here's the quick version. Every computer
- 33:37has a CPU, a central processing unit.
- 33:40That's its brain, and it's good at doing
- 33:43one complicated thing at a time. But
- 33:45video games have a different problem. To
- 33:47draw a game on your screen, you have to
- 33:50calculate the color of millions of
- 33:51pixels many times per second. Each
- 33:54calculation is simple, but there are
- 33:56millions of them, and they all have to
- 33:58happen at once. So, companies like
- 34:00Nvidia built a different kind of chip
- 34:03for that, the GPU, the graphics
- 34:05processing unit. Instead of doing one
- 34:08complicated thing at a time, it does
- 34:10thousands of simple calculations at the
- 34:13same time. And here's the thing.
- 34:15Remember the dials? Training a neural
- 34:18network is basically billions of very
- 34:20simple calculations. Multiply, add,
- 34:23nudge, repeat. Exactly the kind of job a
- 34:26GPU was built for. A chip designed to
- 34:28make video games look pretty turned out
- 34:31to be the perfect machine for AI. And
- 34:34this is why Nvidia is such a valuable
- 34:36company. And I really recommend you go
- 34:39watch the video about chips I made if
- 34:41you really want to understand how these
- 34:43chips are made and why the US and China
- 34:46are fighting over them. But back to the
- 34:48story because there was one big problem.
- 34:51GPUs only understood graphics. By the
- 34:54early 2000s, some scientists had noticed
- 34:57that their work looked a lot like a
- 34:59video game problem. weather models,
- 35:01physics simulations, chemistry, all of
- 35:03it was millions of symbol calculations
- 35:06happening at once. And gaming GPUs were
- 35:09cheap and incredibly powerful at exactly
- 35:11that. But to use a video game GPU for
- 35:14science, researchers had to trick it.
- 35:17They had to disguise their maths as a
- 35:19picture, pretend their numbers were
- 35:22pixel colors, let the card draw them,
- 35:25and then read the image back out as
- 35:27results. It worked, but it was a
- 35:29nightmare. So, in 2006, Nvidia announced
- 35:32CUDA. It's software that lets
- 35:34programmers give a GPU any kind of
- 35:37calculation directly without dressing it
- 35:39up as graphics. And this was a huge
- 35:42gamble. Nvidia CEO Yansen Huang built
- 35:46CUDA support into its ordinary gaming
- 35:48GPUs. That made every chip more
- 35:50expensive for a market that barely
- 35:52existed. It was only a handful of
- 35:55scientists, oil companies, and banks.
- 35:57And for years that gamble didn't pay off
- 36:00and investors kept asking why a gaming
- 36:02company was spending fortunes on this.
- 36:05And then came 2012
- 36:08in Toronto. Draw Hinton the man who'
- 36:10kept neural networks alive through the
- 36:13winter had two students Alex Kresfki and
- 36:17Ilia Satsk.
- 36:19And they had an idea. What if you took a
- 36:22neural network, made it much bigger and
- 36:26deeper than anyone had dared before, and
- 36:29trained it on gaming graphics cards?
- 36:32Quick thing before we move on. There's
- 36:34no team behind this channel. The
- 36:36research, the writing, the filming, the
- 36:38editing, that's all just me, which is
- 36:42why I can only make about one video a
- 36:44month instead of one a week. And
- 36:47honestly, one deeply researched video a
- 36:50month doesn't pay nearly as well as
- 36:52pumping out quick videos every few days.
- 36:55The only reason I can keep making videos
- 36:57like this is my patron supporters. So, a
- 37:00huge thank you to all of you. And if
- 37:03you'd like to support the channel, too,
- 37:06the link is in the description. Anyway,
- 37:09let's get back to the video because we
- 37:12were talking about Joffrey Hinton, Alex
- 37:14Krosfki, and Ilascover
- 37:16and their idea. What if you took a
- 37:18neural network, made it much bigger and
- 37:21deeper than anyone had dared before, and
- 37:23drained it on gaming graphics cards?
- 37:26Well, Alex Kreski bought two Nvidia
- 37:30gaming cards around $500 each, put them
- 37:34into the computer in his bedroom, and
- 37:36wrote the code to make them train a
- 37:38neural network. It took about a week of
- 37:40the cards crunching numbers non-stop.
- 37:44But remember, neural networks don't just
- 37:46need computing power, they also need
- 37:48examples, lots of them. And that's where
- 37:50the second ingredient came in. A few
- 37:53years earlier, the Stanford professor
- 37:55Fifi Lee had built a giant collection of
- 37:58photos called Imageet. Over a million
- 38:01images, each labeled by hand with what's
- 38:03in it. Dog, car, strawberry, a thousand
- 38:06different categories. And all those
- 38:09labels, well, they were added by
- 38:11thousands of low paid online workers
- 38:13clicking through images one by one. And
- 38:15since 2010, there had been a yearly
- 38:17competition. whose program could
- 38:20recognize the most imageet photos
- 38:22correctly. Until then, the best programs
- 38:25were built the old way. Humans had
- 38:28carefully designed by hand what the
- 38:31computer should look for. Edges, shapes,
- 38:33textures, basically the rules camp
- 38:36approach. But in 2012, Alex Kresfk's
- 38:39network later named Alexet entered and
- 38:44it destroyed everyone. The best
- 38:47traditional program got about 26% of its
- 38:50answers wrong. Alex Neck got about 15%
- 38:53wrong. In a competition where the
- 38:56winners usually improved by a couple of
- 38:57percentage points, Alex Net nearly half
- 39:01the error rate. And nobody had told it
- 39:04what an edge or a shape was. It had
- 39:06learned all of that by itself from
- 39:08examples. Rosenblat's idea from 1958
- 39:11finally had what it needed, data and
- 39:14computing power. And big tech noticed
- 39:18immediately just a few weeks later in
- 39:21December 2012. Hinton Kresfki and
- 39:24Satskcover founded a company. It was
- 39:26called DNN Research. Three people, no
- 39:30product, no revenue, no business plan,
- 39:33just one paper. And they did something
- 39:36pretty wild. That December, the world's
- 39:38biggest AI conference was taking place
- 39:41in a casino hotel at Lake Tahoe, Nevada.
- 39:45And from this hotel room, Hinton run a
- 39:47secret auction by email. The biders
- 39:50Google, Microsoft, the Chinese tech
- 39:53giant BU, and a small London startup
- 39:56called Deep Mind. The bids kept climbing
- 39:59by millions at a time. And when it
- 40:02reached $44 million, Hinton stopped the
- 40:05auction and sold to Google. Not even
- 40:08necessarily to the highest possible
- 40:09bidder, who just chose Google. $44
- 40:13million for three researchers and one
- 40:15idea that most of the field had laughed
- 40:18at for decades. And this was a turning
- 40:20point. For 70 years, AI had mostly been
- 40:25something that happened at universities,
- 40:27funded by governments and the military.
- 40:30From now on, the best minds in AI would
- 40:33increasingly work for big tech. And
- 40:35remember that small London startup in
- 40:37the auction, Deep Mind. It was founded
- 40:40by a man called Deus Asabis, a former
- 40:43child chess prodigy who designed video
- 40:46games as a teenager and then got a PhD
- 40:49in neuroscience. His mission was as
- 40:52modest as the Dartmouth crowds. Solve
- 40:55intelligence. In 2013, Deep Mind showed
- 40:58off a program that taught itself to play
- 41:01old Atari video games. Nobody told it
- 41:04the rules. It just saw the pixels on
- 41:06screen and the score unlearned by trial
- 41:09and error. After a while, it was better
- 41:11than most humans. And just weeks later,
- 41:14in early 2014, Google bought DeepMind
- 41:17for around half a billion dollars. So
- 41:20now Google owned the Hinton team as well
- 41:22as DeepMind. A lot of the world's best
- 41:25AI researchers were now working for one
- 41:28company. And then in 2016, DeepMind went
- 41:32after the game that was considered AI's
- 41:34holy grail, Go. It's an ancient Chinese
- 41:38board game, and it has more possible
- 41:40board positions than there are atoms in
- 41:43the universe. So, you can't brute force
- 41:45it like Deep Blue did with chess. You
- 41:47need something closer to intuition.
- 41:50Deep Mind's program, Alph Go, played the
- 41:53world champion Lee Settle in Soul.
- 41:56According to Deep Mind, more than 200
- 41:58million people watched. And in the
- 42:01second game, Alph Go played a move so
- 42:03strange that commentators thought it was
- 42:05a mistake. Move 37. Lee Settle looked
- 42:10stunned and took almost 15 minutes to
- 42:13respond. It turned out to be brilliant,
- 42:16a move no human would have ever played.
- 42:19Alph Go ended up beating Lee Saddle by
- 42:224-1. And 3 years later, Lee had all
- 42:25retired from Professional Go, saying AI
- 42:28was an entity that cannot be defeated.
- 42:31So by 2016, Google had the best
- 42:34researchers, the best AI lab, and a
- 42:37mountain of GPUs. AI could now recognize
- 42:40images, and beat humans at games. But
- 42:42language was still a problem. Back then,
- 42:45AI read text the way you'd read through
- 42:47a keyhole, one word at a time, left to
- 42:50right. And by the end of a long
- 42:52sentence, it had often forgotten how the
- 42:54sentence started. It was also slow
- 42:56because reading one word after the other
- 42:59meant you couldn't use the GPU's big
- 43:01strength, doing thousands of things at
- 43:04the same time. And then in 2017, a group
- 43:08of eight researchers at Google published
- 43:10a paper with a slightly cheeky title,
- 43:13Attention Is All You Need. It's a nod to
- 43:16the Beatles song, All You Need Is Love.
- 43:19their idea. Stop reading one word at a
- 43:22time. Look at the whole text at once and
- 43:24for every word work out which other
- 43:26words matter most for understanding it.
- 43:29That's what attention means. Take the
- 43:32sentence. The cat didn't cross the road
- 43:34because it was too tired. What does it
- 43:37mean? The cat or the road? You know
- 43:41immediately it's the cat. Attention is
- 43:44how the model learns to make that same
- 43:45connection. When it reads it now, it
- 43:50pays the most attention to cat. And
- 43:52because it looks at everything at once,
- 43:54it could run perfectly on GPUs with
- 43:57thousands of calculations in parallel.
- 43:59So you could train it on way more text
- 44:01than ever before. They called it the
- 44:03transformer. And that's the T in chatbt,
- 44:07generative pre-trained transformer. It's
- 44:10the reason chach can look at your whole
- 44:12conversation at once, which we talked
- 44:15about at the start. The paper has since
- 44:18been cited well over a 100,000 times.
- 44:21It's one of the most important research
- 44:23papers of the century and was invented
- 44:25at Google. And what Google did next
- 44:28might be the biggest blunder in tech
- 44:31history. Because here's the irony.
- 44:34Google had invented the technology
- 44:37behind chat GBT, but it didn't release a
- 44:40chatbot. Google did end up using
- 44:43transformers. It quietly built them into
- 44:46Google search to better understand what
- 44:47people were typing. But releasing a
- 44:50chatbot that talks to you, that's where
- 44:53Google hesitated. Two Google engineers,
- 44:56Noam Shazair, one of the eight authors,
- 44:59and his colleague Daniel Defrighas did
- 45:02build a chatbot based on the
- 45:03transformer. The two Google employees
- 45:06pushed for years to release it to the
- 45:07public. Google said no. Officially, the
- 45:11worry was safety. What if it said
- 45:13something offensive or wrong under
- 45:15Google's name? But many people think
- 45:17there was another reason, too. Google
- 45:20makes most of its money from search ads.
- 45:23If a chatbot just gives you the answer,
- 45:25you don't click on any links. So, why
- 45:27would Google built the thing that could
- 45:29kill its own business? So, the authors
- 45:31of the most important AI paper of the
- 45:33century did what frustrated talent in
- 45:36Silicon Valley does. They left. One by
- 45:40one, all eight of them left Google. Most
- 45:43of them started their own companies.
- 45:46Aiden Gomez, who'd been an intern and
- 45:48stellar student when the paper was
- 45:50written, co-founded Coher. Leon Jones,
- 45:53moved to Tokyo and co-founded Sakana AI.
- 45:57And Noam Jazer co-founded Character AI,
- 46:00a chatbot company. And one of the eight,
- 46:03Lucas Kaiser, went to work for a small
- 46:06lab called Open AI. Before we get into
- 46:11open AI, I want to get back to one funny
- 46:13thing. Because in 2024, with Google
- 46:16desperately trying to catch up in the AI
- 46:18race, it paid around $2.7 billion in a
- 46:22deal with Character AI, largely to get
- 46:25Noam Shazer back. So, Google basically
- 46:28paid billions to rehire a man to work on
- 46:30a technology he had invented while
- 46:32working at Google. Okay, but let's get
- 46:35to open AI. What was this small lab that
- 46:38was called OpenAI? In the summer of
- 46:412015, Sam Oldman, who was running the
- 46:44famous startup accelerator Y Combinator
- 46:47at the time, hosted a private dinner at
- 46:49a hotel in Silicon Valley. Among the
- 46:52guests were Elon Musk, Greg Brockman,
- 46:55then the CTO of the payment company
- 46:57Stripe, and Ilia Satskever. Yes, the
- 47:01same Ilia Satskever from the AlexNet
- 47:03team and the $44 million auction. He was
- 47:07working at Google at the time and all of
- 47:10them shared one fear that Google with
- 47:13Hinton's team and Deep Mind was going to
- 47:17own the future of AI. One company
- 47:20controlling what might become the most
- 47:22powerful technology in history. So in
- 47:25December 2015, they launched the
- 47:27opposite, Open AI, a nonprofit. No
- 47:31shareholders, no profit motive. Its
- 47:33mission was to make sure that artificial
- 47:36general intelligence AI as smart as a
- 47:38human benefits all of humanity, and it
- 47:41would share its research openly, hence
- 47:44the name. And to show they were serious,
- 47:47they announced a funding pledge of $1
- 47:48billion. But a nonprofit with a mission
- 47:51to save humanity quickly ran into a very
- 47:53practical problem. Money. Remember what
- 47:56neural networks need? Huge amount of
- 47:59data and huge amounts of computing
- 48:01power. And computing power costs a
- 48:04fortune. And of that billion dollar
- 48:05pledge, much less actually came in. Then
- 48:08came a power struggle. In 2018, Elon
- 48:11Musk left. According to OpenAI, he'd
- 48:13wanted more control of the company, and
- 48:15when he didn't get it, he walked away.
- 48:17Musk tells the story differently and the
- 48:20two sides have been fighting over it in
- 48:22court. But nonetheless, OpenAI had money
- 48:25troubles. So in 2019, OpenAI did
- 48:28something that would have been
- 48:29unthinkable at the start. It created a
- 48:32for-profit company under the nonprofit
- 48:34with investors profits capped at 100
- 48:36times their investment.
- 48:39So if you put in $10 million, you could
- 48:42only make a billion.
- 48:44Very modest. It's a bit like going on a
- 48:48diet and limiting yourself to a 100
- 48:50donuts a day. But the donut diet did its
- 48:52job because now OpenAI could take money
- 48:55from investors and the first big one
- 48:57showed up fast. Just a few months later,
- 49:00Microsoft invested $1 billion. But let's
- 49:03jump back to an important question. Why
- 49:05did OpenAI suddenly need so much money?
- 49:08Well, it was because of one big bet.
- 49:10Inside OpenAI, the chief scientist, Ilia
- 49:13Satskev, had become convinced of
- 49:15something.
- 49:17Maybe AI didn't need some brilliant new
- 49:19idea. Maybe you just had to take the
- 49:21transformer and make it bigger. More
- 49:24dials, more data, more computing power.
- 49:28Satsk was so convinced that he
- 49:30reportedly led staff in a chant at
- 49:32company events. Feel the AGI. And in
- 49:352020, Open AI researchers published a
- 49:38paper that seemed to prove him right.
- 49:40It's called scaling laws for neural
- 49:42language models. And it showed that if
- 49:45you make the model bigger and give it
- 49:46more data and more computing power, it
- 49:49gets better predictably every single
- 49:52time. And you can see that bet in
- 49:55OpenAI's models. They call them GPT.
- 49:58GPD1 in 2018 had 117 million dials. It
- 50:03was trained on about 7,000 unpublished
- 50:06books, lots of them self-published
- 50:08romance novels. Also, none of the
- 50:11authors were ever asked. GPD2 in 2019
- 50:15had 1.5 billion dials and OpenAI
- 50:19announced it was too dangerous to
- 50:20release in full because it might be used
- 50:22to mass-produce fake news, which of
- 50:25course made headlines everywhere. Sound
- 50:28familiar? The Perceptrons press
- 50:30conference all over again. GPT3 in 2020
- 50:34had 175 billion dials trained on a huge
- 50:38chunk of the internet on a supercomput
- 50:41Microsoft build specifically for OpenAI
- 50:44with 10,000 Nvidia GPUs. So we went from
- 50:47117 million to 175 billion dials in 2
- 50:51years. That's around 1,500 times bigger.
- 50:55But there was still a problem. GBT3 was
- 50:58basically a giant autocomplete. Remember
- 51:01my paper cups? It was the same game just
- 51:04at a gigantic scale. So if you typed in
- 51:07a question, it didn't necessarily answer
- 51:09it. It just continued the text,
- 51:12sometimes with more questions as if it
- 51:14was writing an exam paper, which was
- 51:16obviously impressive, but not very
- 51:19useful. So OpenAI added one more
- 51:21ingredient, humans. People wrote example
- 51:24answers and ranked the model's responses
- 51:27from best to worst. And the model was
- 51:29trained to give more of the kind of
- 51:31answers humans liked. This is called
- 51:34reinforcement learning from human
- 51:36feedback or RLHF. And it's what turned
- 51:39autocomplete into an assistant. Who
- 51:41those humans were and how they were
- 51:43treated we'll get to later. But for now,
- 51:45let's look at the moment you probably
- 51:47first heard of chat GPT. The moment it
- 51:49went mainstream. On the 30th of October
- 51:522022, OpenAI released a chatbot built on
- 51:55GPT 3.5 as a free research preview. They
- 52:00called it Chat GPT. Sam Oldman announced
- 52:03it with a pretty casual tweet. Nobody
- 52:06inside OpenAI expected much. 5 days
- 52:08later, it had a million users. Within
- 52:10about 2 months, according to estimates,
- 52:13100 million. At the time, the fastest
- 52:15growing consumer app in history. And at
- 52:18Google, management reportedly declared a
- 52:20code red. The company that had invented
- 52:23the technology behind Chhat GPT now had
- 52:26to scramble to catch up. But let's jump
- 52:29back a little. Remember those humans who
- 52:32helped turn GPT into Chhat GPT? Let's
- 52:34talk about them because the story of
- 52:36Chhat GPT isn't just a story about
- 52:39genius researchers in San Francisco.
- 52:41Remember, GPT3 learned from a huge chunk
- 52:44of the internet. And the internet isn't
- 52:46just Wikipedia and recipes. It's also
- 52:48full of hate speech, violence, and
- 52:50abuse. And a model trained on all of it
- 52:53can repeat all of it. So before open AI
- 52:56could release a friendly chatbot, it
- 52:58needed a filter. A second AI that could
- 53:00spot toxic content and block it. But to
- 53:03build that filter, you need examples.
- 53:05Thousands of examples of the worst
- 53:07things human have ever written. Each one
- 53:10read and labeled by a person. So in late
- 53:132021, OpenAI hired a company called Sama
- 53:17which employed workers in Nairobi,
- 53:20Kenya. Sama marketed itself as an
- 53:23ethical AI company. The workers drop
- 53:26read text describing things like murder,
- 53:28suicide, torture, and child sexual abuse
- 53:31and label it so the filter could learn
- 53:33to recognize it. According to an
- 53:36investigation by Time magazine, they
- 53:37read between 150 and 250 of these
- 53:40passages per 9hour shift. And their
- 53:43take-home pay, according to Time,
- 53:45between roughly 1.32 and $2 an hour.
- 53:49Sama disputes those figures. Workers
- 53:52told time the work left them
- 53:53traumatized. One described as torture.
- 53:56In early 2022, Sama ended its work with
- 53:58Open AI months ahead of schedule. And
- 54:01that was just one drop. Remember RLHF,
- 54:05the humans rating chat's answers so it
- 54:08learns to be helpful? That's another
- 54:09layer of hidden labor. When OpenAI first
- 54:12developed this technique, it hired a few
- 54:14dozen contractors through outsourcing
- 54:17platforms like Upwork and a company
- 54:19called Scale AI to write example answers
- 54:22and rank the models responses. And today
- 54:25that work has turned into a huge global
- 54:27industry. Hundreds of thousands of
- 54:29people in places like Kenya, the
- 54:31Philippines, and Venezuela sit at home
- 54:33labeling images, writing chatbot
- 54:35answers, and correcting AI mistakes task
- 54:37by task, often for a few dollars an
- 54:40hour. The Washington Post reported on
- 54:41workers on Scale AI's platforms, Remo
- 54:44tasks in the Philippines, whose payments
- 54:46were delayed, reduced, or never arrived
- 54:49at all. And the thing is in 2025, Meta
- 54:52invested around $14 billion in scale AI.
- 54:54The company built on gig workers in the
- 54:57global south was suddenly worth almost
- 54:59$30 billion. So behind every intelligent
- 55:02machine, there's a huge invisible army
- 55:04of humans clicking, rating, correcting,
- 55:07making the AI look smarter than it
- 55:09actually is. Remember how tragic he read
- 55:11a huge chunk of the internet, websites,
- 55:13Wikipedia, Reddit, millions of books?
- 55:16Well, here's a question nobody really
- 55:18asked at the time. Did anyone ask
- 55:20permission? Mostly no. The first GPT was
- 55:24trained on thousands of self-published
- 55:26romance books whose authors were never
- 55:28asked. And it went on from there. News
- 55:30articles, blog posts, forum discussions,
- 55:33artwork, code. If it was online, it was
- 55:36fair game. The economist Cecilia Recap
- 55:39calls companies like OpenAI, Anthropic,
- 55:41and Google intellectual monopolies. The
- 55:44knowledge that millions of people
- 55:46created together for free gets swallowed
- 55:48up and turned into private products that
- 55:50a handful of companies own and sell back
- 55:53to us. And it wasn't just stuff that was
- 55:56freely online. According to court
- 55:58documents, MetaTrained its AI on books
- 56:00from a pirate library called Liben.
- 56:02Millions of copyrighted books downloaded
- 56:05illegally. And in a lawsuit against
- 56:07Anthropic, court documents showed it had
- 56:09also downloaded millions of pirated
- 56:12books. and it bought millions of
- 56:14physical books, cut off their spines and
- 56:16scanned every page. And the irony is
- 56:18since companies realized how valuable
- 56:20all this human writing was, the platform
- 56:23started charging for it. Reddit now gets
- 56:26paid by Google and OpenAI to let them
- 56:28train on Reddit posts. Posts written for
- 56:31free by Reddit users who get nothing.
- 56:35Okay, so AI is built on human labor and
- 56:38on human knowledge. But what does it do
- 56:40to the humans who have to work with it?
- 56:43And to explain that, there's a great
- 56:45idea from the writer Cory Dr. Row. And
- 56:47it starts with chess. Remember Deep Blue
- 56:50being Kasparov. Afterwards, Kasparov
- 56:54came up with a new kind of chess. A
- 56:56human and a computer playing together as
- 56:58a team. He called them centators like
- 57:01the creature from Greek mythology. Half
- 57:03human, half horse. The human makes the
- 57:06decisions and the machine gives them
- 57:08superpowers. That's how AI is usually
- 57:10sold to us. You stay in charge and AI
- 57:13makes you faster and better. But Dr. O
- 57:15says that for a lot of workers, it's
- 57:17turning out the other way round. He
- 57:19calls it a reverse centur. The machine
- 57:22is in charge and the human is just there
- 57:24to keep up with it, fix its mistakes,
- 57:26and take the blame when it goes wrong.
- 57:29According to a report in the New York
- 57:31Times, some programmers at Amazon say
- 57:33their jobs have changed. less writing
- 57:36code, more racing to check code written
- 57:38by AI under pressure to work faster and
- 57:40faster. And there's a quieter risk. In
- 57:422025, a study in Poland looked at
- 57:45doctors doing colonoscopies with an AI
- 57:47assistant that helped spot early signs
- 57:49of cancer. After a few months of getting
- 57:52used to the AI when the doctors worked
- 57:54without it, they spotted fewer warning
- 57:56signs than before. Their detection rate
- 57:59dropped from about 28% to 22%.
- 58:03So AI doesn't just need humans. It can
- 58:06quietly change what humans are for. From
- 58:09doing the work to babysitting the
- 58:11machine and what that means for society
- 58:13is what we'll explore in the next video
- 58:15of this series. Anyway, thanks for
- 58:18watching and see you soon. Bye.
About this transcript
This page contains the full transcript of not another AI video by Anna Bocca, generated from the public captions YouTube serves with the video. The transcript has 8,922 words across 1,336 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.