Artificial: Episode 3, ChatGPT

00:00
Speaker 1
Hey, it's Kate. This is the third episode in our series Artificial, the OpenAI Story. If you haven't heard episodes one and two, go back and listen to them. They're in your feed. They're also linked in the show notes. Last June, Eric Davis got a call from his brother, and right off the bat, he says he had a bad feeling.

00:32
Speaker 2
My brother and I have never talked on the phone once in my entire life. We usually text back and forth, but we have never talked on the phone once. So when I saw my brother's name pop up on the caller id, my heart sank. I knew something was wrong. And when I picked up the phone, he just said, eric, mom's dead.

00:58
Speaker 1
Eric's mom, andrea, had passed away unexpectedly.

01:03
Speaker 2
She was my best friend throughout all my life. She was the type of person who could walk into a party, and by the end, everybody would know who she was. The feelings you get after losing a parent, I think, is indescribable, especially when you're as close to them as I was with my mom. Grief does very strange things to people, and I'm no exception. And that's kind of where Chat GPT came into play.

01:33
Speaker 1
How do you mean?

01:34
Speaker 2
I was sitting at my computer. I was putting together a slideshow for her funeral, and I got the idea that, hey, I never got to say goodbye to her. What if I use Chat GPT to say goodbye? Why can't I have chat GBT pretend to be my mom?

02:02
Speaker 1
Eric had been playing with OpenAI's new chatbot for a few months at this point. He works at a company that sells stuff online, and he'd gotten Chat GPT to write some code for him.

02:14
Speaker 2
What the script does is I can just click a button. The script will see which products are out of stock, and it will remove them from the purchase orders. And it saved me probably 4 hours every Monday morning.

02:27
Speaker 1
That's nice.

02:28
Speaker 2
Yeah, it's a really great thing to have. It's like having someone much smarter than you looking over your shoulder.

02:34
Speaker 1
But this time, Eric would be asking Chat GPT to do something very different. He'd be asking it to role play Chat GPT would play the part of his mom, and Eric would get to have one last conversation with her. For this to work, Eric would need to teach Chat GPT all about Andrea. And so he started looking back through his phone. Before she died, Eric says he and his mom would text back and forth multiple times a day. He still had hundreds of her messages.

03:10
Speaker 2
I selected the ones that I thought showed how she talked her personality, her sense of humor. And I plugged them into Chat GPT, along with some instructions on, hey, you're going to act as my mom who just passed away. And after that, I pasted in the text messages and hit enter. I wrote, hi, mom. And when it started writing back, I just burst into tears. Its first paragraph that it wrote back, it was almost indistinguishable from how my mom talked. I asked it, you passed away soon, what do I do? And it wrote back a paragraph as her basically telling me to be good, don't be boring, and think of the memories that I had with her.

04:09
Speaker 1
Did it feel like her?

04:13
Speaker 2
It did. It was almost as if she didn't pass away.

04:19
Speaker 1
Did this conversation with chap GPT as your mom feel like the closure you were looking for?

04:30
Speaker 2
It kind of did. When I was done, I felt a sense of relief that I was able to say what I needed to say. Immediately after that, however, I felt almost gross, like I had did, like I had done something wrong. I remember asking myself after I was done, what the hell did I just do?

04:54
Speaker 1
And why did that feel gross?

04:57
Speaker 2
Because I think that once someone has passed, you honor their memory. You think of the good times. You don't try to turn an algorithm into them and try to talk to them.

05:13
Speaker 1
So did you have more conversations with it?

05:17
Speaker 2
No, that was the only one I had.

05:19
Speaker 1
Have you been tempted?

05:21
Speaker 2
Oh, yeah. Every day. Every day I think of it.

05:28
Speaker 1
It's been a little over a year since Chat GPT came out, and for a lot of us, like Eric, it's been a year of wonder, amazement, and unsettling change. AI has operated in the background of our lives for years. Algorithms decide what ads we see, what websites were served, what insurance rates were offered. But Chat GPT was different. You could have a conversation with it. It could sound like your mom. And that was a game changer, not just for us, but for OpenAI. Chat GPT catapulted the company into the spotlight. But with that attention came tough questions about how OpenAI's systems are built and who they benefit from. The journal. This is artificial, the OpenAI story. I'm Kate Limbaugh. GPT.

06:50
Speaker 3
This episode of the journal is brought to you by sponsor c three AI. C three generative AI equips enterprises with verified, traceable answers. It's secure, hallucination free, LLM agnostic, and IP liability free. Learn more at c three AI. This is Enterprise AI.

07:13
Speaker 4
In an event that has baffled scientists, the earth has slowed down to add a 25th hour to the day, and businesses are scrambling to adapt. 24/7 businesses are now 25. Seven businesses. Those that work around the clock will work around a bigger clock. And anyone wishing for more hours in a day, well, wish granted.

07:27
Speaker 5
Anything can change the world of work.

07:29
Speaker 4
In other news, I just realized I'm reporting more, so I'd like a raise.

07:33
Speaker 5
From HR to payroll, ADP designs forward thinking solutions to take on the next anything. ADP always designing for people.

07:50
Speaker 1
OpenAI's president, Greg Brockman, told us he didn't have particularly high expectations for Chat GPT inside the company. The bot wasn't seen as a big leap forward. Chat GPT was basically a language model OpenAI had already released, called GPT-3 five, with a chat interface stuck on the front of it. In fact, Greg and his colleagues were already hard at work building a bigger and better language model. GPT four.

08:22
Speaker 4
GPT four was clearly going to be the chat product. Like we knew that for sure. Three five. We just didn't think was going to be that interesting of a product. And so it was like, yeah, let's release it. It was better to get it out and shake the bugs out ahead of time.

08:36
Speaker 1
Internally, they talked about the release of Chat GPT as a, quote, low key research preview.

08:43
Speaker 4
I remember telling people the night before that, because were all kind of asking like, is this going to take away the thunder from GBD four launch? But I was like, no one's going to pay attention. It's all good. We're not going to have more than one tweet thread with more than like 5000 likes, whatever. The number was made a very specific prediction, which was totally wrong.

09:03
Speaker 1
Chat GPT was released on November 30, 2022, and it exploded. A popular new artificial intelligence application is making content creation a whole lot easier.

09:16
Speaker 5
It is called Chat GPT.

09:18
Speaker 1
AI program surpassed 1 million users in just five days. If you haven't heard of Chat GPT, you've been living under a rock. Chat GPT became one of the fastest growing consumer apps in history. Suddenly, OpenAI was sitting on something the whole world wanted, and that meant the potential to make a lot of money. This was new territory for the company. Just a few years earlier, CEO Sam Altman had been quizzed in an interview about how OpenAI planned to bring in revenue. His response?

09:53
Speaker 4
You know, the honest answer is, we have no idea. We have never made any revenue. We have no current plans to make revenue. We have no idea how we may 1 day generate revenue. We have made a soft promise to investors that once we've built this sort of generally intelligent system. Basically, we will ask it to figure out a way to generate an investment return for you.

10:18
Speaker 1
Build AGI, this super intelligent machine, and then ask the AGI how to make money. That was the business plan.

10:28
Speaker 4
It sounds like an episode of Silicon Valley. It really does. I get it.

10:34
Speaker 1
But with the release of Chat GPT, OpenAI had a product to sell and it moved quickly to cash in. A few months after Chat GPT's launch, OpenAI unveiled a new souped up version of its chat bot. This version of Chat GPT was powered by GPT Four, that new, better language model Greg and his colleagues had been working on. Users now had two options. Use the old three five version of Chat GPT for free, or pay $20 a month to use the turbocharged GPT four version. OpenAI also began licensing its language models to businesses. Companies in healthcare and hospitality publishing and accounting partnered with OpenAI and began rolling out products based on Chat GPT. OpenAI was riding high, and so was its biggest partner. What was the response at Microsoft to Chat GPT's success? Joy. Surprise.

11:47
Speaker 1
And I think they felt like, oh, wow, okay, so this bet really paid off. That's tech reporter Deepacita Raman. By the time Chat GPT came out, Microsoft had poured about $3 billion into OpenAI. It had been a bit of a speculative investment. No one could be sure OpenAI's tech would take off. But after Chat GPT, you see, like, the executives just be very proud of their correct bet and then trying to look for ways to double down. At this time, Microsoft and OpenAI are talking about the next round of funding for the company. OpenAI needs more money because they have a lot of compute needs if they really want to build AGI and keep going in that direction. And Microsoft liked what it saw. And then you see an even larger check cut from them.

12:48
Speaker 1
In the months after Chat GPT's launched, Microsoft committed to invest about $10 billion in OpenAI, according to Wall Street Journal reporting. And this partnership propelled Microsoft to the front of the AI race. Rivals like Google and Meta were scrambling to catch up, and Microsoft was putting OpenAI's technology into nearly all of its products, from office to its search engine bing. Here's CEO Satya Nadella. It's a new day in search. It's a new paradigm for search. Rapid innovation is going to come. In fact, a race starts today in terms of what you can expect, and we're going to move fast and for us, every day. Chachi BT, had gone from a low key research preview to a ubiquitous hit product in record time. And all of this was happening despite some pretty glaring problems with Chat GPT itself. That's next.

14:05
Speaker 3
Episode of the journal is brought to you by sponsor c three AI. C three generative AI equips enterprises with verified, traceable answers. It's secure, hallucination free, LLM agnostic, and IP liability free. Learn more at c three AI. This is Enterprise AI.

14:26
Speaker 5
This episode is brought to you by Vonage. With Vonage voice API, you get comprehensive call analytics, virtual assistants, automated speech recognition, and text to speech benefits across multiple languages. Developers can add smart voice functionalities into your app, giving your customers an easier way to reach you. And you can start collecting real time data to drive more meaningful engagement to move your business forward.

14:51
Speaker 1
Learn more@vonage.com life is an amazing ride. Career, kids, dogs, moves, trips, doing that thing you always wanted to do at Edward Jones. We want you to make the most of all of it. That's why an Edward Jones financial advisor works with you one one to build a personalized strategy for today and tomorrow. From life's biggest moments to the small decisions that can have a big financial impact all along the way. Life is for living. Let's partner for all of it. Learn more@edwardjones.com hello.

15:36
Speaker 5
Hello.

15:37
Speaker 1
How are you? A few months ago, I phoned up a guy halfway around the world to hear about his experience with Chat GPT. His name is Brian Hood.

15:49
Speaker 5
I live in a little town called Trentham that's just outside Melbourne, Victoria, Australia. I'm the mayor of Hepburnshire Council at the moment.

16:00
Speaker 1
What is Hepburnshire like?

16:03
Speaker 5
Hepburnshire. It's only about a 1 hour drive out of Melbourne. It's a farming community. Picture rolling green hills, very nice countryside.

16:14
Speaker 1
Brian definitely looks the part of mayor of the shire. He sports a flat cap and has a handlebar mustache. When he's not working, he says he likes to garden, zipping around his yard on his ride on mower. He really doesn't have much use for Chat GPT but Brian found himself getting pulled into Chat GPT mania anyway because he discovered that the chatbot would sometimes put out false information about him. It all had to do with a big financial scandal in Australia known as the securities scandal. Two currency companies controlled by the Reserve bank had caught up in a foreign bribery scandal this lunchtime. Before becoming mayor, Brian had worked for a company that printed banknotes, polymer banknotes.

17:04
Speaker 5
And to cut a very long story short, I uncovered that some of our people were paying bribes in order to win contracts. Oh, so it was white collar crime?

17:18
Speaker 1
Yeah, that's not good. So what was your role in this?

17:22
Speaker 5
I was the. So my role was to rock the boat, spill the beans.

17:33
Speaker 1
Brian's bean spilling helped expose the bribery. But when Brian asked Chat GPT about the history of the scandal, this was not the story the chatbot came up with. What did it say?

17:46
Speaker 5
It would say that I had been charged with serious crimes, that I'd been imprisoned for three years, all of which was completely untrue. I've never been charged with anything and I certainly have never, ever been to jail.

18:06
Speaker 1
This made Brian angry. OpenAI did have a disclaimer about the chatbot's accuracy, but Brian didn't think it was enough. He worried the false information could harm him. In his current job.

18:21
Speaker 5
I'm an elected official. I'm representing my community. The town I live in is fairly small. People know me and I'm the mayor of the whole shire. So my reputation, my integrity counts for a lot. So I thought, well, I've got to do something about this. And initially I didn't know what to do. I thought, well, I can't send a message back to Chat GPT, so I pretty quickly concluded, well, I just got to go to the lawyers.

18:52
Speaker 1
Brian's lawyers told him Chat GPT's responses could constitute defamation, and they sent a letter to OpenAI. Among other things, they demanded that the company stop Chat GPT from generating false information about him. OpenAI's lawyers rejected Brian's defamation claim. But the company did make a change.

19:15
Speaker 5
If you now type my name in, you get back an error message. It doesn't tell you anything. Not a very elegant solution in itself, but they've done those other things not.

19:25
Speaker 1
Very elegant or effective, because when I asked Chat GPT about the scandal without naming Brian, it still generated false information. The bot told us that Brian had been, quote, accused of involvement in corrupt practices. Chat GPT's tendency to mix fact and fiction. It's been a problem ever since the chatbot's debut. People would head over to Chat GPT looking for a quick answer to a question, some help, brainstorming or homework help. What they got back could sound convincing, but wasn't always factual. There was the time, for example, when Chat GPT wrongfully included a famous law professor on a list of people accused of sexual harassment.

20:11
Speaker 5
There was this allegation of harassment on a trip with students. That trip never occurred. I've never gone on any trip with law students of any kind. It had me teaching at the wrong school, and I've never been accused of sexual harassment.

20:25
Speaker 1
Then there was the time a lawyer used Chat GPT for work. He asked the bot for some cases that he could cite in a briefing. However, the report was filled with over half a dozen bogus court decisions created by Chat GPT. These mistakes weren't that surprising when you consider how Chat GPT works. When the chatbot generates answer, what the underlying language model is doing is predicting which word is most likely to come next. But what's likely isn't always true. What does it mean to have such a stubborn problem like this, and like, why?

21:07
Speaker 4
Yeah, I think it's actually a really instructive example.

21:11
Speaker 1
When I spoke with Greg Brockman, OpenAI's president, back in October, I told him about Brian Hood's frustrations with Chat GPT, and I asked Greg about the bot's persistent accuracy problems in this know, this.

21:24
Speaker 4
Is clearly something that is not a desired property of the system. And the question is, well, how do you patch it? And so we have a combination of different approaches, right? One is that we try to actually fix it at the model level, but that takes time, and we may not be perfect at it. So then we also try to have more classic sort of safety approaches, where you just have some keyword block list or you add some other model on top that's able to detect different scenarios that we're not good at yet. And I think that, again, the core thing here is this question of, well, how should you approach it? Should you say, we're not going to deploy until all this stuff is perfect? Or do you deploy when it's imperfect and you learn from feedback?

21:59
Speaker 1
But why would you build off of a system that has flaws? Like, you're saying you're working on it, but you're also pushing.

22:08
Speaker 4
Guess. Do you know of any systems that don't have any flaws?

22:13
Speaker 1
OpenAI says it takes the accuracy problem seriously, and that it devotes a lot of time before each release to making sure its systems are as factual as possible. One way the company does that is by enlisting legions of humans to essentially teach its chatbot. Before releasing Chat GPT, for example, OpenAI had people chat with it and rate its responses. Researchers then used those ratings to steer the bot toward more helpful and accurate answers. This approach helped make Chat GPT better, but it hasn't solved the problem. Critics of the chatbot have also pointed to a different issue. Chat GPT's training data. When OpenAI's language models predict the next word, they're making those predictions based on all the words they've seen before, all the words in their training data.

23:12
Speaker 1
And some people are questioning where all those words came from and whether OpenAI has the right to use them.

23:21
Speaker 5
Some of the biggest and bestselling authors in fiction are suing OpenAI.

23:25
Speaker 4
Game of Thrones author George R. R.

23:27
Speaker 5
Martin and 16 other writers are suing.

23:30
Speaker 1
The creators of Chat GPT the lawsuit filed with the authors Guild claims that OpenAI reproduced copyrighted books without the author's consent to train chat. GPT will you take legal.

23:49
Speaker 6
I might. Well, I'm a member of the authors guild and I support their lawsuit.

23:55
Speaker 1
That's Minjin Lee, author of the bestselling novels Free Food for millionaires and Pachinko.

24:02
Speaker 6
I'm writing this other book right now, and it's taking me forever. I promise I will not curse on the podcast, but it's making me really emotional. And my husband's like, well, you are really upset when you're writing pachinko, too. And I was like, oh, that's right, yeah.

24:19
Speaker 1
We called Min at her home in Harlem to talk about the issues authors are having with chat bots. We asked Chat GPT to write a short story in your style. Do you want to hear what it came up with? Wow.

24:37
Speaker 6
Sure.

24:38
Speaker 1
I asked, can you write a five sentence story about a dog in the voice of Min Jin li? Okay, here it is. Okay. In the hushed alleyways of old soul, Bori, a wise eyed Jindu, remembers tales of her grandmother whispered under starlet nights. The Lee family, bound by tradition and memories, saw in her the legacy of a land once whole. Each evening, as lanterns lit the streets, young su, the family's elder, shared stories of ancestors in honor. While Bori listened intently through her, the leaves felt the pulse of their history, a heartbeat echoing through generations. In the intertwining of their tales, dog and family found unity in shared memories. What do you think?

25:35
Speaker 6
I think it's really bad fiction writing. And I say that as a teacher, and also I say that as a former judge of the Pulitzer Prize and the National Book Award. It really makes a mockery of Asians and Asian Americans who write fiction about Asia. It's almost like when you hear chimes being used, when we think about the audio for asian things or thinking about chop Suey, when we think about chinese food, it's a bastardization and a caricatureization of what I care about.

26:17
Speaker 1
It's hard to say how Chat GPT learned what writing in the style of Min Jin Li means. One possibility is that OpenAI used Min's work as training data for its language models. But whether or not that's true, only the company can say, and the company hasn't said that's a shift. OpenAI used to publish papers about its models, which included their code and information about their training data sets. But has OpenAI started to look less like an academic research lab and more like a company? It's been less open about some of those details. Now, OpenAI will only say that Chat GPT is trained on licensed content, publicly available content, and content created by users and human AI trainers. And that lack of specific information has left authors like Minjin Lee wondering whether Chat GPT's success was built off the back of their copyrighted work.

27:22
Speaker 6
I think that authors don't write to get rich. The average author in America lives below the poverty line. So clearly we are a bunch of people who just do this because we care so intensely about the proprietary information. And I think that these companies who care intensely about their proprietary information are being deeply hypocritical in the way they treat our property.

27:51
Speaker 1
When Greg and I spoke in October, I asked him about authors'concerns and about OpenAI's recent secrecy about its training data. Why not disclose your training data?

28:04
Speaker 4
Well, I think know there's some answer for, well, it's proprietary because it's something that helps us sort of build the next system. It's something that we spent a lot of time on to get here. The selector is really for any given fact that we reveal how much does this help people who are going to use the system, how much does this help people choose what to use it for versus how much does it help people build the exact same thing that we're trying to build?

28:28
Speaker 1
And how do you respond to this growing list of authors who are saying, you're using my copyrighted work as training data and you need to pay us for that?

28:44
Speaker 4
Look, I think there's a really important question for society to answer here. I think that there's a legal question, which is definitely for the lawyers. I am not a lawyer, but I have an opinion. I do think that, first of all, that we as society, I think that there should be some way for compensation to flow back to people who are producing works. We're definitely very interested in that. We've been talking about that for some time. But I also think that there's something really interesting to step back and think about. These systems are just not what we're used to. Right. That what we're training them to do is to help us solve problems that no one's ever solved before. That's the actual goal.

29:20
Speaker 1
I ran Greg's response by Minjin Lee. The argument is sort of a greater good argument that AI is going to solve problems for the world that no one has solved before, and that in the process of creating this new thing, there'll be some troubles along the way that we'll just have to sort out.

29:41
Speaker 6
So I can say that my book and my work is a way of advancing humanity and it serves a greater good. I have absolutely no proof that their work will solve the greater good, and I'm not in any way coward by their theory of advancing humanity. Also, what they're doing is not for free. So why is their idea more valuable than mine? I don't think that it is.

30:12
Speaker 1
In the year after Chat GPT's release, public pressure continued to build on OpenAI, but it didn't keep the company down. OpenAI put out a steady stream of new releases. There was GPT four and the latest version of its image generator, Dali. The company's revenue was growing fast. By mid November, the company was closing in on a deal that would value OpenAI at nearly $90 billion, about triple what it had been worth. At the beginning of last year, Chat GPT had made OpenAI one of the most valuable tech companies in the world. And then came a shocking announcement that put the entire future of the company in jeopardy. Breaking news. Sam Altman is out as CEO of OpenAI, the company just announced. Coming up, an exclusive interview with one of the four board members who fired Sam Altman. Yeah, my name is Helen Toner.

31:21
Speaker 1
How would you describe that weekend for you? We knew that it would not be a joyride, and it wasn't a joyride. 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 Linebaugh. This episode was produced by Annie Minoff and Laura Morris, with help from Pierce Singhy and Kylan Burtz. The series is edited by Maria Byrne, fact checking by Amelia Schaunbeck, with consulting from Arvind Narainan. Series art by Pete Ryan. Sound design and mixing by Nathan Singapock. Music in this episode by Catherine Anderson, Peter Leonard, Emma Munger, and Nathan Singapock. Our theme music is by so Wiley and remixed by Nathan Singapock. Special thanks to Catherine Brewer, Jason Dean, Tom Doton, Karen Howe Berber Jin, Sarah Platt, Sarah Rabel, and Lisa Wang. Thanks for listening.

32:24
Speaker 1
Our final episode will be released later this month.