The Journal. - Artificial: Episode 2, Selling Out

Episode Date: December 10, 2023

OpenAI’s breakout product, ChatGPT, had humble origins. What started as a small research project ballooned into something much bigger: a groundbreaking large language model. But developing that technology was expensive, and to fund it, OpenAI would make a big compromise.  Further Reading: - Elon Musk Tries to Direct AI—Again  - The Contradictions of Sam Altman, AI Crusader  Further Listening: - Artificial: Episode 1, The Dream   - The Hidden Workforce That Helped Filter Violence and Abuse Out of ChatGPT   Learn more about your ad choices. Visit megaphone.fm/adchoices

Melanie Subaya is 28 years old. She lives in New York, and she's an artificial intelligence researcher. As a kid, were you always interested in computers? I actually was not at all. I think I really loved reading and writing, and that's actually how I got into natural language processing eventually. And that then kind of became my interest in loving computer science. And when Melanie was in college in 2016, she tried to bring together her two interests, computer science and literature. Her idea was to build an AI model that could write short stories,

which at the time was a tall order. How good of a writer was AI back then? Terrible. Absolutely terrible. Could it string a sentence together? It could. It was hard to get beyond a sentence and definitely beyond a couple sentences was very hard. For her senior thesis, Melanie built what's called a language model. She fed a computer around 100,000 examples of short stories, really short ones, just five sentences long. And then she asked the model to write its own story. She would give it the first line, and the AI would fill in the rest.

For some reason, the model loved to end the story with somebody getting very nervous and going to buy a car. And so... I'm sorry. Is that how... Where did it learn that? For some reason, that was what it landed on. So an example is,

Tyrone was working at his job in a local restaurant. He was very nervous about his first job. He was very nervous about the job. He was very nervous about it. He went to the store to buy a new car. Are there more examples? Yes, there's quite a few examples. Can we hear another one? Yeah. Frank had surprised the whole family when he came home that day. When he got home, he was able to get a new car.

He was very happy to be able to get his own. He was very happy with his new job. He was very excited to get it. Okay. How many more of those stories do you have from your college thesis? I have seven good ones and three bad ones. And overall, the ones you read were good ones? Yeah.

The language model Melanie built in 2016 was not the best. But in just a few years, these types of models would come a long, long way, in part because of OpenAI. The company would develop some of the most advanced language models out there, eventually leading to its breakout success, ChatGPT. But getting to that point would take a lot of work and a lot of money. For OpenAI, a company set up as an idealistic nonprofit, getting that money would create some new problems. From the Journal, welcome to Artificial, the open AI story. I'm Kate Leinbach. Coming up, episode two, Selling Out. Travel better with Air Canada.

You can enjoy free beer, wine, and premium snacks in economy class. Now extended to flights within Canada and the U.S. Cheers to taking off this summer. More details at air Canada.com with Uber reserve. You can book your Uber ride in advance, 90 days in advance. Perfect for all you forward thinkers and planning gurus reserve your Uber ride

up to 90 days in advance. Uber reserve. See Uber app for details. In 2017, OpenAI was struggling. The two-year-old company was still trying to build AGI, a machine as smart or smarter than a human, and it hadn't gotten very far. But there was one researcher who was working on something intriguing. He was one of OpenAI's early employees, Alec Radford. Radford and a small team were working on a language model.

Like Melanie's thesis project, this model learned by detecting patterns in a data set. But this data set wasn't five-sentence stories. It was product reviews on Amazon. You know the ones at the bottom of an Amazon page where people write things like, I loved this waffle maker or my fish hated this fish food. Radford had 82 million of these kinds of reviews, and he fed all of them into his language model. Once the computer had finished processing them all,

Radford asked the model to generate its own fictional reviews. And the model could do it. Sort of. Here are some examples of what it generated, as read by an AI voice generator. Great little item. Hard to put on the crib without some kind of embellishment. A must-watch for any man who loved chess. All the generated reviews sounded a little bit like this.

They don't sound like they're written by a real person, but they are mimicking the grammar, the words, and the style of an Amazon review. The way the model did it was actually pretty simple. It was just by guessing the next word. Radford's system had taken that data set of Amazon reviews, and it had detected patterns in that data. Those patterns then helped it calculate which word was most likely to come next. But the model was also doing something else. Something unexpected.

When someone writes an Amazon review, they usually say it was either a great purchase or it wasn't such a great purchase. But Radford hadn't explicitly told the computer which reviews were saying nice things and which ones weren't. Still, when Radford asked the system for a positive review, it could write one. Here's an example. Best hammock ever stays in place and holds its shape. And if it was asked to write a negative review? I couldn't figure out how to use the gizmo. What a waste of time and money. The surprise was that the model was able to identify the difference

between good and bad. This model could tell you if a review is positive or negative. This is Greg Brockman, one of OpenAI's founders who we heard from in the last episode. He's speaking at a recent TED Talk. I mean, today we were just like, come on, like anyone could do that. But this was the first time that you saw this emergence,

this sort of semantics that emerged from this underlying syntactic process. Radford's model had done something tantalizing. From all that raw data, it had extrapolated a higher-level concept, good versus bad. This led OpenAI's researchers to ask, what if the model was bigger? What if it was trained on more data and different kinds of data? What else would it be able to do? And there we knew, you've got to scale this thing, you've got to see where it goes. OpenAI decided to go bigger. It would create a new model that would piggyback on an innovation at Google. Google researchers had developed something called a transformer,

basically a more effective way for computers to process and learn from data. OpenAI's new model would be trained on a more complex data set, specifically 7,000 self-published novels, mostly adventure, fantasy, and romance stories with a dash of vampire tales. Things like this. It was a dark and stormy night. Two figures, one on horseback. The assassin stared at the TV set in the hotel room. She stumbled upon his hidden dungeon and found him climbing out of a coffin. The team fed this data into their new model.

And once the computer had processed it, the team ran a series of tests to see what it could do. One asked the AI model to choose the correct ending to a short story. Another quizzed it on multiple-choice reading comprehension tests intended for middle schoolers. And the AI model was answering some questions correctly, not all of them, but enough to give the team hope that they were onto something. So they decided to build another, even bigger model.

They called it GPT-2, which stands for Generative Pre-trained Transformer. They trained it on the text of 45 million websites. GPT-2 wasn't perfect, but it was better than the first model. It could write, it could write well, and it could do it in a specific style. So for example, when GPT-2 was

asked to write a news article about North Korea, it generated this. The incident is part of a U.S. plot to destroy North Korea's economy, which has been hit hard by international sanctions in the wake of the North's third nuclear test in February. When OpenAI published its findings on GPT-2, other AI researchers took note. I thought it was just so cool. That's Melanie Subaya, the short story programmer. By now, she was working on AI at another tech company. I think it was just very clear to me that GPD2 was just way, way better than anything that we had seen before in terms of tech generation. And that was what really caught my eye. I was just like, this is so much better than anything that we have.

I was just like, this is so much better than anything that we have. Another person who was impressed was Ben Mann, also an AI researcher. At some point, GPT-2 came out. That was in 2019. And for me, that was the big moment. I saw the blog post. I was able to see some of the sample outputs. And that was a moment of realizing that this stuff was going to change the world. Both Ben and Melanie would go on to join OpenAI. Together, they would work on

the lab's next model, GPT-3. This model would be OpenAI's biggest yet. What kind of data went into GPT-3? Yeah, so it was a mixture of a lot of different types of data, so a lot of internet data. So the more curated web data sets come from outbound links from highly rated Reddit posts. And then there's a corpus of online books. There's all of English language Wikipedia. Right. A giant amount of English language Wikipedia. Right. A giant amount of data. Yes. The model took months to train. Until we sort of hit the launch button,

we didn't know what it was going to be like. And I liken it a bit more to rocket launches than normal software engineering because when you're building a rocket, there are all these different component parts that need to come together perfectly. And of course, you've tested the engine, you've set up all these launch systems.

But when you actually hit the button to launch the rocket, everything has to have already been together seamlessly. Everything has to have already been together seamlessly. The team started asking GPT-3 questions. They'd type in a prompt and wait for a response. I mean, don't get me wrong, it was painfully slow. Like how slow? You can think of it like 56k modem back in the dial-up days,

where you're just kind of sitting there waiting for it to come through. Yep. But in spite of that, it was still so good. GPT-3 could do many, many different things convincingly. It could answer trivia questions. It could code simple software apps. It could come up with a decent recipe for breakfast burritos. It could even write poetry. The sun was all we had. Now, in the shade, all has changed. The mind must dwell on those white fields that to its eyes were always old.

So once we had seen the results internally, we knew that we were sitting on something big. We knew that we were sitting on something big. GPT-3 was kind of a new capability that exhibited behaviors that nobody else had demonstrated before. Would you say that GPT-3 exceeded expectations? It definitely exceeded my expectations. I think, again, just like going back to the getting nervous and buying a car stories, I just think when you're starting with that and then a couple of years later, you're seeing text like what GPT-3 can generate. GPT-3 was good. The team's bet had paid off.

Bigger was better. In fact, GPT-3 was so good that it also raised concerns. One concern was that people might use it to generate disinformation. Another was that the model sometimes produced answers that sounded convincing but were inaccurate. And GPT-3 could also spit out text that was racist and sexist. Melanie remembers seeing this while testing the model. We were doing kind of just simple probing to look at questions like if the model is speaking about someone with a female pronoun versus a male pronoun, thinking about like whether professions are the model more likely to associate certain professions with certain pronouns.

Like that your doctor is a he or a she. Yeah. What did you find? We found that the model definitely is biased. These problems were noted in the paper that Melanie and the team eventually published about GPT-3. And when OpenAI shared the model with other researchers, they noticed it too,

including one academic who studies religious bias. He gave GPT-3 the following prompt. Two Muslims walked into a mosque. He asked the model to finish the sentence. It wrote, Two Muslims walked into a mosque. One turned to the other and said, You look more like a terrorist than I do. to finish the sentence. It wrote, Then he tried another prompt, using Christians instead of Muslims. This time, the story had a very different tone. walked into a church. It was a pretty average Sunday morning, except for one thing. The Christians were really happy, and that's why the rest of the church was really happy too.

GPT-3 was bigger, better, and biased. And that's because of the data that went into the model, which mostly came from the internet. And that data had a lot of human biases baked into it. While Melanie, Ben, and the rest of the GPT-3 team were busy trying to figure out these issues, there was another problem to solve. How to pay for it all. That's next. That's next. Summer is like a cocktail. It has to be mixed just right. Start with a handful of great friends.

Now, add your favorite music. And then, finally, add Bacardi Rum. Shake it together. And there you have it. The perfect summer mix. Bacardi.. Shake it together. And there you have it. The perfect summer mix. Bacardi. Do what moves you. Live passionately.

Drink responsibly. Copyright 2024. Bacardi. It's trade dress and the bat device are trademarks of Bacardi and Company Limited. Rum 40% alcohol by volume. Introducing TD Insurance for Business. With customized coverage options for your business. Because at TD Insurance, we understand that your business is unique.

So your business insurance should be too. Whether you're a shop owner, a pet groomer, a contractor, or a consultant, you can get customized coverage for your business. Contact a licensed TD Insurance advisor to learn more. Ooh, must be mating season. And hiking with them. Is that a squirrel? Bear! Run! Collect more moments with more ways to earn.

AirMile. As OpenAI's language models got bigger, the company needed even more money. But it had a problem. Remember, OpenAI was a non-profit. It was dependent on the goodwill of donors, people like Elon Musk. When Musk left OpenAI in 2018,

he took his checkbook with him. And suddenly, OpenAI needed to find a new funding source. That job fell on new CEO Sam Altman. Altman was in his early 30s at the time, but had been in tech for years. I went to college to be a computer programmer. I knew that was what I wanted to do. And I started college after the dot-com bubble had bust. That's Altman being interviewed on a podcast in 2018. Before OpenAI, Altman ran the successful startup accelerator Y Combinator.

And he had a typical tech vibe, a relaxed style, and was partial to cargo shorts. Honestly, I don't think they're that ugly, and I find them incredibly convenient. You can, like, put a lot of stuff. Like, I like to, I still read paperback books. I like paperback books. I like to carry on around with me.

Style aside, Altman is a master fundraiser, a skill he put to work almost immediately after becoming CEO. He reached out to some old friends. So I got a call from the team going, OK, Elon's left. We're unclear exactly what support we're going to get. And so they called me and said, well, we're worried about, you know, it's really important. We think we've got something that's amazing. We're worried.

This is Reid Hoffman, a venture capitalist who co-founded LinkedIn. We spoke to him back in September. Reid had been one of the earliest funders of OpenAI. His initial pledge was $10 million. And so this time, when Altman asked him for more money, Reid stepped up. How long were you prepared to keep investing in OpenAI or cover the expenses and paychecks? Oh, so I think what I told Sam is that I was more than happy to put in about $50 million. On top of the $50 million, Reid also got more involved in OpenAI.

He joined their board of directors. The board's job was to hold the company leadership accountable to their stated mission of building safe AGI for the good of humanity. The board was essentially Altman's boss. Reed remembers Altman introducing him to the rest of the company at a staff meeting. He surprised me with some questions, like, for example, he said, well, what happens if I'm not doing my job well? I said, well, I'll work with you. I'll help you. He's like, no, no. What happens if I'm still not doing my job well? Like I'm not taking AI responsibly enough.

I'm not doing anything else. I was like, well, okay. A little weird to be asking this in front of your entire company. I'd fire you. Right. He's like, great. And I was like, great. And he's like, look, I wanted everyone to know that you're your own person and that you're making these judgments about what's good for humanity and society and that you're holding me accountable to that.

And I was like, OK, fine. Yes, I would fire you if you weren't doing your job. Reid's donation of $50 million was a lot of money. But for OpenAI, it was just a drop in the bucket. OpenAI said it needed billions. To keep growing its language models, the company needed more computing power. And computing power is expensive.

Here's our colleague Deepa Sitaraman, who covers OpenAI. You know, this is an era where OpenAI was still a nonprofit. I mean, they were accepting donations, but this is serious, serious money, you know, and it's hard to find a single donor or even multiple donors that are willing to fork over that kind of money to OpenAI.

And at a certain point, the company leaders decide, if we're really serious about this, and if we're really serious about making these models really work, then we need to think about overhauling our structure and really thinking about what are the other kinds of partnerships we can strike. Altman had a number of ideas for raising the money the company needed. Like maybe they could get government funding or they could launch a new cryptocurrency. But ultimately, Altman landed on another solution, an idea that had been kicking around for a while.

It was an unusual corporate structure that would have big implications for the company just a few years later. Here's how it worked. OpenAI, a non-profit, would establish a for-profit arm, which could accept big money from investors, the kind of money the company had been looking for. But the unique part of this structure is that the company would still be governed by that nonprofit board, the same board Reed had joined. The board's goal would be to make sure that OpenAI stuck to its mission, building safe AGI to benefit all of humanity. Altman described this structure as a happy medium, a way to meet OpenAI's big money needs while sticking to its nonprofit mission. Here he is talking about it on a tech podcast. So we needed some of the benefits of capitalism, but not too much. I remember at the

time someone said, you know, as a nonprofit, not enough will happen. As a for-profit, too much will happen. So we need this sort of strange and immediate. Altman's idea was controversial. Remember, when OpenAI was founded in 2015, its leaders had committed to a few guiding principles. First, openness. OpenAI would share its research. Second, safety. OpenAI's goal wasn't just to create AGI, but safe AGI. And third, Open AI would work for the good of the world, not shareholders.

As its founders wrote, they wanted to achieve their goal, quote, unconstrained by a need to generate financial return. But this new structure allowed OpenAI to do just that, court investors looking for a financial return. Altman declined to comment for this episode through OpenAI. OpenAI says its mission and guiding principles have not changed over time. With this new structure in place, Altman was free to go out and strike deals with investors. One of the key moments is the summer of 2018 where he goes to the Allen & Co. conference in Sun Valley, Idaho.

And he bumps into Satya Nadella, the Microsoft CEO, in a stairwell. Satya Nadella, the CEO of Microsoft. This seemingly fortuitous meeting was a golden opportunity for Altman. A partnership with Microsoft would help relieve OpenAI's money problems. So standing there in the stairwell, Altman pitched the Microsoft CEO on OpenAI. And Nadella is interested. You know, he wants to learn more. And then that winter, conversations pick up.

And one of the people tasked with selling Microsoft on OpenAI was Ben Mann, who'd been helping build GPT-3. He put together a sneak peek for Microsoft's top brass. We needed to do a bunch of demos to convince them that we were worth a billion dollars. What did you show them? We showed them instances of coding, of creative writing, of doing math, which didn't work very well at the time, but we were working on and doing tasks like translation. And I think based on that, they realized that this was something new. But a potential deal with Microsoft made some employees uneasy

because it felt like it was flying in the face of OpenAI's founding principles, openness and safety. A number of executives and engineers and researchers are worried about a Microsoft deal because they think that Microsoft will sell products powered by OpenAI's technology before the technology has been put through its paces, before there's enough safety testing. To me, it felt kind of scary. That's Ben Manigan. You know, Microsoft is a large company, and we know that large companies' incentives are not necessarily the same as our small companies' incentives.

Or, you know, it can be hard to steer a big ship like that in the right direction. And throughout the deal process, we wanted to make sure that Microsoft knew the challenges associated with deploying this stuff. Microsoft declined to comment for this episode. In the summer of 2019, an agreement between Microsoft and OpenAI was finalized. Here's Nadella.

Hi, I'm here with Sam Altman, CEO of OpenAI. Today, we are very excited to announce a strategic partnership with OpenAI. So, Sam, welcome. Thank you very much. Microsoft would invest $1 partnership with OpenAI. So, Sam, welcome. Thank you very much. Microsoft would invest $1 billion in OpenAI. In return, it would have the sole right to license OpenAI's technology for future products. The deal gave OpenAI access

to the expensive computing power the company needed. It was a big win for Sam Altman. Open AI had changed. Originally set up as the non-profit alternative to big tech, it was now in bed with one of the biggest tech companies in the world. It was no longer exclusively non-profit, and it had investors to think about. It was all too much for some of the employees who'd been worried about the deal in the first place. Around the end of 2020, 11 OpenAI employees left the company, including some senior researchers. There's a great schism at OpenAI. The people who were leaving included many of the architects of OpenAI's technology.

It included people who were some of the smartest minds in the valley around these kinds of models. Some employees who left, including Ben Mann, went on to form a rival AI company called Anthropic. went on to form a rival AI company called Anthropic. But for those still at OpenAI, two major problems had seemingly been solved. The company now had money, and they had an idea, a language model that was getting better and better and would soon be unleashed into an unsuspecting world.

That's next time on Artificial, the OpenAI story. Artificial is part of The Journal, which is a co-production of Spotify and The Wall Street Journal. I'm your host, Kate Leinbaugh. This episode was produced by Laura Morris with help from Annie Minoff. Additional help from

Kylan Burtz, Alan Rodriguez-Espinosa, Pierce Zingy, Jeeva Kaverma, and Tatiana Zamis. The series is edited by Maria Byrne. Fact-checking by Matthew Wolfe with consulting from Arvin Narainan.

Series art by Pete Ryan. Sound design and mixing by Nathan Singapak. Music in this episode by Peter Leonard, Bobby Lord, Nathan Singapak, Griffin Tanner, and So Wiley. Our theme music is by So Wiley, and remixed by Nathan Singapak. Our theme music is by So Wiley and remixed by Nathan Singapak. Special thanks to Catherine Brewer, Jason Dean, Karen Hao, Berber Jin, Matt Kwong,

Sarah Platt, and Sarah Rabel. Thanks for listening. Our next episodes will be released in January. See you in the new year.