#110 - We’re back! ChatGPT, ChatGPT, ChatGPT, and some other stuff
01:13:03 Share

Hello and welcome back to Skynet Today's Last Week in AI podcast. We can hear AI researchers and practitioners chat about what's going on with AI. As usual, in this episode, we all provide summaries and discussion of last week's most interesting AI news. You can also check out our Last Week in AI newsletter at lastweekin.ai for articles we did not and did cover in this episode.

And we are back after four months of hiatus, a bit longer than expected, and we have a new co-host. Woo! Hey, everyone. So my name is Jeremy. I guess some of you might know about Taurus Data Science and the Taurus Data Science podcast. So that's something that I hosted a while ago. I since started working on AI safety stuff.

And and then kind of went off and did my own thing. And Andre, yeah, came to me very nicely and said, Hey, we're we're looking at fire this thing up again and asked if I join in. This seems like a really exciting time to be doing this on top of everything else. I think it's a really cool podcast and cool project. But man, there is a lot of stuff to talk about. And I think it's increasingly important for not just like AI people, not just the like techie nerds like us to be following this stuff, but like for everyone.

And so, yeah, I'm just super stoked to be having this conversation, to be lined up to track this stuff together as the tech evolves. Yeah, it's great to be back now that chat GBT exploded and now there's more news than ever. And you may remember Jeremy actually was a guest in a couple of our episodes. So I think it's a really cool setup to have him back.

And in case you don't remember, I am Andrey Karenkov. I'm a PhD student at Stanford. Almost done, a couple more months. So yeah, it's great to be back. And let's just go ahead and dive back in to the news. As usual, we start with applications and business. And our first story is about how BuzzFeed says it will use AI tools from OpenAI to personalize its content.

So in a memo, Buzzfeed said that it will use things by OpenAI and OpenAI, of course, created ChatGPT to create kind of AI-based content. And that involves enhancing the quiz experience, informing our brainstorming and personalizing our content for our audience, which I think is pretty interesting.

They did say that they will not be using AI in its newsroom. So this is mainly for, I guess, lower impact kind of stuff where errors don't matter quite as much. And related to this, we'll also mention that before this, there were articles about how CNET was publishing entire articles generated by AI kind of in secret until it came out.

And then there was a bit of a controversy about how I did this and how I published about 70 articles with this T-Net money stuff headline.

And then more than half of these articles turned out to have errors. And these articles were things like what is compound interest? How much should you keep in a CD? Does home equity loan affect private mortgage insurance? So a kind of information you want to not get wrong. And yeah, so this was a bit of a mess that CNET created.

So can I ask you about this BuzzFeed story? Because I think it's, this is probably really important. The CNET one is too, but because these are precedent setting moves, right? Like I don't think anyone's under the illusion that this is not going to accelerate as a trend. And so when we look at BuzzFeed, you talked about they're not going to use it in the newsroom. Do they talk about specific examples of, are we talking about content curation or are we talking about content generation?

Yeah, so this was a memo that didn't get too much into the details. There were some examples where Buzzfeed has many of these quizzes. I don't know which Disney princess are you most similar to or which Harry Potter school would you be assigned into? Hidden journalism. I know, yeah. So I think it's going to be more fluffy things like that initially, but I could imagine, of course, they'll be experimenting and trying out various things as it gets going.

And BuzzFeed stock actually jumped a bunch. I think I saw 50% jump in its stock. So it's clear that, as you said, this is a big move and investors are excited to see this happening.

So I think we'll pick up on the investor through line a little bit later because there are a couple articles that point in that direction that are really interesting. Do the investment dynamics that held with startups in the past continue to hold with generative AI companies? I think that's a really interesting dimension to go into. But when we look at this BuzzFeed thing, so the BuzzFeed obviously is like a news company. That's how they've always been framed.

I'm really curious strategically whether this starts to position them as more of almost a gaming company. Like we've seen products like AI Dungeons come out that use like GPT-3 to give a kind of, what's that old Wild West game people used to play just by chat interface with their old computers? Oh yeah, Oregon Trail. Right? So like those kinds of, but like dynamically personalized and kind of that sort of experience.

I'd be curious. And we've seen Netflix play around with gaming as they look at their model, their movie model, not being quite enough. I'd be curious about whether this portends some kind of new direction for BuzzFeed as well, but I guess time will tell.

Yeah, I think the most interesting aspect here is personalized content. And as you said, initially, it'll be kind of the same BuzzFeed content, but with AI. But I could see it coming or going in a more like back and forth direction where you actually make a choice. And then the AI dynamically creates new content, which will be a pretty different experience, like you said.

Wow. Well, I guess good job Buzzfeed, at least for the moment. CNET seems like a bit of a different story. Yeah. I'd be curious to find out how was this first discovered? How did people twig to the fact that CNET was using AI to generate content?

Yeah, so they were publishing these stories under this ambiguous headline of the CNET money staff. And if you clicked on that byline, it says that this article was generated using automation technology and thoroughly edited and thought checked by an editor on our editorial staff. So it was hidden and they didn't do this secretly exactly, but they did this quietly, was what the CEO said.

Did they get into what specifically that fact checking looked like or whether it was in fact applied as advertised? I don't think so. Yeah, they really might. Like many people in CNET didn't know this was happening or like the actual journalists. So it was seemed kind of a quiet experiment that they didn't intend to be a big deal. But they did say that they paused it for now, but they will be going back to using AI as well. And hopefully then they'll

It'll be more spelled out how they're doing it. Right. Well, and I guess that's another one of these things when you talk about corporate strategies. Like, you go in there, you're like, all right, we're going to start using, as so many companies will inevitably, we're going to start using AI-generated content. How do you break that news to your staff? Like, it's one thing to pretend to the world that you're not using AI-generated content. Mike, I

I would wonder if maybe telling the staff was one of the main reasons why they kept it under wraps. Because like the morale implications of that, the clock starts ticking on everybody's jobs. They all start looking for the exits and so on. And just I'm so fascinated by what those conversations would have looked like in the boardroom or among the strategy level people to figure that out. But fascinating.

Yeah, I think I would imagine journalists and writers of news are starting to be a little bit nervous, just like people did last year with visual art. So I'm sure it'll be a quickly evolving thing this year. Actually, next we have the story of who owns the generative AI platform, which is quite related.

That's right. Yeah, this was one I dug up. So I've been looking for an excuse to dive into the economics of generative AI. This, I would argue, this is the single most important technological story currently unfolding full stop. More important than quantum computing, more important than drone tech, generative AI as an industry, as they put it. So this is a post that was written up by Andreessen Horowitz. So obviously big VC firm, know a ton about investing, know a ton about big trends and are really good at getting in early on them.

And they were just breaking down the kind of generative AI ecosystem and asking this very provocative question of who are the big winners going to be? Because we know looking in other industries, the winners, they don't always appear where you think they would.

So if we look at, for example, computing, right? Like back in the day, people assumed, well, I guess IBM is going to be the big winner in computing because they're making the hardware, the machines, and the hardware is what you sell and kind of end a story. But then along came Microsoft and they showed actually what you want, what you end up doing in this industry is you commoditize the hardware. So there's this like race to the bottom on pricing for hardware and

All the margins end up being in software. So Microsoft sells tons of Microsoft Word, Microsoft Office products, games, things like that. And the alpha there, the returns are massive. So it turns out that you got to parse things with a pretty fine-tuned scope to figure out where exactly the revenues are going to accrue in the stack.

And what they do in this article is they break the stack down into three parts. So the first one, actually, I'm just going to, I'm just going to scroll down to my notes here because I want to make sure I get all this juicy stuff. So the first one was the apps themselves, right? The generative AI apps that we consume. So think here of going to chat GPT, just that webpage that's serving you this product. Think about going to, I don't know, mid journey or like

Any website where you can use the thing right away, that's the app. So there's one question that says, well, will app developers end up accruing all the value or will it be the model providers, right? Is it going to be, for example, like the open AIs? Is it going to be the coheres? Is it going to be the Googles ultimately? Things like that. Or is it the infrastructure vendors, the people who build the processor, the computing hardware?

And as they go through the stack, they're really interesting arguments to dig into. We're not going to have time to go into super detail, but the upshot is it really doesn't seem like there's a single point where there's defensible differentiation in the market. If you go at the app level, right? If I decide to make, I don't know, some kind of app that tells you what your age is from pictures and it's really fun and it goes viral, somebody else can replicate that really easily using probably the exact same model that I did.

If I go to stable the fusion, I go to open AI or whatever to build my app. Somebody else can do that too. Now, likewise, we're starting to see competition at the level of model providers too. So open AI, sure they have chat GPT. We've got deep mind that has Sparrow. Google has Lambda. Soon these are going to be out there. Just like we saw the erosion of margins for GPT-3, right?

Right? So there was first GB3 came out. It was the only kind of product that did any kind of human-like text generation, then AI21 Labs, then Cohere, and OpenAI has been forced to drop prices. And so there's an argument that says maybe there's less alpha there than we think. There might still be enough to make massive companies, and that seems quite likely, but are we going to see trillion-dollar companies? Are we going to see $10 trillion companies? Is there going to be another FANG company coming out of this? That remains to be seen. The final thing is with those infrastructure providers, which chips are hard, it's

Semiconductors are hard. It's a tough business from a CapEx standpoint. And a lot of new entrants too are challenging the traditional AWS GCP type players, Lambda Labs, and anyway, a couple others competing on like customization. So really rich article. I highly recommend checking it out if you're interested in the strategy of where all this stuff is going. Yeah, I agree. It's an interesting question. And I think it is like right now it appears probably that OpenAI is way ahead, right?

Just because they have released these user-facing apps and they have the opening API for a while that people have built various apps on top of already and have products. But I think it might be a little bit similar to what we saw in the cloud area where at first Amazon and AWS were the first we've seen and very kind of dominant.

But then Google created their cloud offering, Microsoft created their cloud offering, and now it's a little bit of a competition where different companies offer different benefits. It's a cost benefit analysis. And like you said, Google has this technology and they're going to start offering it. Microsoft is going to offer it through Azure.

But you could definitely see multiple companies getting around to this. And there are some moats to getting this up to speed. So hardware costs are very complicated, very large. And then it's actually not trivial to develop the computation necessary. So Google has discussed how for Palm, I believe, how it was very kind of sophisticated to really get

things working in this super paralyzed compute environment. So for now, it's not going to move super quick, but I do think like with cloud, it's going to kind of gradually become more of a space with a lot of players where there's no kind of clear winner and just a lot of different options that have different kind of benefits and costs and so on.

Yeah, I think that has as well. For those of us who are AI risk hawks like I am, this is really worrying, actually, because your investments in safety, right, as a cutting edge AI lab, they are like a tax. They come out of your margins, your profits. You literally are faced every day with this choice. Do I, for example, do I spend more money screening people before they can use my app like OpenAI used to do with GPT-3 until what?

until AI21 Labs came out with a competitor product that did not screen people. And then people started jumping ship going to AI21 Labs. And all of a sudden we find out, oh, look at this, OpenAI no longer has a screening procedure. What a surprise. And so seeing that kind of safety race to the bottom, both on access, potentially also on alignment is something that you start to think about it when you hear about DeepMind Sparrow and like you said, Palm and Lambda and so on. So

So a definite, I think, macro trend to watch from a safety standpoint too, if that's your cup of tea. But no, totally agree. It's anyone's guess what's going to happen next. Yeah. And it's already kind of playing out with generative AI for images where you do have multiple players.

And some of them have really strong safety safeguards for which data you use, what can you generate. Others maybe have less of that. And I think we'll see that sooner or later how that plays out. And actually related to this story, we have a

kind of a very related thing in our lighting round, which we will go through a bit quicker, starting with here's why Chad GPT is just the start of the generative AI boom. Yeah. So I thought this article was another great way of looking at the macro trends behind generative AI. Just quickly, the thing that I'll mention as an interesting thought that comes to mind when you read this is, is it the case that we are going to see startups properly valued?

as they come out with new generative AI products. And the reason that this is an interesting question, I think at this stage is the rate of the rate at which new capabilities are coming online. So you come out with a new text generation system and you apply it to, or you apply chat GPT to some kind of like medical diagnostic engine and

And you don't know that six months from now, a better system isn't going to arise. Competitors aren't going to be able to jump all over it. You might have massive growth in user base early on in the short term, and then that growth might stall or just completely crash. And this is something, the canonical example of this is stuff that you saw early on in the Web 2.0 revolution era. I know that famously, like Michael Seibel at Y Combinator has this story about his startup called SocialCam. They reached, I don't know how many millions of users really quickly, but they had no retention.

The competition was too intense. Also, the product was just kind of early to market and so on. So it could spread virally, but no retention. And so the big question is, is that going to happen with this generation of products? And is that now just like a new feature of the way that startups are going to work? They're going to have very short boom-bust cycles as new products emerge that just completely overtake them as the underlying models get better. So I think that's a fascinating question.

Yeah, exactly. I think just quickly, we can mention that this has been happening already with GPT-3 where a bunch of companies have emerged, particularly in writing copyright for marketing. And they all offer the same thing. And it's not super hard to replicate that. And now with ChatGPT, whoever builds that in first and gets a better product

might just go ahead and win. But on that note, we can move on to this next story, which is Microsoft to invest 10 billion in OpenAI, the creator of ChatGPT. So regardless whether immediately there will be startups that kind of are valued highly, I think Microsoft definitely believes that having this technology and having it as a cloud provider

will kind of make it the winner in the short term. Yeah. I think it's also telling, you look at the percentage of equity that Microsoft owns in OpenAI. What does that tell you about how many more investments OpenAI plans to take, right?

That's a massive signal that they think that they're close to something very significant. At the very least, they think that they're close to some sort of economic takeoff, like escape velocity, where they're going to be able to make enough money that they don't need to raise money anymore. You can surely guess that Opening Eye is not a company that plans to sell itself outright and release control of itself to anyone else. So

If I were a betting man, I would say this is the last investment that certainly they plan to take from the outside. And you can draw from that the implications that you want about timelines for AI and stuff. But I think it's quite a telling number, that 49%. Yeah. And it's not too surprising because I would imagine they're not making a ton of money on the API so far. So for now they need that money to scale and to get to compute. But that might not be the case for too much longer.

And also related to this previous story, we have a pretty cool article that talks about how generative AI funding exploded over the past two years. So actually, GPT-3 was from 2020. It's been a while since we had technology, at least in an earlier stage, and already in 2021 and 2022 compared to 2020.

It really exploded to more than $1 billion in all of these startups. And we saw this huge jump and many companies, 85, 78 in the past couple of years. And this year, it's just going to be another massive jump, clearly. But it's clear that already there's a lot of excitement among investors for a little while.

Yeah. And circling back to what Andreessen Horowitz are saying, I think they referred to this as the total market cap that we're looking at for generative AIs. I think as they put it, it's like somewhere between all of software and all human endeavors. So when you think about the level of investment, we're seeing $1 billion at this level, when we're seeing things like chat GPT, that might be on the small side. I mean, that may not reflect even the value, the expectation value of the space right now, where to cram the

your billions into this market though is a much tougher question. And well, if I knew that, that I'd be much more than just a little fiddling angel investor. So I have no idea where that goes, but anyway, yeah, really interesting trend of life.

Yeah, and I would imagine as someone who really thinks about AI safety, one thing to note is right now these systems are not super reliable, right? They can produce outputs you don't want. They can be factually inaccurate. And I think it will take a while, at least, like we're not going to get all the value right away. And these tools are going to be more like things...

humans can use to do their work more efficiently than to automate the entire economy. And whether that'll be the case 10 years from now, that's harder to say. Yeah, actually, sorry, just really quickly to piggyback on that. I think there is something like if you think about AI alignment as something that's really important and as something that's like we talked about earlier, kind of a tax on these organizations doing their work and you worry that maybe they won't spend that tax.

One thing I think that ChatGPT does is it shows us that actually already alignment bottlenecked right now in terms of the value we can get from our AI systems. ChatGPT is not more powerful than GPT-3. It's not really more scaled. It's just better aligned. And so this might give you a little bit of hope that maybe market incentives will naturally push more resources, at least to kind of short-term alignment solutioneering, even if that doesn't tackle the kind of like catastrophic risk side of things. But anyway, just something to flag.

Yeah, and I think I want to mention also that to me, the most interesting aspect of chat GPT, first, of course, the UI was super important, it turned out. But also that the only real differentiation from what OpenAI already had was this human-in-the-loop kind of alignment, RHLF.

And that turned out to kind of work and also maybe to make the output better. So alignment might also result in just improving how all this works, but that'll take some time. And I think for our last story we have after

inking its OpenAI deal. So again, OpenAI, that's all we talk about. Shutterstock rolls out a generative AI toolkit to create images based on text prompts. So this is again showing how

Things are taking a little while. Last year, we already have generative images with DALI 2, but it is kind of moving fast where already you have Shutterstock, a super big player in stock images, and they're rolling out a product that anyone can use to generate images instead of licensing images. And here it says the images are ready for licensing. So you can actually imagine that people who create the stock images can also use AI to create additional images

stock images that people can license. And it's actually quicker to just do a search to find an image that fits your needs rather than yourself creating it. So I think this is a very interesting story showing how already this is going to have a massive impact on very large companies, very large markets.

Yeah, it's hard to, it's like always hard to imagine where this can go. Like our imaginations just keep failing us. But yeah, you look at a company like Shutterstock, which I think was doing well. Actually, I'd never looked at really their business model or like how they're doing, but certainly been around for a while. All of a sudden, like adapt or die, like it's like overnight with one model. And I don't think that's going to slow down, but it's such an interesting case study. Exactly. Yeah.

Well, obviously a ton to discuss with Chad GPT and the text generative AI, but let's move on to research and advancements for things that are a little further out and not quite business ready yet.

Starting with a music LM. And this is Google created an AI that can generate music from text descriptions, but will not release it. So this is more of a research paper. They actually released a paper and a demo of pre-generated music.

And this is showing how the space is also moving kind of quickly, although slower than generative images and text, where you've seen a few papers over the years that have really shown progress for user generation, refusion, jukebox from OpenAI.

But this space is harder. It's harder to get data. Generally, generating music that sounds good is hard because there's a lot of ways where humans can detect minor flaws. Whereas with text, it's kind of easier to skip over some of the weirdness that we have already seen. And this is a pretty big jump, I think, for anyone who's tried out these previous models. The quality of the audio is

is pretty impressive. It's kind of hard to pick up on it being artificial in many cases, partially because there's no vocals, no lyrics involved. But yeah, the other thing is it can generate things that are very long, like minutes long, and you have coherent melodies and compositions, which is another very hard thing to do with music just because it's so much data.

But yeah, this is a big jump. And this is a case where, again, we are seeing maybe it's like 2020, where we see a kind of a GPT-3-esque jump. And within a few years, we're going to see a lot of these models and we're going to have friends and composers starting to worry about these models and how they will impact their jobs. Yeah. I mean, at a certain point, doesn't everybody become a prompt engineer? I'm really, I mean...

I'm curious about the shoot because the last time I tuned in, sorry for the pun, totally unintentional. The last time I tuned into music generation was I think with like Newsnet, which was like this opening. What was that? Was that 2019, maybe 18? It was a long time ago. So back then, already pretty impressive stuff was happening, but I think the generated...

audio was like 15 seconds or something. Is the big, as far as you understand, is the biggest innovation here, the context window, like the length of the generated content, or is it something else or many things? I think it's a mix of things where the complexity of a prompts is higher. You can really tell it's a pretty impressive mix of stuff and it will, you can do death metal with something like rap.

And it would just mix those two genres pretty impressively. They have a really interesting example where you can prompt it with descriptions of paintings like a scream famously, and it can generate, I tried it out and thematically, it sounds pretty good. It kind of fits those paintings and there's just a lot of different things it can do similar to GPT-3 in a way, right? Where it turned out to generalize very well.

And yeah, I think part of it is this very long generation and part of it is just the sound quality and the generality of what it can do. And I feel like I almost don't need to ask this question because it's 2023, but is it based on a language model?

That's actually an interesting question where looking at the paper, we've seen GPD3, all these models, even image generation kind of, it has evolved where it's just one big transformer kind of model. And so it's one big model, you throw a bunch of data to it and it just works. Here, it's a more complicated setup where there are three different models that are trained separately.

And they have different functions where one of them kind of breaks down the different components of music into smaller chunks. One of them handles the decision of what to include, and one of them does the final kind of conversion back to sound. So it's a slightly more complex system of things with each of them being built off of recent research. So it includes

Something from a previous paper from 2022 and another paper from 2021 and another one from last year. And it just combined these things to make this work, which is, I think, really interesting. And I think might be true of what we'll need for longer text generation or video generation, where it will not be as simple as an end-to-end model because transformers, ultimately, it's hard to scale them in terms of the input space and the outputs.

output space for technical reasons. So this is kind of showing that maybe you will need to think a bit more carefully and kind of compose different modules to make it able to generate this more complex kind of chunkier data.

Yeah, isn't this always how it goes, right? There's like this, the crux of the argument between the scaling maximalists like Rich Sutton and then the like Yann LeCarré style people who will say, oh no, you need like a custom brain like architecture with different modules and stuff. And it's,

Really interesting. It's going to be interesting to see if, especially if music turns out to be an early proof point, perhaps for a Yann Nacar style view where, hey, this is just one of those areas that really shows the weakness of a single kind of dense NAD or even a sparse model, but one that has no custom adaptations. But who knows? I think the scaling people would be like, well, just throw Transform at it and see what scaling does in a year or so.

Yeah. And I guess it's worth noting that DALI 2 from OpenAI was also a little bit more complicated. They didn't just have one big model. They had Clip in there and they had kind of a two-stage thing. And then ultimately Google, I think, showed that you can really simplify it down to one big transformer. So it may end up being something like Transformer Excel. We already have models that

are meant to handle longer text, or it may wind up being harder and more complicated like this. And it's just a matter of time until we get more papers, but it's going to be stored just because the data is harder to access. Right. The computer's bigger. Yeah. I almost wonder if we end up with language models that like are pre-trained on text and then applied to music generation as the ultimate solution. But anyway, that's its own side quest.

Yeah. And enough about music. Let's go back to Chad GPT again, because that's really the big story. So the big story which just happened, I think yesterday was the new AI classifier for indicating AI in text. Yeah, this is kind of cool, right? Because we've all heard that we're seeing the news stories where people, especially educators are just freaking out. I think there was a poll that was done showing the vast, like more than 80% or something of like college students

admit to having used chat GPT. Touching that in a bit. Yeah. Yeah. Sorry. I won't tread on that yet, but so educators are, they're pissing their pants right now. Understandably so, but there's this question going around as to whether the right approach is to say, Hey, you know what? We're going to ban this from classrooms and we're going to build really good detectors and patch people for cheating. Or do we just say, Hey, you know what? This is like a calculator. We're going to accept it. It's going to be part of the toolkit. Everybody needs to learn how to use.

And in fact, through that lens, maybe educators have a responsibility to introduce chat GPP and kind of encourage its use. So open AI, I think are quite correct on this. Sam A has said many times that, you know what, you're going to be forced to, it's going to be adapt or die. You can't pretend these things aren't out there. We have a duty to educate people about how to use them and not to keep them awake.

But that doesn't mean we have to do it overnight. In the interim, we can benefit from classifiers like this one that OpenAI just put out that can at least help. It's not decisive. One of the key things that they say in the paper they publish is that they have a decent false positive rate.

So decently often, about 9% of the time, this thing will say, oh, this piece of text was AI generated when in fact it was not. And you don't want to be doing that, calling people out for plagiarism and stuff and like making them fail out of school because your AI classifier screwed up. It's also worth noting that only I think 26% of the time or so, it'll actually give you a true positive. The rest of the time in between that 9% and that 26%, it's kind of neutral. It's not going to give you an output because it's not constant enough.

So this is a highly imperfect tool, something that you should use, they flag that you should only use in concert with other strategies to demonstrate plagiarism. So I think the battle is almost already lost when it comes to generation versus discrimination. I don't think we're going to be seeing an era of really good

text classifiers that can tell with confidence that an AI generated text is actually AI generated. I think that's a lost cause. I think generation will probably overtake it, but we'll see. Maybe I'm wrong and educators will have a little bit more time to breathe, but we'll just have to find out. Yeah, yeah. And curiously, I found this kind of fun that also very recently Stanford released another research result called DetectGPT.

which uses a totally different method. So this thing from OpenAI is very interesting because it kind of perturbs your output just a little bit away. You can detect this does something totally different, but also claims to have that. And all of this, I think, connects very strongly to something that we've seen happen years ago already, where

People were very concerned about deepfakes, video deepfakes, voice deepfakes. And there was a lot of discussion around watermarking and seeing if we can, for photography, kind of watermark and create kind of a source of origin kind of thing. And kind of the nice thing is, I think,

There's an argument to be made, like we discussed before, that as API providers, it's much harder to run this on your home computer. You cannot form ChatGPT as just a normal person with your laptop.

So I think it is pretty feasible that these companies, just even due to pressure from many different places, will introduce these techniques that will get better and better. But it's going to be a battle probably back and forth, and it's going to be a very much active area of research this year.

Yeah, I think this is another area where who knows, maybe it's not a lost cause, but you can imagine people taking chunks of their AI generated text, running it through a text summarization engine, like a different generative AI and make it like screwing up any kind of watermarking or things like that. I don't know. It's a tough engineering challenge. But as with all these things, I keep finding myself saying, I guess we'll just have to see. I guess we'll just have to see.

Yeah. And then let's move on to a lighting round where we'll have a couple more stories that are a little more diverse. The first one is Versadal Robodog runs through the sandy beach at three miles per second.

So we got to get back to the physical world a little bit. And so Keist, this company, announced that their research team developed this Quadra Petal Robot, which is a dog-like robot that you've seen quite a bit of in the past decade. And there's a fun video where there's a person running on this beach and a robot dog can keep up.

on this sandy terrain, which is really pretty tricky even for humans to run in. And so it's pretty cool that they have this little robot doing it. Yeah. And it's cool to see that I think these kinds of robots, the quadrupedal dog-like robots are really starting to be very robust and usable potentially in many areas. Three meters a second is no joke. That's pretty fast. And I don't

I don't know. I was sometimes wonder if some of the people at Boston Dynamics literally look at like episodes of Black Mirror and kind of go, oh, that's a good product idea. Is that what's going on in Silicon Valley right now, man?

Yeah, I don't know. We're still having a hard time selling it to people for various things. So it's not there yet. Robotics is always going to be kind of slow, but it's nice to see on the research front, we are seeing more and more robotics like this. But we have to get back to chat GPT, obviously, because that's all we're talking about. We've been away from it for too long already. Yeah, already. So let's touch on graph GPT.

Yeah, this is kind of cool. It's in retrospect, I think this is like kind of an obvious application that of course, chat GPT could do this, but it still surprised me. So there's this repo, this GitHub repo that I saw called graph GPT. So you go to it, it's actually deployed as a live app too. So you can check that out. Basically paste in like a Wikipedia description of the plot of some movie.

And it will generate a graph that shows you how all the characters, all the entities in that movie are related to each other. So basically reads text and then extracts. It'll basically convert that text into a JSON formatted thing with a graph with nodes and edges and all that stuff. And I just thought this was super interesting because you think about all the analytics that this makes possible like overnight.

And it's a lot. You can think about malicious use too. If you're doing network analysis, trying to figure out who knows who and all that stuff, take a bunch of people's Facebook posts, a bunch of people's LinkedIn posts, pump them through this in a scalable way. And you can do some interesting things, not to give people ideas here, but I think that's yet another layer of thing that would have been science fiction like 20 minutes ago and boom, we've got it. Yeah. And it's pretty cool because in this more structured format of knowledge graphs, it's easier to do things like fact-checking

in the output. And also there are, there's quite a bit of research on combining knowledge graphs, which are just structured facts that are things like President Barack Obama was born in this particular year and combining it with machine learning models where the knowledge graph is part of the input. And that might be a useful thing for AI safety to make sure that facts are preserved and things like that. Yeah.

Yeah, this is a cool new result on the opposite way of creating knowledge graphs. And they do, I will say also, they actually show the prompt that they use for this. So if you go to the GitHub repo and you're curious about what kind of prompt would reliably lead to this output, check it out. I thought it was kind of an interesting dimension of it too.

Yeah. And then we got a couple of stories on another area that has been developing for about a decade. And there's a lot of startups in that space and they are starting to emerge. So we're going to go to medical AI, starting with AI has designed bacteria killing proteins from scratch and they work. So we have a biotechnology startup in this case, Olmadani at Fluorofluid.

And they used AI to design millions of new proteins and then pick a small sample of them to test whether they worked. And in fact, they're starting to test it in real life and shown that they work. And I don't know too much about this, but I assume that bacteria killing might be a good thing. So yeah.

Yeah, now Jeremy steps in, does his Doomer thing. I will say this is from the standpoint of designer pathogens. Damn, that makes my skin crawl. There was actually, I think, already a Nature paper published that did a review on like

How you could just, if you've got a loss function, roughly speaking, make these proteins healthier for humans. You can just put a minus sign in front of it and now it becomes like a killing machine. So this is a dual use all over the place, but very cool. Yeah, that was a big story last year. And I think that is an area of concern where you can use this to design very effective viruses, bioweapons. But let's stay positive and this is good. For sure. This is phenomenal. And proteomics is fantastic.

a notoriously difficult area for people to kind of wrap their minds around and fold their proteins around. And it's really exciting to see it, especially think about diseases like Alzheimer's disease, where, you know, a lot of it is thought to revolve around protein folding issues, at least as I understand it. So some of these things, who knows, maybe we're finally getting close to breaking open a bottleneck in a lot of research areas with this sort of stuff.

Yeah, and that goes to our final lightning round story, which is again in this medical AI space, machine learning identifies drugs that could potentially help smokers quit

So this research team at Penn State College and University of Minnesota found that you can test medications like destometrofen that could be repurposed to help quit smoking cigarettes. And you can test it with a machine learning method to analyze data sets for patterns and trends.

And some of them are already being tested in clinical trials. So this is another case where a positive use case where we haven't had a miracle drug to help smokers quit. They easily, a lot of them have bad side effects. And obviously this will have a massive impact for many people in terms of helping their health. So this seems pretty exciting.

And do they talk about things other than smoking too? Because I assume the same kind of brain addiction pathways might be in play for other drugs, things like that. Yeah, that's interesting. I would imagine that it has to be a little bit application specific, but probably the same techniques could help with other sorts of behaviors. So as with a lot of medical stuff, probably you can create something similar for another application.

Right. So cool. Another silver lining to all this stuff, amazing MedTech advances. And I would expect the next year to be hopefully as wild for MedTech as it has been for everything else. Be nice to share the wealth a little bit. Yeah. And I think VAT space has been, it's exciting where a lot of things are hitting like FDA approval and hitting trials. So there it'll be, it's a hard thing to do, but it's already starting to get there.

And we touched on a lot of things that are a little more worrying already with ChatGPT. And now we're going to move on to policy and societal impacts where we're going to do that some more. Starting with our first story, 4chan users embrace AI voice clone tool to generate celebrity content.

hate speech. So an AI startup called 11 Labs has released a demo, a free tool where you can basically give it a small sample of speech, I think a couple minutes and something you could do. And then you can go to text to speech where you can make this voice say whatever you want.

So it's a pretty impressive model as we talked a bit about how this has been around for a while as well. But like with MusicLM, this speech sounds almost real. It's hard to tell that it's not that person, which is a pretty important distinction.

And yeah, 4chan, famous kind of forum where there's a lot of hate speech of all kinds. It kind of went a little bit viral there where people talked about it and show how they use it. And then we got things like Emma Watson reading a section of Mein Kampf and Ben Shapiro making racist remarks. And yeah.

various things like that. So obviously not a good thing. And then this kind of went kind of viral. You saw a lot of this sort of thing and the lab, the company has already kind of clamped down and has said that they will now have additional restrictions in terms of who can do it so that this sort of thing doesn't happen.

Yikes. You think about all the policies that people are having to invent on the fly for this stuff. Now, I guess this is what OpenAI was getting at. I don't know what it was back in 2018 with GPT-2. They said, okay, we're going to start releasing this in bigger and bigger increments. This idea of staged release. People are going to have to start doing this sort of thinking on a deadline to figure out like, how do you even learn what the malicious uses are of these systems without releasing? Or like,

How do you get those uses? Anyway, predicting the future is really hard and a malicious actor obviously has a lot of incentives to figure it out once you give them the model. Yeah, yeah. I think it's kind of an interesting case where

for GPE2, but also for video creation, there was a lot of concern around deepfakes. Also, for instance, creating photography of fake things. And this was discussed a lot in 2018 and 2019. And we didn't really see that much happen in the deepfakes. There's a couple of examples, but they're kind of easy to tell. They were not real. So that kind of thing didn't

really turn out to be problematic. And then now suddenly that exact thing that people were worried about years ago already, and we're talking about a lot,

happen, where you have a public tool that's really easy to use and a bunch of fake speech was created. And this article also has an example where you have President Joe Biden saying some crazy stuff. And obviously, that's a case where it's not just something that's obviously weird, like Emma Watson reading Mein Kampf, but this could really spread misinformation even more

than we've already had with like kind of easy to refute things. So this is worrying, but hopefully companies like Eleven Labs can make it harder to do this sort of thing.

Yeah. And in the process, if they can pull that off, one of the virtues of what they're doing is it does help to make policymakers aware that this technology exists. It's a good thing, I think, that technology like this is at first way too expensive. Like the average person can't build their own chat GPT, but they will fairly soon, like in five years or whatever, thanks to Moore's law or whatever version of Moore's law ends up applying.

probably processing power and the packaging of software and hardware and all that stuff, it's going to allow people to do this. And so we're getting these early warning shots. That's valuable. It's helpful. It's good. And then people just have to actually pick up on the hint and do something. But

Yeah, exactly. And I think it's interesting where this is another case where similar to ChatGPT, where this was kind of talked about a little bit by a few people and then suddenly you have this public tool that makes it easy for a lot of people to use it. And immediately you have a problematic use cases being displayed. So we had, you know, especially you had a deepfake porn on Reddit and that was a big issue.

And the last year, there were a lot of concerns among voice actors where you saw some usage of this, these sort of techniques. There's a couple of companies in the space to create voice acting.

And that's a huge area of concern where voice actors and actors have a right to their voice, right? You can, it's not okay for someone to just clone Emma Watson's voice and do anything with it. Emma Watson doesn't want that. So that's another big issue that this company hasn't seemingly thought about of you shouldn't be able to clone anyone's voice. And that's a really tricky thing to, to think about and limit.

Yeah, 100%. And I think voice actors, it might be some of the lowest hanging fruit right now in AI, come to think of it. Because having heard, I mean, you were showing me just before we started recording the Joe Biden thing and like, these are mind-blowing, the quality is insane. It's at the point where you almost could just feed a written book and get an audio book type thing. So-

Yeah, exactly. And yeah, it's already starting to happen kind of on a small scale. You have, I believe you had some audio books that were AI narrated being published on Audible. And there was some kind of discussion last year about voice acting for video games.

And yeah, in some sense, it's really nice as with image generation, because now as a small creator, it's hard to afford voice actors and studios. You can now use these techniques to have characters speak and make your game richer. On the other hand, voice actors have a job. It's their livelihood. So there is always kind of a positive aspect and a more worrying aspect.

Yeah, 100%. And malicious applications are not hard to think about for this one either, right? You call your bank in your voice or they call your bank in your voice and engage in dialogue. You can even imagine like a chat GPT type bot coupled to something like this. And you could call a thousand different banks in 20 minutes and just see where, you know, which ones just have a teller who didn't get enough sleep or something and who gives you access to something they shouldn't. Like this is...

There are opportunities for scale bad shit. Yeah. And I recall, I believe last year we talked about something like that where there was a phishing attempt. Yeah. I think against a bank. So this already happened. Cyber security experts are talking about phishing.

And that's just another area where AI is going to make it harder on some people. And as you said before, this is also a concern for educators with ChatGPT. So let's talk more about ChatGPT. And yeah, we have an article called Educator Considerations for ChatGPT from OpenAI.

Yeah, and this is really where they're laying out a lot of the challenges. So this is documentation that OpenAI itself put out because they're reading the headlines too, and they want to help educators understand what the hell to do with chat GPT.

And yeah, they kind of go over all the things we know about ChatGP's capabilities. They talk a lot about the benefits of it, ways you could use it, for example, for streamlined and personalized teaching. I imagine having like kind of a tutor and a lot of people are already using it that way. So that's good.

especially if you have a teacher in the mix, because then they can kind of correct things, right? If chat GPT teaches somebody the wrong law of physics, okay, maybe the teacher can step in and be like, actually, you know, F equals MA or whatever the thing is. There was this section on academic dishonesty and plagiarism detection that I thought was interesting

both because of the philosophy behind it and because my guess is going to be most academic institutions, they're going to be referring to this text. This is in some sense like the definitive text from the people who built the system on how to live with the system in a healthy way. One of the things they flagged was like, make sure that you have

policies that don't retroactively punish people for using this? Because we don't know, like nobody has a sense of what the right thing to do is right now. Is it kosher? Is it not kosher? What's the policy? So that was kind of one aspect. The stuff we talked about as well earlier, like this thing cannot, don't rely on classifiers to tell you with confidence that a piece of text is AI generated. It's often the case that you get a false positive. So it's one part of a complete breakfast type thing.

And then they go into things like truthfulness, right? We know about this issue with chat GPT. It's known to generate false outputs sometimes, decently often. And so you want to think about that, like how much trust are you going to put in the system? What kind of teacher oversight is there on the learning process? That sort of thing. Anyway, it's quite interesting. I would say over-reliance, by the way, is a whole section on its own worth reading too. And then job opportunities and outlooks, which was the, in a sense, I found it to be the funniest part. They're kind of saying, look,

Humility is the best policy, folks. And like, I totally get that. I think it makes perfect sense. But it's just one of those things where for a long time, the job of a teacher was in part to be a job market Sherpa. And all of a sudden, they're just not in a position to do that anymore. There is no certainty in terms of what jobs are going to be human doable in five, 10 years.

And so kind of seeing that crystallized in writing was pretty interesting. And then finally, they just have a section on disclosing the use of chat GPT. So open AI here, really trying to make it easy for people to declare that they have used the system. I'm sure some self-serving reasons, because that's good marketing for them. But as well, I definitely buy their argument and I'm sure their heart's in the right place as they look to kind of make it easier for well-meaning people to be open about their use of the tool.

Yeah, I think this has been already a very active conversation. There's a lot of dimensions to it.

Ultimately, it's not so different from previous changes where, for instance, students got access to the internet, they got access to Wikipedia. It was easy to plagiarize quite a bit from resources. Now it's easy to plagiarize in a way that's harder to detect. And it's easier to just make the AI do the work for you if there's no essay that exists for that. You could make an argument also that you could always pay someone to write you

an essay as well. So now this is bringing those kinds of issues to a forefront. As it gets easier to do that, then again, there's that it's good to adjust. And now educators can focus on the harder things that this cannot do, that chat GPT still cannot do. And there are kind of assignments you can imagine where it

It is not doable by GPT and you can use chat GPT as a tool, but not as a full solution. So yeah, it's very nice to see OpenAI publishing this document so quickly in response to the concerns. And yeah, I think this year we'll probably see more and more kind of discussion around this. Some solutions that ultimately probably benefit everyone.

Very true. Yeah. And the solution side, I think is you're right. Very important to flag. There's a big economic incentive here for people to find ways to help at least help teachers cope, at least help teachers navigate it. And hopefully they do, because otherwise you can't really do this cold jerky. I think that's going to be a little hard on the whole system.

Yep. And yeah, let's go to our lighting round. Once again, we have a very related story from Stanford, from Stanford Daily titled Scores of Stanford Students Used ChatGPT on Final Exams, Survey Suggests. So it said that a majority of students used ChatGPT. I think it was, like you said earlier, 80%.

And a majority also reported there was kind of a mix of things that people used it for. So a lot of it was generation, but some of it was also brainstorming and outlining. Only 5% reported having submitted written material directly from chat GPT with little to no edits.

So yeah, this is a pretty interesting article showing that I don't know, Jet was not out for that long. Final exams were in mid-December. But it's Medford, man. Yeah, well, yeah, people are a little bit savvy about AI here, I guess. But yeah, it already shows that it's clear why. And actually, I'm a course assistant for a class right now.

And we always have, we have code assignments and we always have cases of, did you find this code on GitHub? Did you copy it from another student? And we can detect if you have very similar code to someone else. And this time we saw that and I actually did go to ChatGPT and I was like, is this the output of ChatGPT? Is it why these several students have the same code?

So it's kind of funny where even I as a courses assistant at Stanford or am already impacted by this. It's really interesting actually then to ask, I hadn't thought of asking you, I would have done this offline. So have you seen, let's say an increase in the quality of submissions that you've been getting on average? Have you seen any indications at the macro level that this is happening or? I wouldn't say so. I think most students

I think still would be hesitant to do that in general. I think there is a pretty strong honor code here and very dedicated students. So there's other ways to cheat. Honestly, there are usually solutions on GitHub. You could just Google. So yeah, I don't think it's impacting most students, at least for this sort of thing. I would imagine for written assignments, I also tried it to be like, well, what is the solution to this mathematical problem?

And it answered it correctly. So that was an interesting case where it provided the conceptual approach, but you still had to implement it yourself. So yeah, I think it's likely going to make students' lives easier when they're blocked and don't know where to go. But I don't think it will likely lead to a lot of difference in what students submit.

Okay, that's really interesting. And at least for now, it's making enough mistakes that people need to know their stuff enough to be able to assess whether it's wrong or right. So they're learning something, at least for the moment. Yeah, and for our CS classes, it's not just a case of a chat GPT is a big deal. We already had GitHub Copilot.

that could potentially generate code from comments, entire functions. So yeah, it's a case where it's a little ambiguous how to do this, but I think we'll figure it out more or less. Yeah. We always do. Yeah, yeah.

And let's switch to another area of ethical concern that I think is maybe not discussed quite enough personally. And so this article says, when may a robot kill? New DoD policy tries to clarify. So there was 2012 doctrine on lethal autonomous weapons by the Department of Defense.

That was a little bit ambiguous. And now they have changed it to have a more clear statement about how to build and deploy these lethal autonomous systems safely and ethically and with oversight. So the biggest difference is now there's an explicit need to approve and review any autonomous weapon system. And

Yeah, it's good to see that this area that I think really does need clear guidelines and even needs regulations and international agreements. This is moving pretty slowly. We've seen some efforts from the UN last year. And I do hope that ultimately,

As we see new weapons, new autonomous weapons, we've seen some partially autonomous weapons, not anything kind of fully automated. I think this will also be an area of concern for sure. Yeah, absolutely. And as I recall, like the article revolves as well around DoD directive 3000.09, which is for, so for anybody who's spent any time like in the kind of DoD military space and the intersection of that with AI, this is like the go-to directive that dates back from 2012.

It lays a lot of these ground rules. But, Andre, as you said, like it is kind of ambiguous. It doesn't clearly at least say, hey, there's a review process required. And if you do that review process, you actually can use lethal autonomous weapons. And so for a long time, a lot of people, it seems, including people at the DOD itself, thought that there was a blanket case.

ban on lethal autonomous weapons. So you can think of this as a win some, lose some, depending on your perspective on the stuff. On the one hand, okay, now we have clarity. There are rules. On the other, now we have clarity that you can, if you follow those rules, deploy these systems. And so

Again, like lethal autonomous weapons probably are, it's a foregone conclusion they're going to be used anyway. I think that train left the station a while ago. Maybe on that basis, if you throw in the towel on that, maybe this is just a win. We have clarity now that these rules are going to be enforced. So that's good. Yeah, it's a bit of progress. And yeah, it's similar to something like deepfakes a few years ago, I think, where people are talking about it, at least in a policy space. It's not in the public consciousness quite yet.

But it's only a matter of time until we have some big story, some big event that really blows us up. And then people realize we are close to having real robotic systems, not terminators, little tanks with targeting systems that are going to be deployed in wars. And it's going to be a big deal, obviously.

And you think about what alignment failure means there, right? Like an accident in that context, boy, that could get ugly. Yeah. And so there's a lot of discussion. There's human-involved systems where it's like semi-autonomous. You need to approve things. And I think definitely having less ambiguity is very important. So this is good news.

And moving on to a slightly lower stakes scenario, also robots. We have a new story. Robot cars are causing 911 false alarms in San Francisco. Local to us. Or local to me, at least. You're right. So yeah, we've seen actually Cruise and Waymo

deploy robot taxis with no human driver as a product, as a thing like Lyft or Uber you can order finally after a decade of development. But there's been a lot of issues where this actually points out that there's been some calls about unresponsive

drivers. And there's been especially problems with these cars getting in the way of firefighters, where not only did police and firefighters have to go to these false calls, but there were other incidents where Robotoxy ran over a fireplace hose that was used at a fireplace scene.

And these are the sorts of things that you may not expect, may not think about, but these are the long tail scenarios where it's kind of, we're going to have these problems. And now that we're deploying them for real independently, it's going to, it's going to be a rough patch for probably a while.

Yeah. And even once you get to the point where you can patch all these things and get the long tail for today, like what happens to the long tail for tomorrow? Right. We constantly have an evolving landscape of safety concerns on the road. Bird scooters became a thing overnight, like a couple of years ago in San Francisco and elsewhere. And self-driving car, like how would they respond to that? What if there's a new, I don't know, a new Segway scooter or something that's

These kinds of things, new Boston Dynamics dogs. How about that? This sort of thing is going to be an ongoing issue. And I think this is why some skeptics also are saying like, hey, well, we won't have full self-driving cars until we have our full-on human level intelligence from AI systems. Regardless of what you think of that, you can certainly appreciate the argument that institutional shifts and the environment changing around these systems as technologies evolve, it makes it a really tough problem.

Yeah, and it's a really tough problem for regulators too. We've seen this last decade with Uber actually, where they were kind of aggressive, they kind of skirted against regulations and it was kind of a big fight. Now it's another kind of high stakes scenario where there's a ton of commercial interest in making this thing real. And it's obviously an enormous kind of market opportunity

So a company is going to try to push to deploy these systems, even if it might be dangerous to some extent. Regulators are going to have to balance that with impact on the public where it's not constant. There's not too many cases, but there's maybe enough cases to be concerned in this case. Yeah. And how do you balance it against human accidents? What's the relative impact, importance and value? Yeah. Yeah.

Well, we've had a lot of discussion of concerning stuff throughout this, which we often do. So let's go to our last section, art and fun stuff, where we kind of take it a little bit more easy. Actually, the first story here is not a new story. It's just something I want to share, which is I actually used Chad GPT partially to write a little short story this last weekend where I

We have this ongoing project with another person called Stories by AI, where we just publish kind of a short story, like a thousand words or something as part of a newsletter every week. And the whole idea is to see how we can use AI tools, generative AI tools to use it as a co-writer, as a collaborator, as a tool for writers for these sorts of short things. And

ChatGPT recently hit the scene and we've been experimenting with that. And in some ways it's been better than anything we've used before. And I tried it out this time, went ahead and wrote something that I just had as an idea. And it was interesting where I wrote half the story and I tried to get ChatGPT to write the second half. And it was kind of underwhelming for me in a way where it...

kept on theme in terms of the content, but it really didn't capture the voice I wanted or the tone I wanted. It was a little bit sappy, a little bit too kind of positive. It ended with a little bit of a kind of very happy note that didn't feel right.

So it was interesting where I could keep some of the sentences and little bits of dialogue, but ultimately I had to rewrite most of it, which was helpful in a way because it told me what I didn't want. It actually helped me have the idea of what, well, I don't like this. How about I do this? But it's interesting. I think these experiments with writing short stories of AI have been really interesting in terms of there's

A lot of different ways you interact with them and a lot of different ways that they augment your capability, but they can't replace. Ultimately, they cannot produce. You have an idea in your head of what you want this to be. And the AI will not do that because it doesn't know what's in your head. So it's going to make a version and then you're going to work with that to do really what you want.

And do you find, like, how do you find the prompting game there? If you have an idea in your head, do you try to convey it through the prompt and then it just gets ignored or? I think it depends really. So we had another story last week that was pretty different by another person and I talked to them and it was an interesting case of where there wasn't as kind of defined an idea of what this should be.

So it was more of a back and forth of, well, I want to write this sort of story, these sorts of themes or genre, and let's brainstorm some ideas of where we could go around this general idea. Right. So you have more of this brainstorming session and you could then kind of agree with Chajipati, well, this is interesting.

kind of the idea, this is a rough outline, let's say, well, go ahead and start writing it, right? And then it's more of a tweak scenario. And this was a case more where I already knew what I wanted in a way, I already knew where it was going and I already wrote half of it. So basically we told it, well, complete the short story. I just said literally the prompt was write the second half of this and I provided the half that I did write. And it turned out that use case might not be as good

as when you don't really know what you want to write, you have kind of a general direction and then ChatGPT can act as a collaborator in a pretty cool way. Okay, so human refinement, but the generation side, the kind of the ideation piece is maybe more where the value is right now. Yeah, I think so. And it's interesting to see where, I think this is a case where you really, it's not, you would think that it's as simple as saying, well, write the next paragraph. But in fact,

To really enable these different use cases, you need a different kind of UI where there's another tool called PseudoWrite that I would say is probably still better than ChatGPT because you tell it like write to next paragraph and it provides a few options instead of just one generation. So you can think of, oh, I like this better than that, or I can get some ideas out of this. It also...

has an option where you can tell it complete right to next paragraph with this sort of kind of direction, which I guess you can also do with chat GPT of like right to next paragraph with this rough idea in mind. And that's another kind of technique. So I think, yeah, there's different ways to use it that are quite interesting. And I think

It is going to be probably a case where this year, because of ChatsGPT, a lot of authors and writers will start experimenting. Yeah. Yeah, I don't doubt it. Yeah. So this is kind of a fun little thing. We also generate images of mid-journey to illustrate things. We also have narration by AI. So it's been a really fun project. And you can check it out, storiesby.ai. It's a cute little URL.

And let's finish up with another fun thing. And this is something that I think a lot of you, a lot of us know about, which is one of Boston Dynamics' famous demos. So they released a new video where it's Robot Atlas does some crazy impressive stuff. In this case, it was in the sort of construction scenario of scaffolding and it was tasked with delivering some tools to someone who was

up on the scaffolding and it grabbed the tools, it walked over some planks and it threw the tools over to a human and did like a weird somersault back to the ground. So it's always really fun to see this from Boston Dynamics. And also they had a behind the scenes video that was really cool because it went into what was involved in developing something like that.

Yeah, and I actually haven't seen the behind-the-scenes video. I'd be really curious. But the thing that I always wonder about anytime I see demos like this is how...

How custom-y, demo-y is this? Are we talking how scripted is the demo? Because obviously we have a long tradition, especially in AI, of people going up on stage and giving these highly staged demos that look super impressive. I remember Vicarious did a bunch of these before it more or less came out that there was nothing, no there there really, but that looked super impressive because they were these very carefully controlled demos. Do we have any indication of that with the Boston Dynamics thing?

Yeah, I think so. I think the behind the scenes video actually does quite a good job of getting into it. In general, these demos are like partially staged, of course. So there's no like handwritten motions for doing whatever you want to do. The robot does need to figure out kind of on the fly how to walk and so on. But then what it needs to do, it's not autonomous, right? It's not deciding, okay, I should deliver these tools because the human told me to.

That's more kind of pre-written where it should go to the tools and it should walk to this area and you should throw it. And it's really kind of focusing on showing, well, it can walk in a very stable way and do these like backflips. Or in this case, they focused on how

It can manipulate things in its environment. So we haven't seen these robots kind of actually carry things or manipulate things. So in this case, it takes this plank of wood and places it as a thing to walk on to cross some gap.

And it takes the toolbox and then throws it. So it kind of is meant to demonstrate that capability that they've been working on. And that involves, that's kind of actually pretty hard. So...

pretty much largely staged as far as what's going on, but demonstrating a capability that's really still fascinating and incredible that we're there. I remember visiting the MIT Museum in 2007 or something, and people were like, walking is an unsolved problem and is likely to remain unsolved for another 50 years or something. And just like blasting past all those benchmarks, just incredible stuff.

Yeah. And that's in an area where it's been decades in the work and then Boston Dam has kind of nailed it. And interestingly, I think the walking part is not that much based on machine learning. You don't need, yeah. Whereas here they're starting to do more perception, more manipulation, which is kind of requires more machine learning techniques. So interesting under the hood. Yeah.

Yeah. So as usual, we'll include the links to all these stories and this video in last week's AI in our newsletter and also in the podcast description. So if any of this was interesting, you can read the full stories there. But with that, thank you so much for listening to this week's episode. We're back and we're excited to keep going. So we'll try to keep posting weekly. And if you are a new listener, you just got this link, then go ahead and subscribe.

Okay. Okay. Good. Yeah. So I think we went a little over time.