#117 - Google’s Bard Rush, BloombergGPT, ChatGPT King, Balenciaga Harry Potter
01:36:59 Share

Hello and welcome to Skyner Today's Last Week in AI podcast, where you can hear us chat about what's going on with AI. As usual, in this episode, we'll provide summaries and discuss some of last week's most interesting AI news. You can also check out our Last Week in AI newsletter at lastweekin.ai for articles that we did not cover in this episode.

I am one of your hosts, Andrey Karenkov. I'm currently finishing up my PhD at the Stanford AI Lab, and I'm about to work on generative tools for video games. Oh, okay. So just so everybody knows, that was the first time I found out that Andrey is going to be working on generative tools for video games. Yeah, yeah. I'm just about to start work, so it's brand new. Oh, look at that. Look at that. Can you share the company name?

It's a small kind of stealthy thing. Oh, okay, okay, okay. Oh, Stealth Co. I love Stealth Co. Okay, great. Yeah, yeah. Sorry, I got distracted. I'm your other co-host here, Jeremy Harris. So I do a bunch of work on AI safety stuff with a company called Gladstone AI. We work with AI alignment researchers and policy researchers at some of the world's top labs. The names you'll probably know. And yeah, I work on AI safety. I have a book also, I should mention.

My editors will kill me if I don't. It's called Quantum Physics Made Me Do It. What the hell does that have to do with AI? Well, you can find out by reading the book. And anyway, it actually is connected kind of in a cool way. Think AI sentience, that sort of thing. Anyway, if that's your cup of tea, please feel free to check it out. And without further ado, we have a big week, I guess, to cover here. Eh, Andre? Yeah, we sure do. As usual this year, every week has been...

It's crazy. This podcast used to be like an hour or shorter and now it's just an hour and a half every time. So we'll see how this one goes.

Let's go ahead and dive in. First, our application and business stories, starting with Google shuffles assistant leadership to focus on barred AI. So this is about an internal memo that basically showcases how Google is freaking out about what's going on with ChatGPT and how

Their release of BARD, their version of Bing Chat has seemingly not gone very well. So as a result, there's a bunch of basically turmoil within the company. What did you take away from this, Jeremy? Yeah, I think you're right. It feels...

It feels like turmoil from the outside. It's always hard to know. The narrative, the public narrative is like, oh, OpenAI is setting the course here with chat GPT and GPT-4, and Google is kind of responding, and it seems panicky. Obviously, we don't know. We can't tell what's happening under the hood. But I will say there are kind of two big categories of facts.

that often can give you insight into what a company is doing that a company really can't hide. And one of them is major personnel transfers. That's this. And the other is major equity shuffles. We saw that when Microsoft, for example, took a 49% stake in OpenAI. That's a signal you just can't hide for various legal reasons. There's just no way around it. So I think this falls into that category of very reliable companies

unhideable information. I think the implications are interesting. It was a leaked internal memo too, so worth noting this isn't something that Google just broadcast. But yeah, I think one of the really interesting things is that we're seeing people get pulled off the Google Assistant team and moved into the BARD team, so senior executive in this case, getting pulled off Google Assistant, which is a pretty successful thing in and of itself.

So the theory here is that maybe that executive is now going to be focusing on how to build BARD in such a way as to help with the Assistant product. So we might see Google Assistant potentially moving in the direction of less like an expert system type of setup where it has pre-programmed responses for things and more towards a kind of AI first chat GPT style bot.

Which in its own right, I think is fascinating, right? This is like a fundamental shift in the kind of AI technology, the kinds of home AI systems that we've seen, where we're moving from these pre-programmed response streams that are more controllable, but less flexible to these like AI first strategies. So that was, I think, some of the high level thoughts that I had.

Yeah, I think it's interesting how this highlights Assistant in particular. If anyone doesn't know, if you're an iPhone person, Assistant is basically Google's Siri equivalent. Both text and audio, it kind of does both. So you can easily imagine how something like ChatGPT would plug into something like Siri or Google Assistant in a lot of ways.

And yeah, so to me, it kind of makes sense for them to try and get those two to meet up and sort of glue their code together. As you said, the technology behind Assistant is probably a bit old school. There's a bunch of parsing and querying databases and things like that. So it'll be interesting to see how quickly they can roll out this refined version that's based on something like BARD.

And related to this, I also noticed there was an article in the information that is also not some sort of public release. This is interviewing from employees at Google and DeepMind. So DeepMind is a subsidiary of Alphabet that is a big AI lab, effectively. It does very little business, AlphaGo,

AlphaFold, all these things you probably heard of what they do. And this article shows how now there is teamwork between Google and DeepMind to compete with ChashGPT, with this project that's codenamed Gemini.

So that's another sign of how big a deal it is for Google. Usually Google and DeepMind have been pretty separate. Avuna had some kind of divisions and some disagreements, but now it looks like there's teamwork, all hands on deck for this.

Yeah, and connecting it to the story we just covered too, as we look at this transition from these more expert systems, like you said, pre-programmed, in a sense, pre-programmed responses, querying databases, as we shift from that to these AI-first systems,

The risk starts to go up a little bit too, right? You can imagine you ask this thing for health advice, you ask it for psychological advice, travel advice. It's going to start to affect your behavior in the real world in a way that's more significant. And so now, yeah, with this story, I think we're seeing another consequence of this acceleration that was caused by ChatGPT's super public release.

And Google now sort of being confronted with like, what do we do? And historically, there has been tension between DeepMind and the mother org here, right? We know, for example, that DeepMind was founded by, among others, Demis Hassabis, who was very focused on AI safety, who didn't want to risk losing control over the org in the Google activity.

acquisition in 2014. And so they had a whole bunch of like independent oversight requirements and things like that, that were introduced that caused friction with the mother org. And it's kind of interesting now, if you're a deep mind, it's like, okay, we want to design and develop AI safely. We set up a bunch of kind of blockers, kind of points of friction with the parent work just so that we could do that. But now we have open AI. They're just going to go out and like

like seemingly launched their own stuff anyway. So is it better if we just say, okay, screw it, we're going to partner now fully with the mother org? You know, I'm glossing over a whole bunch of detail here, I'm sure, and there's going to be nuance under the surface. But at a high level, it kind of seems like, you know, we're seeing that race to the bottom on safety and people are making bigger and bigger bets just to gain the extra couple of weeks or months in releasing these new products.

Yeah, exactly. I think it is kind of interesting to note this parallel between DeepMind and OpenAI because going back to 2015, DeepMind was, I think, around 2013 acquired by Google, OpenAI was started in 2015. And initially, they were kind of the same basic thing of a big lab in industry that sought to create general AI with the goal of doing it safely.

roughly speaking. And then over time, their approaches have really diverged where OpenAI put all its chips into scaling up and in particular scaling up language models. In the first two years, OpenAI was doing all this reinforcement learning, playing Dota and having robot hands, and then they sort of just threw it all away and almost exclusively are working on language models. DeepMind

has not gone that course it has done work on this language stuff but it has also done work on you know healthcare ai and a lot of reinforcement learning i think we discussed one of their works recently with the general uh human uh level adaptation or whatever it was so yeah i think

It's not super relevant, but it's kind of interesting to see how over time the two big AI labs diverged and to create different courses.

I think it is actually totally relevant on the safety side. I think you're exactly right. You have these two orgs that kind of agree on the risk, but they have completely different philosophies about where the risk's going to come from. And to add to what you just said, OpenAI on risk seems to have this view that the best way to de-risk these systems is...

is to release them, let the world play with them and learn the ways in which people find to like jailbreak these systems, to kind of make them, to use them for malicious purposes and things like that. And they're hoping, they're essentially putting all their eggs in the basket of hoping that by doing that, you can kind of iteratively work your way to a safe superhuman AI system. Whereas DeepMind seems to have the view that like, well, wait a minute, there might be like kind of a phase transition at some point.

where these things become intrinsically dangerous and we don't know when that's going to happen, but we want to kind of think in terms of that leap and not necessarily default to releasing things. And it's not like we don't know who's right here. We just know that the answer is going to be really important. And that's kind of a weird space to be in, like these two philosophies being so different.

Yeah, no, there's a lot of interesting parallels where, you know, DeepMind is owned by Google or Alphabet and OpenAI is, you know, largely...

You know, Microsoft is doing a lot of steering just because they have the money for the compute. And so Google hasn't really pushed DeepMind to be very profitable or really at all profitable. They've done some revenue generation via different kind of systems, but they never really did try to go for that. And they're still publishing dozens of papers. OpenAI

doesn't really publish papers. They publish reports, but they don't have a lot of details. They used to publish AI safety papers, but I haven't seen that lately. I don't even know if there's a team that's doing that research anymore. DeepMind has published a lot of papers on AI safety. So

I don't know. It's a good contrast and it does seem like maybe AI research is not the priority for either of them right now. Yeah. To what you were saying about DeepMind and profitability, I believe they were officially profitable for the first time about two years ago.

So it's like, and when we say profitable, it's not like they're selling stuff to other companies. They're finding ways to save Google money. And one of the projects that, Andre, you were alluding to just for listeners, like one of the big ones, the big success stories was finding ways to basically cheapen the cost of deep mines like servers running and making them burn less, like cool. I think it was cheaper server cooling or something like that.

So, those internal use cases being really important. And OpenAI does have a safety team and alignment team in their defense on this, and they are focused on it fairly significantly. I think one of the challenges is too, it's starting to get really hard to tell what is safety research and what's capabilities research.

It used to be really clear. Like it used to be, you could say, okay, this paper is about controlling the AI system. And now this paper is about capabilities. But if you look at chat GPT, that arguably is a success in scaling. So capabilities, but also tuning the behavior of that scaled model through human feedback or LHF. And that's, that used to be a,

safety thing. And so I think we're seeing a blurring of the lines and I think people are starting to worry about giving away effectively trade secrets, even just by sharing so-called safety stuff. So not necessarily a great thing if you want safety knowledge to be very widely distributed, but it's an interesting dimension of this for sure. Yeah, that's a good point. I think when this is a commercial product and one of the ways that

that you can be better than the other product is to be more reliable, right? Or to not say bad things. It's a competitive advantage to have better AI safety. So that is kind of awkward. It's good and bad, I guess, because at least people are competing to do it better, but then they're not sharing it. I don't know. I don't know where to fall on that. Yeah. Yeah.

And on to the next story, moving away a bit from language models, we have something about robotics. So the Robotics Business Review has published the kind of set of investments that showed that in February 2023, robotics investments totaled $620 million. This was just published end of March, to be clear. So...

This is a bit of a few months away. It has a pretty cool breakdown showing that

There's 36 major investments. So the total for 2023 so far, I guess for first few months has been over 1 billion. And some of these sectors include drones that are used for a lot of stuff, data analytics for farms or for infrastructure or various things like that.

And then of course there's construction and healthcare, a lot of these different things. And they have some nice breakdowns of investment amounts, stage, country, things like that. So yeah, I found it kind of neat to see this breakdown of what's happening in robotics.

Yeah, it's also like it's always tough when, you know, you're in an industry and you're looking at like 36 deals in total because it can be so noisy from, you know, from one quarter to the next. But overall, yeah, it seems like, you know, good, healthy activity and definitely countering that narrative a little bit that we've seen. I know it's a narrative I myself have been guilty of propagating on the show specifically. So I do apologize to our listeners.

But historically, people have said stuff like, hey, language models are just taking all the air out of the room for these robotics companies. And it is kind of cool to see that, in fact, it's very much an active area of investment. Maybe a good kind of moment to throw out there, just an idea that I've been mulling over about AI companies and robotics.

The purpose to a significant degree of venture capital is to back very fast growing businesses that have long payback time horizons.

And the challenge with a lot of these software-oriented products is, yeah, they might go super fast on the way up, but how long will they actually last until there's a complete paradigm shift? And all of a sudden, there's some version of ChatGPT that is, I don't know, not an embodied thing, but anyway, we just find an audio-based system or something like that that just completely takes out this entire class of product.

It's very much not inconceivable that happens on the order of months now, just given how fast things are moving. Robotics, on the other hand, still has that characteristic of like, it's a moat. It's harder to compete. So it can make more sense to invest in arguably...

think, and this is where I'm just throwing out a half-baked idea that I've been mulling over and talking to some of my angel investor friends about, but maybe hardware becomes the place where VCs see the most returns, paradoxically, in the coming five years. Maybe it's not software so much despite all the hype, but anyway.

Anyway, just a random thought there. Yeah, it's interesting, I guess, if you are into investing, there's always a risk return trade-off for robotics. You probably have less risk just because there's fewer players, but it'll take way longer to get to revenue and profit as well, obviously.

So, yeah, I'm curious to see what will happen for the rest of this year. There's some fun breakdowns here. It shows that, for instance, 25% of these investments were seed and 14 were A. So these are pretty early stage investments for the most part. A lot of them on the US and China, things like that. So I think...

And one other thing that I am curious about, and I don't talk about that here, but it seems to me that probably when you talk about venture capital firms, they have different specializations. So some of them specialize in hardware industries, right? Manufacturing and logistics and things like that. And other ones are more software oriented. And I do feel there is still...

a set of investors that their wheelhouse is just in the physical world. And that means that we're not going to suddenly switch over to invest in ChagGPT or whatever, right? And in a way, there's real problems that you can only do with actual physical robots. So I think, yeah, let's not discount the importance or I guess...

popularity of this stuff. It's not exploding or skyrocketing like all these other things, but I think we will continue to see more and more robotics out there. Yeah, there's an old saying in YC, they say hardware is hard. And I guess this is why. As a last quick question, oh, sorry. Yeah. So speaking of that, I noticed another story I just felt like mentioning related to this. So

Bloomberg just had this article with human metalworkers hard to come by robotic blacksmith step up. And this is covering a particular company, Machino Labs, which is developing these robotic systems that can do what humans have traditionally done, this prototyping work, blacksmithing essentially.

And that is very tricky because you have to work with actual materials and you need some of that sort of, I guess, intuition and hand skills. So now with computer vision and AI, you can actually shape these materials in a robust way.

So yeah, that's just one example I thought was interesting where you would think that blacksmithing is not so much of a thing anymore. But no, actually, you still need people to do it for a lot of these prototyping steps early on. MARK MANDEL: Yeah, and another example of how hard it is to predict which jobs are going to be vulnerable to automation in the coming years. We have yet to see if they'll actually succeed with this at scale, but kind of interesting.

Also, looking back at this, there's a chart that comes with the robotics investments article that we were talking about a minute ago. I noticed China is about a fifth of the investment amount that the US is. I'm trying to decide whether that's surprising.

There's definitely a more developed VC environment in the States, but China tends to be really good on hardware. The whole Shenzhen ecosystem is just really well-developed. I don't know. Did you find that surprising? Now that you mention it, yeah, I think so.

I looked at the pie chart and there's fewer investments, but by amount, it's pretty surprising because you have like 450 million investment in the US and just 80 million in China, which is surprising. I mean, they have a lot of manufacturing going on. So as usual, when you crunch for data, there's a lot of questions as to how it was compiled and so on. But

Yeah, interesting. And I think good to keep an eye on just if you really want to have a comprehensive idea of what's happening in the AI economy, it's worth it to not forget about robotics.

And on to the lightning round of stories. First up, we have Microsoft's Bing chatbot is getting ads. Pretty much what this says, we're getting ads to pay for what is kind of expensive to service models.

Isn't it a rare treat when the actual article title just tells you what's it? Yeah. I mean, if we're going to pick this apart, and I think there are a couple of little nuggets here, one of which is the specific way in which they're showing the ads. So they show an example of a person asking, which is the cheapest Honda car? And you see the response and it says, yes.

The cheapest Honda car for 2023 is blah, blah, blah, according to Truecar. And then there's this little kind of superscript, a little blue box, and it has the word ad in it. And I don't know, I kind of thought this was a fairly nice, not too intrusive, but very clear indication.

I think it's just a cool moment. We should not lose sight of the fact that this is another fundamental change in the way we consume ads and that a whole bunch of micro decisions around user experience and interface design have had to be made in order to make this a reality. And I'm sure we're going to see a lot of iteration on this and we're going to be redefining what it means to serve ads on the internet through this kind of chatbot, but kind of cool to see the first manifestation of that here.

Yeah, it's interesting just to talk about this a bit more. There's a screenshot and basically there's a question of which is the cheapest and then there's a text answer. And then Bing does this thing of providing citations. So the citations here seem to be kind of the ad. So basically what looks to be happening is, right, if you Google or Bing at the top, you'll have a couple of links that are ads, right?

And being part of what it does is retrieve information from websites to answer a question. So what appears to be happening to me is they just retrieve that ad website and, you know, put the text in. So to me, it's kind of interesting, like, well, what is the return on investment for the people doing the advertising? You don't get a click, you don't get ad money. So...

I guess it's a good response and you get your answer out there, but is this what you actually want? Yeah, it's very interesting to see how they can...

make this stuff profitable. And you can just, you can feel that trade-off, like you said, between wanting to serve an explicit ad that you can get credit for making them click on and giving people content that they actually want. And it's, there's just this thing that happens. I don't know. I found myself wondering this as I read that screenshot, you kind of get this like increased cognitive load because you're reading what the chat bot wrote for you and you're sort of treating it like they're a salesman on commission. And you're like, ah,

"Okay, how much of this are you saying because you're trying to serve the ad?" It's already in a context where there's all this question about how accurate chatbots are to begin with. You're constantly experiencing this epistemic challenge where you're like, "What does this thing actually know and can I trust it?" I think something we'll have to see ironed out over the next few months.

Yeah. Yeah. It'll be very interesting. I mean, just thinking of this example, it's asking what is the cheapest Honda car? I could see myself asking, well, what is playing in the movie theater near me? And then the ad is just to switch the ordering of the answer to be like, list this movie first. And that's, I don't know, it feels a bit weird, right? That...

You're talking to this thing in a way, and you may not even be able to tell how its response is shaped by marketing, but it may just be shuffling around the text and emphasizing certain things and de-emphasizing other things based on payments. That's not even getting into reviews and asking about the qualities of products. So yeah, very...

interesting to see where this goes. So far, it's pretty simple. Next up, we have an article from TechCrunch that is about how YC's winter class is oozing with AI companies. YC is Y Combinator.

accelerator of startups. So this is where startups get some investment and basically get started. And yeah, it's showing how about 35% or 91 startups have something to do with AI, which is a pretty big fraction. It used to be much smaller. And I think it goes into some of the details of what companies there are.

Yeah, and I think it is quite noteworthy this is Y Combinator because just like some inside baseball for people who don't track Silicon Valley, like stupid politics and party gossip or whatever. But back in the day, at least back when I went through Y Combinator actually with my company, Sam Altman, who now is the CEO of OpenAI, was then the, I think he was the CEO of Y Combinator at the time.

And so he really, for 10 years, or 10-ish years at least, kind of set the course of Y Combinator, grew it massively, and had a huge influence on it. And so...

As a result, OpenAI enjoys this special, somewhat privileged relationship with Y Combinator. It may be less surprising that you find OpenAI's products may be preferred by the Y Combinator startups. Also, they're just amazing products, of course, but that's some inside baseball that's maybe relevant. Also, I should flag, YC does not tend to fall for hype.

I remember back in 2017, 2018, that was when I was there. You saw that it was the crypto craze was on full steam. People were going nuts for this. Investors could pour money into basically anything that had .eth as a domain.

But the YC partners were really level-headed. They backed some really solid founders, very infrastructure-level startups, stuff that was unglorious. I remember Quantstamp was one that became really big. And actually, in...

Funnily enough, I think in my batch, we had Devin Finzer and Alex Atala from OpenSea. And they were very level-headed, very kind of clear-eyed. NFTs were not a thing at the time, but they sniped it out. So all which is to say, I don't think this is a coincidence. I don't think this is reflective of just hype. I think, if I had to guess, it's that these tools are genuinely just that impactful, and the partners at Y Combinator see that.

Yeah, exactly. And this article also notes that of that 34% of startups that are AI, of that 20%, so 54 companies are in particular generative AI. So stuff like Jadgbt or Midjourney. And I think...

Yeah, I think it does make sense to me that unlike crypto, you can actually build something that's very useful fairly quickly. So there's a lot of sort of more niche applications. There's like SEC compliance, domain specific knowledge, finance, large language models, AI powered bookkeeping and finance, helping on-call developers,

All this sort of stuff. So all these different applications where AI can be a tool that you can leverage for whatever problem or domain you're in. Yeah. So it's not surprising perhaps, but definitely...

It just goes to show how big this trend is. It's also, I think, worth noting, YC had its demo day and its alumni demo day a couple of days ago, over the weekend. One thing that I did notice, we're talking about this idea that these companies tend to grow freakishly fast, the generative AI ones.

I, you know, looking through some of the presentations on demo day, I gotta say, holy shit. I mean, there was a company I remember seeing 400% month over month growth, which like,

If you're not used to investing in startups, a normal amount of growth that's really exciting is if you see even 20, but really more like 40% month-over-month growth, you're like, oh shit, that's really exciting at seed stage. We're talking 400%, 4X every month growth.

I don't know if that's sustainable. I think it could be, but there certainly are these fundamental issues where the landscape might shift under you and your fundamentals get wiped out that we talked about. I think this is such an interesting time to be watching the AI startup space. Yeah. And as a startup, it's an interesting time because all these non-AI companies are

are kind of being pressured by their investors to do something with AI, right? So at this point, you know, even if it's like a small startup, maybe the CEO is just like, well, we need to use AI. So let's figure something out. What's our GPT-4 strategy? Yeah. Yeah. So I can see how this growth is partially, you know, fueled by the hype. So it's an interesting time.

And speaking of startups, our next story is how Google partners with AI startup, uh, replete to take on Microsoft's GitHub. So it's a goal, uh, code completion, what, uh, GitHub co-pilot already does where it has this ghostwriter technology and, uh,

Yeah, I don't know. What did you notice about the story, Jeremy? Yeah, well, so Replit is actually like one of the darlings of YC. It actually came out of Y Combinator itself. I think it was backed by Paul Graham himself, who was the founder or one of the founders of Y Combinator itself.

and has just outrageously quick growth. I mean, okay, outrageously quick. ChatGPT got to 100 million in two months or whatever, so let's calm down. But, you know, they got to 20 million users very fast, and their users are very valuable. They tend to be developers and people of developer flavor. So it's kind of like, you know,

away getting into competition with GitHub. It's designed to make it super fast to basically spin up a programming environment so you can get to coding really quickly. They have this ghostwriter that they're developing, which is their response to GitHub Copilot. Just like to zoom out for a second, we're starting to see this tech access form where you've got Microsoft-backed companies on one side, and that includes GitHub and it includes OpenAI.

And on the other, we're seeing Google-backed companies. And that includes Anthropic and includes Cohere, which are building large language models, but it now also includes Replit. So Replit seems to be the Google Axis' response to GitHub on the Microsoft side. Everybody's got to have their own version of the damn thing. So here what we're seeing is Amjad Massad, who is this really, really capable CEO. I mean, at least I'm a...

Very impressed with him personally, just based on what I've seen from Twitter. And yeah, he's seeing the full potential of AI, very much an accelerationist

And we're seeing... He's talking about how Google is going to help them compete, essentially, with GitHub and with GitHub Copilot by using Palm and Lambda, some of the heavy-duty models that Google is working in the backend, and focusing them on code generation. So, yeah, I think... Anyway, really important player. I think this is one to keep an eye on. You're going to be seeing a lot more of Replit in the near future, I guess. Yeah, and it's an interesting point, I think, that this is kind of a competition because...

You could think of this as one of the examples of why competition is good. Microsoft has Word and PowerPoint, and Google has its Google Drive products. They both have data storage. They have email. They have all these things that both of them have their own version of. And now they both are going to add support for AI features as quickly as possible to the benefit of its users, I guess, right? Yeah.

Yeah, exactly. If you're a fan of automatically generated code, this is just good for you. It's also interesting to see how different this might be from Copilot. We don't know yet the kinds of decisions that are going to be made here. Also, one open question is GitHub was acquired

by Microsoft wholesale, right? They bought the whole company. Replit is partnering with Google. So now you've got these questions about like, okay, what amount of strategic risk is Replit going to incur by basing their product now increasingly on just Google products?

on potentially sharing their data, what kind of data sharing is going to happen there? Is replet data Google data and vice versa? What does that look like? There's a lot of stuff that we don't know based on this agreement. And the difference between an acquisition and a partnership starts to become really important here. And anyway, it's kind of cool to see Microsoft and Google trying to get their hands all over these companies and throw them at each other so dramatically.

Yeah, it's almost like the TV show Succession, but the companies are just all these moves and it's like a game of chess going on.

Onto the last story, we have a birth of a salesman. OpenAI sheds its lab coat to seek big deals. This is from the information. And this is a bit of an overview with various interviews from a reporter here covering how OpenAI has turned into more of a for-profit company, in particular because they started doing sales starting last year. So it used to be they were pretty much research companies.

Then they started offering these APIs for GPT-3 where you could pay to get access to these models. And now they actually have employees to go and get customers, right? Which is what you typically have in software companies. And it goes into some of the major customers they got. So it covers how Khan Academy is now doing.

paying quite a bit for access to ChatGPT. It says how OpenAI is now on pace to generate hundreds of millions of dollars in revenue, apparently, according to some source, which is pretty impressive, I would say. Yeah, so it's a pretty interesting article with various details of the sort.

Yeah, I think it's also going to be interesting to look at what do price points end up looking like when you commoditize intelligence like this? For all of human history, intelligence has been this limiting resource, and we don't yet know what that looks like at civilizational scale because we can't automate all intelligence. But certainly specific tasks like tutoring or teaching, those seem to be low-hanging fruit, and for better or for worse, they look like they're very much...

in the crosshairs of at least this generation of system. Yeah, I think it's also kind of interesting to note the amount of money that OpenAI is already starting to make. Because when you talk about hundreds of millions of dollars, you think about how much GPT-4 costs to train, right? Estimates vary, but it's certainly in the hundreds of millions is usually what you hear. So what we're now talking about is OpenAI getting to the threshold where they can start to self-finance.

And I think this is consistent with a little while back, we talked about Microsoft buying a 49% stake in OpenAI. And we talked about why that's an interesting number. Why 49%? Well, if you're OpenAI, you're not going to give away another percent. You're going to lose control of your company. And given the mission that OpenAI sees itself as being on, it's not going to be keen to do that.

That's consistent with a move that says, "Hey, we think this is the last fundraise we'll ever have to do." This aligns with that too. Now they're bringing in enough money that they can self-fund their next generation of big AI system. Maybe another indication that open AI is hitting that point of economic takeoff, like escape velocity. It'll be interesting to see if that holds up more broadly around the ecosystem.

Yeah, exactly. And, you know, I think we just discussed how there's all these YC startups. There's really no other API to use something like ChatGPT or GPT aside from OpenAI right now. They're the only player. We may have competitors later this year with Google and possibly Amazon, maybe Anthropic.

But they do get a lot of this kind of baked in revenue from just all these startups that are starting to make profit themselves and pay for this stuff. So, yeah, they're going to be making billions soon, I guess that's probably fair to say.

Onto research and advancements stories. First, we have introducing Bloomberg GPT, Bloomberg's 50 billion parameter large language model, purpose-built from scratch for finance. So in case you don't know, GPT is...

stands for generative pre-trained transformer. It's like a name for a neural network that OpenAI has been using since 2018. And now we are seeing all of these variations on GPTs, which is for different domains. I think we've discussed medical GPT last week, and now we have Bloomberg GPT, which is for finance. And yeah, there's a bit of detail how it's basically

all about finance and is, as we might expect now, super good at all this stuff and is actually being integrated into Bloomberg's products.

Yeah, I'm old enough to remember back when machine learning papers were exciting because they were an algorithm breakthrough. And people were like, oh, look at this algorithm. Look at what I did. And it was so clever. And you would dive in and you'd be like, oh, I'm so excited to find out how they did this clever thing with the algorithm. And now it kind of seems like the innovation is all in the data. It's just like, yep, we threw this like

giant pile of highly proprietary curated data into the same old stupid system that we've been building over and over. That seems, roughly speaking, to be the case here, where they literally just used the code even that was used to train Bloom, which is this open source trained language model that came out a couple months ago. They make a 50 billion parameter

version of that and they train it on this mixed data set. So I think one of the innovations here is they've got a data set that's one part kind of natural language, general purpose stuff from all over the web, and then one part finance specific domain specific.

And what they seem to find exciting about this is that this model manages to perform really well, both in general purpose reasoning, without losing its finance kind of edge. So it's able to achieve something like state-of-the-art in close to both domains, which...

It kind of is this interesting middle ground, right? We've talked in previous episodes, like you said, about the difference in strategy between generality, making a model that can do everything decently well, versus making specific models, like MedGBT, I think it was called. I'm trying to remember. But it's specifically trained for medicine and can't do anything else. This maybe is a bit of a middle ground, and it all comes down to the training data, or at least it seems. Yeah.

Yeah, exactly. I think last week we talked a bit about how there seems to be this emerging template on how to engineer these systems. You really don't need to think too hard about all the specifics. You take kind of a standard approach and it's probably good.

So this is another demonstration that more or less, they didn't do anything too interesting here. They just put together a model and showed that it worked with this new focus, new data set.

Do you think, by the way, Andre, as a question to you, as a fancy Stanford PhD in AI, what do you take home from the fact that we're seeing everything be GPT something? Even if you have to especially train it on domain-specific data, what do you think listeners, if this is the

They've not really thought about this aspect. What do you think is the take-home message of GPT for everything seems to be the solution here? I think it's kind of exciting. There's been excitement about this in academia and research for a few years, I think. People were already thinking, oh, transformers seem to be able to do everything.

And, you know, opening, I started doing GPT because back in 2018, that was kind of all the rage, doing language modeling with transformers and making them bigger and bigger. Like, they did not invent that. They jumped on the hype train in a way. So I think in general, AI researchers find it pretty cool that this one model type seems to be able to do everything. And I think in terms of research,

people were kind of already, there was a lot of criticism on the type of research where you just want to get the best number on a certain benchmark.

And so my take as a researcher is that if we are no longer putting in all this work on tweaking a model or collecting a data set or whatever, that might be a good thing for Academia because now the research will be more on things like safety or interpretability or user interface or things that...

We need, right? As we start to deploy these systems, we don't need better performance. We need better understanding. And that's where academia can really step in and do the things that industry is not so much incentivized to do.

Cool. Yeah. I always find it interesting, especially here academics with that perspective, like what can academia do here? Because that's obviously the perennial discussion as these things start to enter $200 million price range territory to train a scaled model. It's like, okay, how is a little academic lab going to do relevant research? Yeah, I totally agree. I think safety is a really interpretability, that sort of user experience stuff starts to get more and more relevant.

Yeah. Another thing to note about this I find interesting is they didn't just do a press release for this Bloomberg GPT, they released a full on paper.

And they plan to release training logs. And this paper is very detailed. It has the details of the data set, the model, the evaluation, of course, but also it has sections for ethics and openness and a whole appendix about the architecture. So it's kind of interesting that Bloomberg decided to

release such a detailed paper. I don't know if this was because they are trying to get researchers on board and recruit talent or what motivated this move, but it's kind of cool to see that we can actually know what's going on here, unlike GPD4. Well, yeah, to your point, I think it's a data point that at least suggests that the state of the art in the top hedge funds

in the top investment banking firms is actually going to be well beyond this, right? Because you simply do not give away the true cutting edge techniques or tools as a firm that makes money based on your information advantage over other firms, based on your superior ability to model markets.

And so I think if anything, this gives us an indication that Bloomberg thinks, okay, like our models are so much better than this, that it's fine for us to incur the risk of being pretty open about what these systems are and how they work.

Yeah, and this, I guess, is also, to be clear, sort of assistive in a way. It's built into terminals to help users of a terminal. And, you know, Bloomberg kind of has a monopoly on these terminal things for investors, so they don't have much to lose on sharing details. That's true. Date over models seems to be the trend too, yeah. Yeah.

And onto the next major story from TechCrunch, we have takeaways from Stanford's 386-page report on the state of AI. So the...

2023 AI Index Report from Stanford's Human AI Institute just came out. And this is a yearly report that's been coming out for a few years, maybe since 2020. I don't remember exactly. That just is sort of a summary of what's been going on and a picture of the landscape and industry and academia and society, just a whole lot of stuff. And yeah, so there's a lot of

to take away from this. Some of the stuff that's highlighted is, of course, AI development is more and more industry led in terms of, you know, obviously the larger models and more impactful models, but also in terms of research industry has increasingly been doing more of the publications. In fact, they also have an interesting bit here on how

The number of AI incidents and controversies is growing, which is a kind of fun thing to track. But there's a bit of a lag. So their data goes up to 2021. We don't see last year or this year. So yeah, they show how there's a bit of an exponential trend since 2010, 2020.

2012, I think when they start, there's like 10 stories or something. People weren't really that worried about AI. Now it's all over the news and so on. And yeah, there's just dozens and dozens of different charts. And the general picture is things are moving fast and everything is growing fast.

That is actually, there you go, saved about 300 pages of reading. Things are moving fast. Yeah, and everything is growing, it seems, except for policy responses. People are trying to come up with these coherent ways to grapple with what the hell all this means. We've got policy cycle times that are like...

years and years at best in some cases and technology cycle times that now feel like they're sometimes weeks. I don't know. Like, you know, on this show, it seems like we just keep going up the, up and up the curve. So, yeah, I mean, I think one of the things they're highlighting here is, is this, this issue, this risk and the fact that people are trying to write like

the definitive AI bill. We've seen that in Canada, by the way. We have up here in the great north of the wall, we have Bill C-27 and the Artificial Intelligence and Data Act, which has been very controversial, but it is trying to poke at a very important problem. And anyway, obviously the EU has their own thing, which we'll actually get to later today. But

Really interesting to see people try to grapple with what it means to have increasingly human level systems that can, I mean, GPT for anything. We just talked about that. Like, what the hell does that mean if you're a policymaker and someone can turn around a new capability like that almost overnight? Yeah, I do like regarding policy, they do have this one table that showed that

There is kind of a line graph of how many bills that are AI related have been proposed in the United States. And that showed that there has been a steady increase in the past five, six years going from basically zero to closer to 100.

And at least in 2022, nine of them passed. So there's some laws that are starting to come out. We've also seen some policymaking on a more regional front with Seattle and some states, I think Washington, passing different laws regarding face recognition. So yeah, it's a case where at least the government is trying to catch up. But-

Yeah, it's definitely the case that everything is moving very quickly. And they also have an interesting table of AI-related legal cases, speaking of policy, which same thing. In 2022, it's up to more than 100. From even five years ago, it was maybe 10. So...

Maybe not the general trends are maybe not surprising, but there's a lot of sort of interesting tidbits you can get out of it if you do actually read some of those 380 pages. But it's a very readable report. You don't need to be any sort of academic. It's very easy to just look into if you're interested. So as always, we'll have links to all these things in the description and last week in .ai. So feel free to follow up to dig more into it.

On to the lightning round. First up, we have robots using legs as arms to climb and push buttons. So this is a new paper from CMU called Legs as Manipulator: Pushing Quadrupedal Agility Beyond Locomotion. And it's kind of fun. So quadruped robots are these little dog-like robots. You probably have seen the Boston Dynamics one.

And they basically showed how, aside from just walking, you can make them do manipulation, which is to say they can push buttons or move balls or climb if they really need to. And there's some more details on reinforcement learning we used. But I guess the takeaway is if you want to integrate these robots into human environments,

If they can push a button to open a door, that goes a long way towards making them able to operate more flexibly. Yeah, I think it's also an interesting moment in the history of robotics where historically, we talked about expert systems before and human-designed, carefully coded hard rules that increasingly are being replaced by systems that are just trained end-to-end.

It seems with this particular research, we're in a middle ground now where we're not going to code in. Humans are not going to code in the rules for every little thing. We are going to use AI training, like machine learning training.

we're going to split the training up into little subtasks. And so here they have, you know, they mentioned having separate manipulation and locomotion policies that are kind of distinct. So not like one end to end process that you're training, but instead like have sub problems that you identify as a human. And in a way that's kind of like coding a little bit, some hard rules into the system, some priors. Obviously the way this ends up going, I shouldn't say obviously,

My suspicion is, you know, eventually, as we've seen in other areas, more and more of this gets subsumed into like one single big giant network. But we seem to be gradually moving in that direction robotics wise too. It's kind of cool to see. Yeah, I think we keep going back to like, oh, last week or two weeks ago. It's just all of this is so interrelated. Yeah, we discussed this more general issue.

robotics paper a couple of weeks back that was basically on that front where you have this general purpose reasoning thing, and then it was using a transformer that could do general purpose manipulation. So in time, I could see these mobile manipulators, things that can not just move things, but also move around. Eventually, they'll be part of this integrated system that will have some sort of hierarchy,

and not be quite so specific. But this is a step towards that.

Next, we have instant videos could represent the next leap in AI technology from the New York Times. So this is pretty much like an overview article on the trend of generating AI from text. We discussed the new runway model, I think, last week, and this includes some more examples from it. And yeah, just talking about sort of a trend and how

Most likely we'll be able to see high resolution videos that are AI produced and pretty hard to spot in a couple of years.

Yeah, and I guess, again, a story we talked about earlier, but that Pope Francis photo that was going viral on Facebook, they're showing that image generation has crossed that threshold now where people will actually kind of take it at face value. So we're, as a population, rather, we're going to have to start developing antibodies to this stuff. And, you know, image generation is already at that point. Video generation...

It kind of seems like it's coming, and that's what one of the MIT professors who's quoted in the article is saying here. With video generation, we're going to see that our ability to trust these systems is going to start to change pretty quickly. Anyway, another interesting data point. First they came for the text, then they came for the images, and now they're coming for the video. Yeah, and I guess one thing I'll say is

The video is maybe surprisingly not as good as you might think based on images. It's actually very wonky. You can go to the article and see for a few seconds. These are kind of very weird looking. So for now, at least, it's pretty far away.

And then we have in our article, can AI predict how you'll vote in the next election? So this is pretty interesting. A group of students at BYU tested the accuracy of if you tell a GPT-3 to answer a survey given a certain kind of demographics of race, age, ideology, they test to see how those

uh, personas, artificial personas would respond, um, to certain surveys and election elections compared to humans. And they found that, you know, even without directly telling the model, just by telling it, you know, act like this demographic, the responses of a model matched the response responses of a population pretty well. Uh, which is yeah, kind of an interesting result.

Yeah, from a malicious use standpoint and election interference type applications, you can really see how this could be useful. Getting people to mirror the style of people in a certain demographic, getting algorithms to figure out who to target as well. So that's sort of like something that the Center for Security and Emerging Technologies at Georgetown flagged a couple of weeks ago anyway, in their big report on where all this stuff was going.

But more sort of philosophically, I just find this sort of thing fascinating. Like if an AI algorithm can take the, just the, the raw kind of demographic data about you, you know, your skin color, your gender, your, uh, your, your net worth, all this stuff. And if it can successfully predict what your political views are, shouldn't you ask yourself some questions about the extent to which you're reasoning properly about the world?

I mean, okay, let's just say to be fair, right? You can make a decision tree that's pretty accurate. That's true. For that as well. That's true. Yes. There's a question of degree of accuracy though, right? As we start to hone in on like, you know, if a decision tree could predict my behavior, my voting behavior to like,

I don't know, within 20%, I might be like, yeah, you know, okay. But as you start to climb down that percentage ladder and hone in on a person's views with high precision, I'm not saying we're there yet. I just think it's an interesting question. What does this say about human beings? We've seen this play out with language, right? Where people often frame it as like, we're not learning that AI is super capable. What we're learning is that humans are less capable than we thought.

And I think this is an interesting dimension of that that applies to kind of rational judgment and political philosophy. I think there's a lot to chew on here just from a philosophical standpoint.

Yeah, I agree. I think in addition to this question of who you'll vote for, what is also interesting to me is that they looked into responses to a survey. And once you get into that, you're digging into more specifics of these individual issues. And it gets into that question of theory of mind, of how accurately can it imagine the way you think. And it seems like

Like without being trained to do it in any sort of direct way, just by being trained with a regular GP3, it can sort of infer theory of mind for demographics, which is definitely at least interesting, if not concerning. Everything's fine. Everything is interesting and concerning at the same time right now. That's just how it is. Just accept the duality. Yeah. Yeah.

And last story, we have another GPT. We have Hugging GPT solving AI task with Chai GPT and its friends in Hugging Face. So this is from the company Hugging Face, which is...

Pretty big presence in AI. Basically, they're sort of where you store your models and how you share them over worlds. And there's also forums and there's kind of a bunch of things, but they're quite mainstream, let's say, within AI developer circles. And this paper presents how you can essentially have some sort of user query

to do something with text or an image. And then you can have...

A language model like ChatGPT decide how to use the various AI models that are included in this Hugging Face repository to answer. So if you ask like, count how many objects are in this picture, instead of specifically telling you how to do it, you can have a ChatGPT say, well, for this text query, it makes sense to do this. And that's how you'll provide the reply.

Yeah, it's a bit of a, you know, I guess systems paper, but it does kind of suggest a way where you can now have these very powerful APIs where...

you know, you can support just general queries and this GPT will just pick how to do it and hopefully work, but maybe it won't, you know? Yeah. And it comes at a time when it seems like this is on everyone's mind, right? We talked about the chat GPT API that OpenAI launched. Like I have to say it, what was it? Two weeks ago? Like this was news two weeks ago. It feels like forever, but

but yeah, like this idea that you have a, a model of some kind that uses either tools directly. That was the chat GPT API. So it can use Expedia to book a trip or, or whatever. Um, and now we have models using models. And so, uh, you know, again, this question of like, uh,

what is the architectural evolution of this going to look like? Are we going to keep having models that use other models, which is kind of more of that expert systems vibe, humans at least kind of coding in hard coding wires, if you will, connectors between different systems, or eventually does this all become one big model? What variations on this theme are going to persist? I think it's just so interesting to see this through that lens as another point along that axis. And we don't know where that axis leads, obviously, but-

Yeah, definitely. So, yeah, it's a pretty cool paper and I could definitely see this, you know, becoming a thing that Hugging Face provides. I mean, Hugging Face is also a company that

needs to make money actually. So this is a paper, but I could see this being an API that he'll provide at some point. Yeah. And there's also like, when you look at the safety properties of a system like this too, kind of interesting design questions. Like if something goes wrong because the chat GPT model routed your query to a model, then screwed up the response. Like

What's to blame in there? Should the chat GPT model have sent it to a different submodel? Should the submodel itself have gotten the thing right? Is it the fault of the submodel? Anyway, tracing fault in these systems becomes also a stickier thing. And from a policy standpoint, the developers of those two models might be different. Maybe they perceive themselves as having different responsibilities. Anyway, it's a whole beautiful thorny thing, but kind of another interesting footnote there.

Yep. Well, let's move on to policy and societal impacts. And our first story is the Chad GPT King isn't worried, but he knows you might be. Chad GPT King is a pretty interesting title for the CEO of OpenAI. I could have just said the CEO of OpenAI, but okay. And yeah, this is kind of an overview article summarizing in a way

the views and sort of history of Sam Altman, who is, I guess you could call the king of chat GPT. And I guess if you don't know the history and the background of OpenAI and Sam Altman, this does provide a lot of those details that are interesting.

Did you find anything kind of new that you thought was interesting, Jeremy? A couple of little nuggets. But like, yeah, I wrote this in our... So we have a shared kind of notes document on this. And Andre expertly kind of puts it together. And then I add some notes and stuff. But one of the things that I noted at the very top was that this is like the most incredibly New York Times-y article that I've read in a long time. You know, those articles that start with like,

On a rainy day, in the middle of a dirt road, the carriage approached the... You know, it's like none of this is relevant to the story. It's just seems that there's so much... This is dripping with that. And I'm, you know, I'm here for it. It was entertaining as a read. There were a couple of interesting little notes here. So...

One is the duality, the fact that Sam A. has to hold, Sam Altman that is, has to hold in his mind all these things that might seem on the surface like contradictions. For example, he has to believe as he does that AI is going to change the world at the same time, and that that change could be incredibly positive. At the same time, he also believes that it could end the world.

And at the same time, he also thinks that current AI is potentially overhyped. And so like all these things that it's funny to see people react to this stuff on Twitter because they're trying to figure out, no, but what do you really believe? And the through line here seems to be Sam A believes all those things. They are not intrinsically contradictory.

But the things that make those ideas compatible involve a lot of nuance that doesn't necessarily come through in 240 characters. And so this was, I think, a useful dive for that reason into all of those factors. In terms of revelations, one thing that was broken by this piece was the idea that Sam Altman actually holds no equity in OpenAI. So at least to people who come from the startup world, that's really surprising.

reflects his commitment, let's say, to the mission over the, not over the company necessarily, but it certainly paints a complex story. And then they talk about his childhood and the challenges of growing up gay as Sam Altman did in St. Louis. That was quite interesting. So some of the challenges he had there.

Then there's this quote from Paul Graham. Paul Graham was the founder, as we mentioned earlier, of Y Combinator, who passed Y Combinator on to Sam Altman. A journalist asked him, why do you think Sam A is working with no equity? Why do you think he's doing this? PG's response was, why is he working on something that won't make him richer?

One answer is that a lot of people do that once they have enough money, which Sam probably does. The other is that he likes power. So anyway, that kind of like, you know, left the article hanging on this consistent note of ambiguity, which is really what the whole article is about. It's this like set of apparent contradictions and complex ideas. And I thought it was interesting as a read.

Yeah, yeah. It's somewhat lengthy and just browsing it a little bit more, I think, you know, if you think opening high school, then you probably would find Sam Altman at least interesting. He does tend to be pretty interesting to listen to, I find, just the way he expresses his thoughts. And as you said, this piece is...

Let's see, literary. Let me just read a paragraph here. Later, as Mr. Altman sipped a sweet wine in lieu of dessert, he compared his company to the Manhattan Project. As if he were chatting about tomorrow's weather forecast, he said the US effort to build an atomic bomb during the Second World War had been a project on the scale of OpenAI, the level of ambition we aspire to. And that's like a flashback to three years ago, just setting up a piece. So...

The take home there is that Sam A. skips dessert. That's what I got out of that. Anyway, yeah, it's a cool piece. And I think, you know, if you really are interested in chat, knowing more about OpenAI probably does make sense. Next, we have the European Union's Artificial Intelligence Act explained, which is an explanation of the EU's effort to

create this very, very ambitious AI act to do regulation throughout the European Union that is fairly comprehensive and is really beyond what anyone has done so far. I mean, it's comparable to maybe to Canada, but that's about it. So,

Yeah, if you aren't aware of the act, we'll summarize it a little bit. It's quite complicated, as you might expect, and this article does dive into that. But at the high level that we can provide, the idea of the act is that you will take different applications of AI, you will...

categorize them as having you know some risk you know no risk some risk high risk and then depending on the risk there's different requirements of how much you know auditing you need to do how much transparency things like that and obviously there's a lot of questions on how that would work and so on and this has been in the works for years so it's it's a fairly you know

ambitious effort that seems to be coming along.

Yeah, and I think one of the challenges too that they're facing in doing this is just like defining their terms. You know, like this is an act, it's legislation, right? That the EU is going to pass or in some form. And once you define that legislation, then the question is, okay, so you've said that, for example, as they write, you know, AI enabled video games and spam filters are a minimal risk application of AI that therefore has no restrictions on it.

But what do you mean AI-enabled video games? What if I have a video game with properties X, Y, and Z that makes it actually kind of dangerous? So what's actually going to happen, I suspect, is they're going to have to punt the actual implementation and definition of these terms to the regulators themselves, the actual agencies that are going to have to figure out what in practice they do about this.

It's not clear to me how much of the original vibe of the thing is going to persist at that level. That's going to be a very tricky translation task, but it's going to be important for everybody to follow because, in part, the regulation and legislation that comes out in Europe is known to trickle out into the wider world. There's this thing called the Brussels effect, where

where the EU passes a thing and basically because everybody does business in Europe, Facebook, for example, can't just ignore European regulation. All of a sudden, the floor on everybody's activities on their compliance has to go up. No matter what, you're busy complying with Europe's GDPR regulation or something. You have to do that in Europe, so you have to do that period. You need that infrastructure.

So in that way, Europe does tend to have an outsized impact on policy. It is really important to follow what happens there because it inspires policy responses elsewhere too. So don't think of this, I guess, as just a thing that's going on in Europe. It's something that is going to affect our experiences online probably reasonably soon. It's just an open question as to how.

Yeah. And on the point of defining things, I think one of the easy sort of criticisms of regulation that many might have immediately is like, well, this is technology and politicians don't know technology, right? But

I haven't looked into this in the past. In fact, I think Yann LeCun made this point at some point on Twitter, just like, how do you define AI? How do you measure risk? How do you do this or that? How do you categorize? And we actually just published papers on this where academics in Italy, I think, wrote a whole story on how you might be able to quantify what is an AI system, how you can try and

you know, a value to level of risk, et cetera, et cetera. So to me, that was a good sign of this is being very carefully kind of considered and it has been years and works. And, um, yeah, I think I'm, I'm kind of optimistic because we, even aside from chat GPT and all these various things, you know, we have self-driving cars, we have facial recognition. They already are part of society and having widespread effects. Um,

And VUS is not regulating it very much yet. So hopefully this sets a template that others can follow. Well, and to your point, I think there's this interesting question of whether people will often say bad regulation is way worse than no regulation at all. In some cases, that's certainly true. It certainly depends on the way in which the regulation is bad.

But in this case, I think we have systems that are changing the world right now. And we know by default that if no regulatory attention is paid to this stuff, things are probably going to go pretty badly sideways. And then we get to play the game of trying to guess exactly how the world gets fucked up. But if you believe that you'd rather not explore that possibility, some kind of regulation is going to be required. And if you only think of this as a test...

then at least we're running that test. So maybe on a more bullish on regulation perspective to end that comment on.

Yeah, yeah, exactly. And I think, again, you know, having been doing this podcast for like three years or something, it's worth pointing out that it's not just about the future, it's about the present. And there's a reason this has been in the works for years. It's like AI is already here. It may seem like it's just starting to arrive if you haven't been keeping up with it and you mainly just, you know, got interested through chat GPT, but it's...

everywhere and in various ways. So, we need at least some regulation. I think we can probably agree on that. And onto the lightning round, speaking of regulation and policy, there's an article from the Washington Post titled, "How a Tiny Company with Few Rules is Making Fake Images Go Mainstream."

So the company here is Stability AI, the maker of Midjourney, which is one of the major text-to-image services similar to DALI. And there's really only a couple of other ones that are leading the pack. And this is a bit of a summary of the company and in particular, its stance on policy, I guess, or moderation. And there's some interesting tidbits here where, for instance--

You can generate images of President Biden or Putin or other people, but not China's president, Xi Jinping. And the company's founder and CEO, David Holes, in talking about this last year on Discord, said that they just want to minimize drama and political satire in China is pretty not okay. And so they want to be able to...

have people in China use this tech than their ability to generate satire. So the overall vibe that goes through this is that they haven't grappled very much with, you know, moderation and general policies, it seems like. Um,

And yeah, now they actually have discontinued the free tier recently because there's been so much abuse and because you have had these viral things like, you know, President Trump being arrested that we discussed last week, which some people I'm sure got very angry about. So yeah, it's an interesting article that kind of gets into a lot of this policy stuff and, and,

Yeah.

And what usually happens with companies that grow slowly, historically, right, it like took seven years to go from seed to like IPO or something like that, roughly plus or minus like 20 years. But anyway, so historically, that would be the case. So founders would get more and more experience gradually climbing that ladder. You know, they get beaten down, they get humble, they sort of like learn, oh, actually, the world is more complicated a place than I thought.

And now what we're seeing with these outrageous growth curves for these companies is you've got a couple of yahoos, like Stability AI or MidJury or whatever these companies, just getting propelled out of nowhere into the front pages of everything. And they haven't been polished or they haven't had the time to kind of chew the fat on this and think deeply about the philosophical implications of this stuff yet.

And the view from up there looks very different from what it does when you're just starting the company. And if you started the company like six months ago and now you're already up there, it's like you got to adapt pretty fast. So I think this is a sort of interesting almost culture of company story where we're seeing people who were not necessarily fine-tuned to be in this position have to perform it.

Yeah. And it is kind of interesting also this article highlights that this company is very tiny. It is like 10 people. It's the CEO and then there's an engineering team of eight and some advisors, but it's having a giant impact and it's a pretty small set of people that are making decisions as to what's allowed and not allowed and so on.

which is pretty different from say open AI, which has been at this for years. And going back to GPT-2 was, you know, at that point, very careful about potential misapplications. So it's, yeah, it was an interesting read to see behind the scenes of this AI to image tech and how it's being shaped maybe a little bit ad hoc. And next up we have FTC is reviewing competition in AI. So FTC,

This is from FTC chair, Lina Khan, basically saying that they are paying close attention to developments in AI and are supposedly trying to ensure the field isn't dominated by the major tech platforms. But I don't know, Jeremy, I feel like... Mission accomplished. Yeah, I'm pretty sure this is going to be an oligarchy situation myself.

Yeah, no, I completely agree. And I think that it's not, look, there's some problems that competition is great at resolving. Those problems usually look like problems of accessibility. So, you know, you want mass production of a good, you want commoditization and cheapness of a good, have competition crash the cost of that thing, make it a commodity that everybody can pump out.

We've definitely seen that happen with AI as more and more competitors have entered in the space. We've seen open AI have to cut their prices and so on. So, okay, good for accessibility. But if you look at it through the lens of risk, in particular, if you look at it through the lens of catastrophic risk, you know, catastrophic accidents, alignment failure, that sort of thing.

It's not necessarily the best thing in the world for there to be 20 companies, all with the capacity to build human-level AI in a context where we don't know how to control those systems and where safety researchers specifically, not AI researchers generally, but safety researchers specifically at the world's top labs have this consensus that, yeah, it

So it's probably not less than a 10% chance that these things destroy the world. So, you know, when you're looking at that, like, I definitely appreciate the FTC's perspective here. They have to look out for consumers. And that means looking at prices. And that's what their culture has wired them to care about.

But there is another dimension to this that may not fall within their mandate. And this is where I think a lot of these different departments are going to be kind of stepping and agencies will be stepping on each other's toes. Because what the FTC says is like, yeah, we want competition. We want proliferation, democratization of AI. Right.

But if you look at security-minded organizations, they'll be like, no, no, no, we need to prevent proliferation. There are malicious use and accident risks here. So I think it's just going to be an interesting thing to see where the government ends up settling if they have time to settle, given how fast curves are moving. But yeah, definitely an interesting kind of chess piece on the board here with the FTC's move.

Yeah. And actually, the next article is very relevant. It's from the MIT Tech Review. It is Three Ways AI Chatbots Are a Security Disaster. So it's more of an overview article. And it says, for example, it explains jailbreaking, which you may or may not know. Basically, it's getting these chatbots to do things they aren't supposed to do.

through some clever inputs. But there's some other ones I was less aware of. So there is assisting, scamming, and phishing. So phishing is pretending you're someone else. And with AI, it's a lot easier. We already touched on it. There's also a thing called hidden prompt injection. So you can put something in a website to mess with an AI and make it do something you don't want it to do. That has been demonstrated.

And another thing that is described as data poisoning, where you can put something on the web for the company to scrape and put in their training set. And maybe that makes the model give you money somehow. So yeah, I found this to be a pretty interesting overview that actually taught me a few things I wasn't aware of myself.

Yeah, it's really cool to see this, the data poisoning one especially. There was this paper, I think this is the one that's actually linked here. Yeah, yeah, yeah. It's a paper by a bunch of researchers at Google and NVIDIA and Robust Intelligence, which has, anyway, done some work in the space. And yeah, for 60 bucks, they're able to

let's say, buy domains and fill them with images of their choosing, which were then scraped into large datasets. And they were able to edit and add sentences to Wikipedia entries that ended up in an AI model's dataset. And so just for the cost of

to fuck with these models is collapsing and the capabilities of the models is increasing and they're proliferating. And it's just really interesting to see like what dependencies are we going to end up with in our brave new world that we're building where, you know, there's...

I think there's this meme of somebody shows a super complicated and valuable thing like, I don't know, ChatGPT at the top. And then it's supported by an elaborate, complicated structure. And at one point, there's this razor-thin support that's like a few lines of open-source software or something. Well, in this case, it's sort of similar. If you make that razor-thin support maybe like data integrity or scraping reliable data online,

people can mess with your model. And yeah, I'm really curious what this leads to. I mean, there's so many possibilities, it's almost hard to imagine. - Definitely. And to me, I think we've discussed it already, but I'd be really curious to see to what extent phishing is impacted by AI this year, where especially with generative voice, but also with text,

you could easily see how, you know, foreign scammers who do all this phishing via email and via phone call can suddenly be much better. And if that happens, you know, tons of people are going to lose money. And at that point, politicians have to care, right? Because people are being harmed. So maybe VAT will kickstart at least some policy changes sooner rather than later.

Last up, we have another kind of overview story, also from the tech review. ChatGPT is going to change education, not destroy it. So this is a...

conversation summary from different educators about sort of evolving viewpoints of educators on how ChatGPT might become part of education and really shape it. So it touches on a few things like how some tech companies such as Duolingo and Quizlet, and you mentioned also Khan Academy,

are now integrating ChatGPT into their education. And personally, I think it's really exciting. For Duolingo, you can practice talking with a chatbot

And OpenAI has also worked with educators to put together a fact sheet about how chat GPT can be used in the schools and create a tool to spot text written by a chat bot. So lots of little different details here that I think now we are starting to calm down a bit and not freak out about AI written essays and like the end of education or whatever. It's just like education needs to change and evolve.

Yeah, yeah. And I think the, you know, the dust is, well, ordinarily, the dust would take like years to settle on something like this, you know, when the calculator was invented, we had to kind of fuck around with it and find out. The weird thing here is I don't think we have time to fuck around and find out. I think we're gonna have to fuck around as we find out.

And as more systems come online too, right? Like it was ChatGPT four months ago. It was GPT-4 like one month ago. What's coming out next? And I think the new normal is going to be maybe at a meta level, just like what new tools have come out and how can we be more productive and teach ourselves better with those tools? I don't know, but that's like one angle that I imagine these things reduce to just because you don't have the option to just...

Even thinking about lesson plans. You're a teacher. You want to teach relevant tooling. I remember back in the day, we'd all get dragged. The grade six class goes into the computer lab, and we're talking about, here's how you work with Microsoft Word. Well, that fundamentally has to change now.

You cannot have a lesson plan that a teacher prepares six months in advance, even, or three months in advance, and hope that it'll continue to be relevant. It calls for more plasticity and all kinds of other things, and it could, if it's done well, it could lead to incredible benefits. Right?

Right? Like, I mean, the Duolingo example you cited, like imagine that on a rolling basis. That could really be amazing, personalized education and all that, but we've got to be set up for it and have the right sort of philosophy and attitude approaching it. Definitely. One tidbit I found interesting in here is they noted that there's a US survey of 1,000 K-12 teachers and 1,000 students.

And it found that more than half of the teachers have used ChiaGPT and a lot of them, 10%, used it every day. And actually fewer students, only a third of the students use ChiaGPT in the ages 12 to 17. So I think kind of another thing that's worth noting is aside from the student side, I think this will be

a great tool for teachers to make their lives easier, right? Because teachers, that's a hard job. You overstretch, you're underpaid. And it's kind of easy to imagine how something like this can just make it easier to teach and teach better.

And onto our last section, art and fun stuff. Starting with this article, Chinese creators use Midjourney's AI to generate retro urban photography. So we're just talking about Midjourney. Here is more of an artistic application of it.

where, you know, there's a bunch of images. This is audio, so we can't fully describe them. But basically, it's sort of, if you can imagine, generating photos from a few decades ago of like,

weddings out in a rural setting or, you know, minors at work or railroad workers and so on. So this article showcased how several different people have sort of made it their hobby or practice to generate these types of things. And they look very realistic. I mean, as you might imagine, but still just seeing them, it's kind of impressive. Yeah.

Yeah, and it's again that story of the data being the interesting differentiator to some degree. The Chinese, I remember when GPT-3 came out, one of the big questions, almost nationalistic questions in China was, how can we get a Chinese language version of this so that we can use it internally? Obviously, a whole bunch of national security and more defense-oriented stuff.

technology advantage stuff went into that decision too. But just like generally, this desire to have a homegrown version of these things to reflect the culture and the interests of that country. And I think that's what we're seeing here, right? Like so many of these images are clearly informed by priors, by data collected that

carries information about what it was like to live in China at that time. I don't know how this data would compare, for example, to the data that Midjourney was trained on, but plausibly that's part of what's going on here. Yeah, just interesting how art starts to depend on the data set that's collected as well. Yeah, and the other thing I would note in this that I found pretty cool is these images...

At least some of them, sort of what the article starts with, are by Jang Hai-Joon, who is a street photographer in the city. And so, this street photographer got the stool and had this interest of going back to his childhood, essentially, where he grew up and saw this...

these miners and weddings and things like that. And now you can't go back in time and take photographs, but you can use mid-journey to sort of generate your memory of that place. And as a street photographer, it's kind of maybe not too hard to imagine why that's appealing. So from an artistic angle, I think it's interesting, this whole conversation of

AI art and is something like mid journey a tool and is it a human artist? In this case, you have a street photographer using this tool to create essentially street photography in the past. And I think obviously kind of having that aesthetic taste plays into what these look like. Onto the last story, we're going to end with something a little more silly and

The story is AI deep fake video of Harry Potter characters in faux Balenciaga fashion show go viral. So it's about how a YouTuber used AI to recreate characters from Harry Potter, but make them look as if they're part of a Balenciaga fashion show.

And this YouTuber actually has been doing this for a while with various videos that make use of AI to create these kind of like sort of

weird mashups, I suppose you could say. And this one, I guess, just went viral. And you can click on the link and see how you have these characters in these very model-looking stances, and their faces also look really model-like. So it's a bit uncanny and kind of neat.

Yeah, yeah. Another one of those, those deep questions about how much of creativity is just combining ideas together that already existed. This is sort of like the ultimate embodiment of that, like, give me Harry Potter as a fashion show and boom. But yeah, it does. It does look like

I can even see the Ron Weasley version seems like something out of an 80s music video. It's like this weird aesthetic thing. But yeah, fascinating to see in another instance of these image generation systems making very entertaining products. Yeah, and it's interesting. We keep discussing these things that go viral and...

It's like, at what point does... Is this still novelty that these things look so real or... Right. You know, at what point is virality just the idea itself and not the uncanny nature of, look, we can do this, you know? And I don't know. Yeah, I guess for now, it's still kind of just... Still to me, as someone who's been around this stuff for a while, looking at these images, it's kind of crazy.

to know that you can just input some text and you get something this realistic looking. Well, that's it for this episode of Last Week in AI. Once again, you can head on over to lastweekin.ai to find the links to all of these articles and to get our text newsletter with even more articles we don't get to talk about.

If you enjoy our podcast, please do share it with your friends or on social media or with your parents, whatever you feel like. And we do appreciate if you review it on Apple Podcasts so we can take your feedback into account. And yeah, we'll be back next week. Be sure to tune in.