Advertisement

transcript

[MUSIC PLAYING]

From New York Times Opinion, this is “The Ezra Klein Show.”

[MUSIC PLAYING]

Before we get into the episode today, we are doing, or getting ready to do, our end of the year “Ask Me Anything.” So if you have questions you want to hear me answer on the show — I suspect a lot of them are going to be about Israel-Palestine and A.I., but they don’t have to be about Israel-Palestine and A.I. — send them to ezrakleinshow@nytimes.com with A.M.A. in the headline. Again, to ezrakleinshow@nytimes.com with A.M.A. in the headline.

[MUSIC PLAYING]

If you follow business or tech or artificial intelligence news at all, in recent weeks you certainly were following Sam Altman being unexpectedly fired as C.E.O. of OpenAI, and then a huge staff revolt at OpenAI where more than 95 percent of the company said it would resign if he was not reinstated. And then he was reinstated, and so this whole thing seemed to have happened for nothing.

I spent a lot of time reporting on this. And I talked to people on the Altman side of things. I talked to people on the board side of things. And the thing I am now convinced of — truly convinced of — is that there was less to it than met the eye. People saw — I saw — Altman fired by this nonprofit board meant to ensure that A.I. is built to serve humanity.

And I assumed — and I think many assumed — there was some disagreement here over what OpenAI was doing, over how much safety was building into the systems, over the pace of commercialization, over the contracts it was signing, over what it was going to be building next year, over something. And that, I think, I can say conclusively — and has been corroborated by other reporting — that was not what this was about.

The OpenAI board did not trust and did not feel it could control Sam Altman, and that is why they fired Altman. It’s not that they felt they couldn’t trust him on one thing, that they were trying to control him on X but he was beating them on X. It’s that a lot of little things added up. They felt their job was to control the company, that they did not feel they could control him, and so to do their job, they had to get rid of him.

They did not have, obviously, the support inside the company to do that. They were not ultimately willing to let OpenAI completely collapse. And so they largely — although, I think in their view not totally — backed down. One of the members is still on the board. Altman and the president of OpenAI, Greg Brockman, are off the board.

Some new board members are coming in who they think are going to be stronger and more willing to stand up to them. There’s an investigation that is going to be done of Altman’s behavior, that will be at least released to the board so they’ll I guess know what to think of him. It’s a very strange story. I wouldn’t be surprised if there’s things yet to come out. But I am pretty convinced that this was truly a struggle for control not a struggle about x.

But it has been a year since ChatGPT was released. It was a weird way to mark the year, but it has been a year. A year since OpenAI kicked off the whole modern era in artificial intelligence, a year since a lot of people’s estimations of what humanity’s future looked like began to shift and cloud, and darken, and shimmer.

And so I wanted to have the conversation that many of us thought was a conversation happening here about what A.I. was becoming, how it was being used, how it was being commercialized, whether or not the path we’re on is going to benefit humanity. And so I asked my friends over at “Hard Fork,” another great New York Times podcast, to come on the show.

Kevin Roose is my colleague at The Times. He writes a tech column called “The Shift.” Casey Newton is the editor of Platformer, an absolutely must-read newsletter about the intersection of technology and democracy. And they have been following this in and out, but they’ve been closely following A.I. for the past year. So I wanted to have this broader conversation with them. As always, my email ezrakleinshow@nytimes.com.

[MUSIC PLAYING]

Kevin Roose, Casey Newton, welcome to the show, my friends.

Hey, Ezra.

Thanks for having us.

All right, so we’re talking on Monday, Nov. 27. ChatGPT, which kicked off this era in A.I., was released on Nov. 30, 2022. So the big anniversary party was that Sam Altman got temporarily fired and the company almost collapsed and was rebuilt over at Microsoft, which I don’t think is how people expected to mark the anniversary.

But it has been now a year, roughly, in this whole new A.I. world that we’re in. And so I want to talk about what’s changed in that year. And the place I want to begin is with the capabilities of the A.I. systems we’re seeing. Not the ones we’re hearing about, but that we know are actually being used by someone in semi real-world conditions. What can A.I. systems do today that they couldn’t do a year ago, Kevin?

Well, the first most obvious capabilities improvement is that these models have become what’s called multimodal. So a year ago, we had ChatGPT, which could take in text input and output other text as the response to your prompt. But now we have models that can take in text and output images, take in text and output video, take in voice data and output other voice data.

So these models are now working with many more types of inputs and outputs than they were a year ago. And that’s the most obvious difference. If you just woke up from a year-long nap and took a look at the A.I. capabilities on the market, that’s the thing that you would probably notice first.

I want to pull something out about that — is that it almost sounds like they’re developing what you might call senses. And I recognize that there’s a real danger of anthropomorphizing A.I. systems, so I’m not trying to do that. But one thing about having different senses is that we get some information that helps us learn about the world from our eyes, other information that helps us learn about the world from our ears, et cetera.

One of the constraints on the models is how much training data they can have. As it becomes multimodal, it would seem that would radically expand the amount of training data. If you can have not just all of the text on the internet, but all of the audio on YouTube, or all the podcast audio on Spotify or something — or Apple podcasts — that’s a lot of data to learn about the world from, that, in theory, will make the models smarter and more capable. Does it have that kind of recursive quality?

Absolutely. I mean, part of the backdrop for these capabilities improvements is this race for high-quality data. All of the A.I. labs are obsessed with finding new undiscovered high-quality data sources that they can use to train their models. And so if you run out of text because you’ve scraped the entire internet, then you’ve got to go to podcasts or YouTube videos or some other source of data to keep improving your models.

For what it’s worth, though, I don’t think the availability of more training data is what is interesting about the past year. I think what was interesting about ChatGPT was that it gave average people a way to interact with A.I. for the first time. It was just a box that you could type in and ask it anything and often get something pretty good in response.

And even a year into folks using this now, I don’t think we’ve fully discovered everything that it can be used for. And I think more people are experiencing vertigo every day as they think about what this could mean for their own jobs and careers. So to me, the important thing was actually just the box that you type in and get questions from.

Yeah, I agree with that. I think if you had just paused there and there was no new development in A.I., I think it would still probably take the next five or 10 years for society to adjust to the new capabilities in our midst.

So you’ve made this point in other places, Casey, that a lot of the advances to come are going to be in user interfaces, in how we interact with these systems. In a way, that was a big advance of ChatGPT. The system behind it had been around for a while. But the ability to speak to it, or I guess write to it in natural language, it created this huge cultural moment around A.I.

But what can these A.I. products actually do that they couldn’t do a year ago? Not just how we interface with them, but their underlying capacity or power.

I mean, as of the Developer Day Update that OpenAI had a few weeks back, the world knowledge of the system has been updated to April of this year. And so you’re able to get something closer to real-time knowledge of world events. It has now integrated with Microsoft Bing. And so you can get truly real-time information in a way that was impossible when ChatGPT launched.

And these might sound like relatively minor things, Ezra, but you start chaining them together, you start building the right interfaces, and you actually start to see beyond the internet as we know it today. You see a world where the web, where Google is not our starting point for doing everything online. It is just a little box on your computer that you type in and you get the answer without ever visiting a web page. So that’s all going to take many years to unfold, but the beginnings of it are easy to see now.

One other capability that didn’t exist a year ago, at least in any public products, is the ability to bring your own data into these models. So Claude was the first language model that I used that had the ability to, say, upload a PDF. So you could say, here’s a research paper. It’s 100 pages long. Help me summarize and analyze this. And it could do that. Now ChatGPT can do the same thing. And I know a bunch of other systems are moving in that direction too.

There are also companies that have tried to spin up their own language models that are trained on their own internal data. So if you are Coca-Cola, or B.C.G., or some other business, and you want an internal ChatGPT that you can use for your own employees to ask, say, questions about your H.R. documents, that is a thing that companies have been building. So that’s not the sexiest, most consumer-facing application, but that is something that there’s enormous demand for out there.

So one thing it seems to me to be getting better at, from what I can tell from others, is coding. I often ask people whether they’re using A.I. bots very often, and if so, for what? And basically nobody says yes unless they are a coder. Everybody says, oh, yeah, I played around with it. I thought it was really cool. I sometimes use DALL-E or Midjourney to make pictures for my kids or for my email newsletter. But it is the coders who say, I’m using it all the time, it has become completely essential to me. I’m curious to hear a bit about that capability increase.

I think where it has become part of the daily habit of programmers is through tools like GitHub Copilot, which is basically ChatGPT for coders that finishes whatever line of code you’re working on or helps you debug some code that’s broken.

And there have been some studies and tests. I think there was one test that GitHub itself ran where they gave two groups of coders the same task, and one group was allowed to use GitHub Copilot and one group wasn’t. And the group with GitHub Copilot finished the task 55 percent faster than the group without it. Now, that is a radical productivity increase. And if you tell a programmer, here’s a tool that can make you 55 percent faster, they’re going to want to use that every day.

So when I see, functionally, chatbots in the wild, what I see is different versions of what people used to somewhat derisively call the fancy autocomplete, right? Help you finish this line of code, help you finish this email. You ask a question that you might ask the search engine, like why do I have spots all over my elbow? And it gives you an answer that hopefully is right but maybe is not right.

I do think some of the search implications are interesting. But at the same time, it is not the case that Bing has made great strides on Google. People have not moved to asking the kind of Bing chatbot its questions as opposed to asking Google. Everybody feels like they need A.I. in their thing now. There’s a — I don’t think you can raise money in Silicon Valley at the moment if you don’t have a generative A.I. play built into your product or built into your business strategy.

But that was true for a minute for crypto, too. And I’m not one of the people who makes a crypto A.I. analogy. I think crypto is largely vaporware and A.I. is largely real. But Silicon Valley is faddish and people don’t know how to use things and so everybody tries to put things in all at once. What product has actually gotten way better?

I’ll just use one example. There’s an app you might be familiar with called Notion. It’s productivity collaborative software. I write a newsletter. I save every link that I put in my newsletter into Notion. And now that there is A.I. inside Notion, Notion can do a couple of things. One, it can just look at every link I save and just write a two-sentence summary for me, which is just nice to see, at a glance, what that story is about.

And most recently it added a feature where you can just do Q&A with a database and say like, hey, what are some of the big stories about Meta over the past few weeks? And it’ll just start pulling those up, essentially querying the database that I have built. And so while we’re very early in this, you’re beginning to see a world where A.I. is taking data that you have stored somewhere and it’s turning it into your personal research assistant. So is it great right now? No, I would give it like a C. But for 1.0, I think it’s not bad.

And I’ll share another example that is not from my own use, but I was talking a few weeks ago with a doctor who’s a friend of a friend. And doctors, you get tons of messages from patients. What’s this rash? Can you renew this prescription? Do I need to come in for a blood test, that kind of stuff. And doctors and nurses spend a ton of time just opening up their message portal, replying to all these messages. It’s a huge part of being a doctor and it’s a part that they don’t like.

And so this doctor was telling me that they have this software now that essentially uses a language model — I assume it’s OpenAI’s or someone very similar to that — that goes in and pre-fills the responses to patient queries. And the doctor still has to look it over, make sure everything’s right and press Send. But just that act of pre-populating the field, this person was saying it saves them a ton of time, like on the order of several hours a day.

And if you have that, and you extrapolate to what if every doctor in America was saving themselves an hour or two a day of responding to patient messages? I mean, that’s a radical productivity enhancement. And so you can say that that’s just fancy autocomplete, and I guess on some level it is. But just having fancy autocomplete in these paperwork-heavy professions could be very important.

Well, let me push that in two directions, because one direction is that I am not super thrilled about the idea that my doctor, theoretically here, is glancing over things and clicking “submit” as opposed to reading my message themselves and having to do the act of writing, which helps you think about things and thinking about what I actually emailed them and what kind of answer they need to give me. I mean, I know personally the difference in thought between scanning things and editing and thinking through things. So that’s my diminishing response.

But the flip of it is the thing I’m not hearing anybody say here, and the thing I keep waiting for and being interested in is the things A.I. might be able to do better than my doctor. I was reading Jack Clark’s “Import A.I.” newsletter today, which I super recommend to people who want to follow advancements in the field, and he was talking about a — I mean, it was a system being tested, not a system that is in deployment — but it was better at picking up pancreatic cancer from certain kinds of information than doctors are.

And I keep waiting to hear something like this going out into the field. Something that doesn’t just save people a bit of time around the edges. I agree that’s a productivity improvement. It’s fine. You can build a business around that. But the promise of A.I. when Sam Altman sat with you all a few weeks ago, or however long it was, and said, we’re moving to the best world ever, he didn’t mean that our paperwork is going to get a little bit easier to complete. He meant we’d have cures for new diseases. He meant that we would have new kinds of energy possibilities. I’m interested in the programs and the models that can create things that don’t exist.

Well, to get there, you need systems that can reason. And right now the systems that we have just aren’t very good at reasoning. I think that over the past year we have seen them move a little away from the way that I was thinking of them a year ago, which was a sort of fancy autocomplete. It’s making a prediction about what the next word will be. It’s still true that they do it that way. But it is able to create a kind of facsimile of thought that can be interesting in some ways.

But you just can’t get to where you’re going, Ezra, with a facsimile of thought. You need something that has improved reasoning capabilities. So maybe that comes with the next generation of frontier models. But until then, I think you’ll be disappointed.

But do you need a different kind of model? This is something that lingers in the back of my head. So I did an interview on the show with Demis Hassabis who’s the co-founder of DeepMind and now runs the integrated DeepMind Google A.I. program. And DeepMind had built this system a while back called AlphaFold, which treated how proteins are constructed in 3-D space, which is to say in reality. We live in 3-D space. [LAUGHS] It treated it as a game. And it fed itself a bunch of information and it became very good at predicting the structure of proteins. And that solved this really big scientific problem. And they then created a subsidiary of Alphabet called Isomorphic Labs to try to build drug discovery on similar foundations.

But my understanding is that Google, during this period, became terrified of Microsoft and OpenAI beating it up in search and Office. And so they pulled a lot of resources, not least Hassabis himself, into this integrated structure to try to win the chatbot wars, which is now what their system Bard is trying to do. And so when you said, Casey, that we need things that can reason, I mean maybe. But also, you could say we need things that are tailored to solve problems we care about more.

And I think this is one of the things that worries me a bit, that we’ve backed ourselves into business models that are not that important for humanity. Is there some chance of that? I mean, are we going too hard after language-based general intelligent A.I. that, by the way, integrates very nicely into a suite of enterprise software as opposed to building things that actually create scientific breakthroughs but don’t have the same kind of high scalability profit structure behind them?

I would stick up for the people who are working on the what you could call the non-language problems in A.I. right now. This stuff is going on. It maybe doesn’t get as much attention from people like the three of us as it should. But if you talk to folks in fields like pharmaceuticals and biotech, there are new A.I. biotech companies spinning up every day, getting funding to go after drug discovery or some more narrow application.

We talked to a researcher the other day, formerly of Google, who is teaching A.I. to smell. Taking the same techniques that go into these transformer-based neural networks like ChatGPT and applying them to the molecular structures of different chemicals, and using that to be able to predict what these things will smell like. And you might say, well, what’s the big deal with that? And the answer is that some diseases have smells associated with them that we can’t pick up on because our noses aren’t as sensitive as, say, dogs or other animals. But if you could train an A.I. to be able to recognize scent molecules and predict odors from just chemical structures, that could actually be useful in all kinds of ways. So I think this kind of thing is happening. It’s just not dominating the coverage the way that ChatGPT is.

[MUSIC PLAYING]

Let me ask you, Kevin, about I think an interesting, maybe promising, maybe scary avenue for A.I. that you possibly personally foreclosed, which is, at some point during the year, Microsoft gave you access to a OpenAI-powered chatbot that had this dual personality of Sydney. And Sydney tried to convince you you didn’t love your wife and that you wanted to run away with Sydney. And my understanding is immediately after that happened, everybody with enough money to have a real business model an A.I. lobotomized the personalities of their A.I.s. That was the end of Sydney.

But there are a lot of startups out there trying to do A.I. friends, A.I. therapists, A.I. sex bots, A.I. boyfriends and girlfriends and nonbinary partners, just every kind of A.I. companion you can imagine. I’ve always thought this is a pretty obvious way this will affect society, and the Sydney thing convinced me that the technology for it already exists. So where is that, and how are those companies doing?

Yeah, I mean, I’m sorry, A, if I did foreclose the possibility of A.I. personalities.

I think what’s happening is it’s just a little too controversial and fraught for any of the big companies to wade into. Microsoft doesn’t want its A.I. assistants and co-pilots to have strong personalities. That much is clear. And I don’t think their enterprise customers want them to have strong personalities, especially if those personalities are adversarial, or confrontational, or creepy, or unpredictable in some way. They want Clippy but with real brain power.

But there are companies that are going after this more social A.I. market. One of them is this company Character A.I., which was started by one of the original people at Google who made the transformer breakthrough. And that company is growing pretty rapidly. They’ve got a lot of users, especially young users, and they are doing, essentially, A.I. personas. You can make your own A.I. persona and chat with it or you can pick from ones that others have created.

Meta is also going a little bit in this direction. They have these persona-driven A.I. chatbots. And all of these companies have put guardrails around — no one really wants to do the erotic — what they call erotic role-play, in part because they don’t want to run afoul of things like the Apple app store terms of service.

But I expect that that will also be a big market for young people. And anecdotally, I mean I have just heard from a lot of young people who already say, like, my friends have A.I. chatbot friends that they talk to all the time. And it does seem to be making inroads into high schools. And that’s just an area that I’ll be fascinated to track.

I mean, this is going to be huge. A couple thoughts come to mind. One, I talked to somebody who works at one of the leading A.I. companies, and they told me that 99 percent of people whose accounts they remove, they remove for trying to get it to write text-based erotica. So that, I think, speaks to the market demand for this sort of thing.

I’ve also talked to people who have used the models of this that are not constrained by any safety guidelines, and I’ve been told these things are actually incredible at writing erotica. So what I’m telling you is there is a $10 billion —

So you’ve done really a lot of reporting on this, Casey. You’d say maybe a personal interest?

I’m so interested in this. Look, I write about content moderation, and porn is the content moderation frontier. And it’s just very interesting to me that it’s so clear that there are billions of dollars to be made here and no company will touch it.

And I asked one person involved, I said, well, why don’t you just let people do this? And they basically said, look, if you do this, you become a porn company overnight. It overwhelms the usage — this is what people wind up using your thing for and you’re working at a different company then. So I sort of get it.

But even setting aside the explicitly erotic stuff, Ezra, you well know and have talked and written about, just the loneliness epidemic that we have in this country. There’s a lot of isolated people in this world. And I think there is a very real possibility that a lot of those people will find comfort and joy and delight with talking to these A.I.-based companions.

I also think that when that happens, there will be a culture war over it. And we will see lengthy segments on Fox News about how the Silicon Valley technologists created a generation of shut-ins who wants to do nothing but talk to their fake friends on their phones. So I do think this is the culture war yet to come, and the question is just when do the enabling technologies get good enough and when do companies decide that they’re willing to deal with the blowback?

I also think this is going to be a generational thing. I mean, I’m very interested in this and have been for a bit, in part because I suspect, if I had to make a prediction here, my five-year-old is going to grow up with A.I. friends. And my pat line is that today we worry that 12-year-olds don’t see their friends enough in person. And tomorrow we’ll worry that not enough of our 12-year-old’s friends are persons. Because it’s going to become normal.

And my sense is that the systems are really good. If you unleashed them, you are already good enough to functionally master this particular application. And the big players simply haven’t unleashed them. I’ve heard from people at the big companies here who are like, oh, yeah, if we wanted to do this, we could dominate it.

But that does bring me to a question, which is Meta kind of does want to do this. Meta, which owns Facebook, which is a social media company. They seem to want to do it in terms of these lame-seeming celebrity avatars. You can talk to A.I. Snoop Dogg.

It’s so bad.

But Meta is interesting to me because their A.I. division is run by Yann LeCun, who’s one of the most important A.I. researchers in the field, and they seem to have very different cultural dynamics in their A.I. shop than Google DeepMind or OpenAI. Tell me a bit about Meta’s strategy here and what makes them culturally different.

Well, Casey, you cover Meta and have for a long time and may have some insight here. My sense is that they are up against a couple problems, one of which is they have arrived to A.I. late and to generative A.I. specifically. Facebook was, for many years, considered one of the top two labs along with Google when it came to recruiting A.I. talent, to putting out cutting edge research, to presenting papers at the big A.I. conferences. They were one of the big dogs.

And then they had this funny thing happen where they released a model called Galactica just right before ChatGPT was released last year. And it was supposed to be this L.L.M. for science and for research papers. And it was out for I think three days. And people started noticing that it was making up fake citations. It was hallucinating. It was doing what all the A.I. models do. But it was from Meta and so it felt different. It had this tarnish on it because people already worried about fake news on Facebook. And so it got pulled down. And then ChatGPT, just shortly thereafter, launched and became this global sensation.

So they’re grappling for what to do with this technology that they’ve built now. There’s not a real obvious business case for shoving A.I. chatbots into products like Facebook and Instagram. And they don’t sell enterprise software like Microsoft does, so they can’t really shove it into paid subscription products. So my sense from talking with folks over there is that they’re just not sure what to do with this technology that they’ve built. And so they’re just flinging it open to the masses. What do you think?

That tracks with me. I basically don’t get it either. Basically what you’ve just said has been explained to me. They are investing a ton with no obvious return on investment in the near term future. I will say that these celebrity A.I. chatbots they’ve made are quite bad.

Horrible.

It’s truly baffling. And the thing is they’ve taken celebrities, but the celebrities are not playing themselves in the A.I. They’ve given all of the celebrities silly names. And you can just follow their Instagram and send them messages and say, hey, character that Snoop Dogg is portraying, like what do you think about this? So it’s all very silly. And I expect it’ll die a rapid death sometime in the next year and then we’ll see if they have a better idea.

What I will say is if you’re somebody who wakes up from A.I. nightmares some mornings, as a lot of folks in San Francisco do, go listen to Yann LeCun talk about it. No one has ever been more relaxed about A.I. than Yann LeCun. It’s just like an army of superhuman assistants are about to live inside your computer. They’re going to do anything you want to do and there’s no risk of them harming you ever. So if you’re feeling anxious, go listen to Yann.

Do you think he’s right? Because it also has led to a policy difference. Meta has been much more open source in their approach, which OpenAI and Google seem to think is irresponsible. But there is something happening there that I think is also built around a different view of safety. What is their view of safety? Why does Yann LeCun, who is an important figure in this whole world, why is he so much more chill than, you know, name your other founder?

I mean, part of it is I just think these are deeply held convictions from someone who is an expert on this space and who has been a pioneer and who understands the technology certainly far better than I do. And he can just not see from here to killer robot. So I respect his viewpoint in that respect given his credentials in the space.

I think on the question of is open source A.I. safer, this is still an open question — not to pun. The argument for it being safer is, well, if it’s open source, that means that average people can go in and look at the code, and identify flaws, and see how the machine works, and they can point those out in public, and then they can be fixed in public.

Whereas if you have something like OpenAI, which is building very powerful systems behind closed doors, we don’t have the same kind of access. And so you might not need to rely on a government regulator to see how safe their systems were. So that is the argument in favor of open source.

Of course, the flip side of that is like, well, if you take a very powerful open source model and you put it out on the open web, even if it’s true that anyone can poke holes and identify flaws, it’s also true that a bad actor could take that model and then use it to do something really, really bad. So that hasn’t happened yet, but it certainly seems like it’s an obvious possibility at some time in the near future.

Let me use that as a bridge to safety more generally. So we’ve talked a bit about where these systems have gone over the past year, where they seem to be going. But there’s been a lot of concern that they are unsafe and fundamentally that they’ve become misaligned or that we don’t understand them or what they’re doing. What kind of breakthroughs have there been with all this investment and all this attention on safety, Kevin?

So a lot of work has gone into what is called fine-tuning of these models. So basically if you’re making a large language model, like GPT-4, you have several phases of that. Phase one is what’s called pre-training, which is just the basic process. You take all of this data. You shove it into this neural network and it learns to make predictions about the next word in a sequence.

Then from there, you do what’s called fine-tuning. And that is basically where you are trying to turn the model into something that’s actually useful or tailored. Turn it into a chatbot. Turn it into a tool for doctors. Turn it into something for social A.I.s. That’s the process that includes things like reinforcement learning from human feedback, which is how a lot of the leading models are fine-tuned. And that work has continued to progress. The models, they say, today are safer and less likely to generate harmful outputs than previous generations of models.

There’s also this field of interpretability, which is where I’ve been doing a lot of reporting over the past few months, which is this tiny subfield of A.I. that is trying to figure out what the guts of a language model look like and what is actually happening inside one of these models when you ask it a question or give it some prompt and it produces an output.

And this is a huge deal, not only because I think people want to know how these things work — they’re not satisfied by just saying these are mystical black boxes — but also because if you understand what’s going on inside a model, then you can understand if, for example, the model starts lying to you or starts becoming deceptive, which is a thing that A.I. safety researchers worry about. So that process of interpretability research I think is really important. There have been a few minor breakthroughs in that field over the past year, but it is still slow going and it’s still a very hard problem to crack.

And I think it’s worth just pausing to underscore what Kevin said, which is the people building these systems do not know how they work. They know at a high level, but there is a lot within that where if you show them an individual output from the A.I., they will not be able to tell you exactly why it said what it said. Also, if you run the same query multiple times, you’ll get slightly different answers. Why is that? Again, the researchers can’t tell you.

So as we have these endless debates over A.I. safety, one reason why I do tend to lean on the side of the folks who are scared is this exact point. At the end of the day, we still don’t know how the systems work.

Tell me if this tracks for you. I think compared to a year ago, when I talked to the A.I. safety people, the people who worry about A.I.s that become misaligned and do terrible civilizational-level damage, A.I.s that could be really badly misused, they seem to think it has been actually a pretty good year, most of them.

They think they’ve been able to keep big models, like GPT-4, which, of course, are much less powerful than what they one day expect to invent, but they think they’ve been pretty good at keeping them aligned. They have made some progress on interpretability, which wasn’t totally clear a year ago. Many people said that was potentially not a problem we could solve. At least we’re making some breakthroughs there.

They’re not relaxed — the people who worry about this. And they will often say we would need a long time to fully understand even the things we have now and we may not have that long. But nevertheless, I get the sense that the safety people seem a little more confident that the technical work they’ve been doing is paying off than at least was the impression I got from the reporting prior.

I think that’s right. I mean, Sam Altman, in particular, this has been his strategy, is we are going to release this stuff that is in our labs and we’re going to wait and see how society reacts to it. And then we’ll give it some time to let society address and then we will release the next thing. That’s what he thinks is the best way to slowly integrate A.I. into our lives.

And if you’d asked me maybe 11 months ago, like a month into using ChatGPT, what are the odds of something really, really bad happening because of the availability of ChatGPT, I would have put them much higher than they turned out to be. And when you talk to folks at OpenAI, they will tell you that that company really has taken A.I. safety really seriously. You can see this yourself when you use the product. Ask it a question about sex, it basically calls the police. So there is a lot to be said for how these systems have been built so far.

And I would say the other thing that I’ve heard from A.I. safety researchers is that they are feeling relief not just that the world has not ended but that more people are now worried about A.I. It was a very lonely thing for many years to be someone who worried about A.I. safety because there was no apparent reason to be worried about A.I. safety, right? The chatbots that were out, it was like Siri and Alexa and they were terrible. And no one could imagine that these things could become dangerous or harmful because the technology itself was just not that advanced.

Now you have congressional hearings. You have regulations coming from multiple countries. You have people like Geoff Hinton and Yoshua Bengio, two of the so-called godfathers of deep learning, proclaiming that they are worried about where this technology is headed. So I think for the people who’ve been working on this stuff for a long time, there is some just palpable relief at like, oh, I don’t have to carry this all on my shoulders anymore. The world is now........

© The New York Times