#166 - new AI song generator, Microsoft's GPT4 efforts, AlphaFold3, xLSTM, OpenAI Model Spec

I finished my PhD at Stanford last year and I now work at a generative AI startup, and this week we do not have our usual co-host, Jeremie. He is on vacation, but we do have a wonderful guest co-host.

I am a diehard listener to this podcast, and so it's such a treat to be able to come on here, in addition to spending as much time as I can every week listening to this podcast, I also host the Super Data Science podcast, which is, as far as we can tell, the world's most listened to data science podcast. And, that's more of an interview show than a new show if people are interested in that kind of stuff. I'm also co-founder and chief data scientist at a machine learning company called Nebula.

So, yeah, so that's me and Andre. Should we should we talk about the beers now? This reminds me of the podcast that I used to love, and it inspired me to get into podcasting. About data science was a show called Partially Derivative. And on that show, partially derivative. They always were drinking beer on air. And so they would open up every episode talking about what beer they're drinking. And so in today's episode. The sound of me opening, my beer, which is but we

have a bit of a twist here. This is kind of like, I don't think I even qualify as a millennial, but it's like a millennial twist on having a beer on air because we're drinking Athletic Brewing Company beers, both of us. And these are basically nonalcoholic. They have half a percent of alcohol, per beer.

So thank you for the feedback. We'll try to incorporate it as we go forward and we'll far. And I guess, John, I, I try.

as usual with the tools and apps section. And the first story is about AI generated music yet again, the kind of one of the highlight trends of a year, it seems. And this time it is coming from 11 labs, 11 labs. If you have been following it for a while, you probably know is a leading source of, AI generated voice. You can give it text and get really good, high quality voice outputs.

And now they have previewed, they have given us a. A sneak peek of their work on AI generated music via some posts on Twitter or ECS. So, not quite out yet. Not clear when it'll be out, but they did post a few examples and the examples I listen to them really clean, really convincing results where, you know, we now have two big players. We recently discovered a new entry with UEO.

There's also now 11 labs, seemingly also entering into a space or soon to be entering into the space of music generation with yet another model that is like when I listen to it, it's hard to tell it's AI generated. So yeah, we are very much in the realm of generating full songs that are indistinguishable almost from AI music or human music.

In case you missed it, episodes that happened once a month, and they recap the most interesting parts of conversation from the preceding, month. And the production team is like, let's make your life easier so that at least one episode of the month, you don't have to be involved. But in the end, it's one of those interesting things where you're like, I could do it, and it absolutely sounded like me.

But because to make the transitions between those conversations really good anyway, I've got a scripted and then I might as well just recorded it anyway. It takes like an extra minute. So yeah, one of these interesting things you guys were talking last week about how in generative AI, the chat application is kind of well-worn through, but there's lots of other areas.

It was GitHub, I think specifically last week. You're talking about that with where there's new ways that you can be using generative AI, but it's interesting how it can get you so close. But just be a little bit off of really, at least today, saving you time in reality.

And I guess that's just going to get better and better and better as we have more and more tools and we become better at, syncing them all together and linking them all together and kind of having a web of AI systems that act as agents, act as generators, and really do make our lives easier and save us a lot of time.

AI now. And similarly, I would imagine with music it won't replace popular artists, but for TV shows, for movies, for things where you might want to quickly generate a song, especially for indie or, you know, less, financially backed projects like podcasts, like podcasts. Yeah, exactly, exactly. So anyway, exciting news from other labs. And you can go to the link in the description, as always, to actually listen to the song. It's pretty impressive.

Next story. This one actually happened last week, but we do want to make sure to touch on it since we did last week, and it is about the mysterious GPT two chat bot I modeled appear suddenly and confused. Exports. So this was on the LMS, this chat bot arena, which we mentioned quite a lot. It's where people can, basically vote on which of our best chat bots and give a score. And it appeared and got. Pretty good scores. And everyone was wondering, is this from OpenAI?

Is this some kind of a new model they're working on? There was no details at all as to, what it's from, but it did have pretty impressive reasoning abilities and, did pretty well, although also had some weaknesses compared to GPT four. And all of this is happening, as we are waiting, OpenAI has announced that, next Monday. So next week, just a couple of days from when we are recording, they have some sort of announcement comping, on Monday morning.

So people have been speculating that it's search and OpenAI. I said, no, it's not search and it's not GPT five. Is this gpt2 thing related? Maybe it's some sort of extra small model that's really good and nobody knows. But, it's interesting that this happened and, yeah, it's a fun little story to be aware of.

This is it's kind of I don't know if this is a shameless plug in. You can pull this out, Andre, if you want to when you're editing. But, the the team that created this, we had the head of that team, Professor Joey Gonzalez, on my show Super Data Science and episode number 707, and he talks about Elon Says a lot, which was new at that time. And yeah, it's a really cool way to be comparing models in a way that is, crowdsourced and leveraging real human feedback.

It's definitely the biggest, the biggest player for doing that. I think the most well thought through, website for comparing, chat bots in that manner.

So losses and things like it are pretty much the best way to know model quality, or at least one of the best, better ways. And again, later on in the show, we'll have another interesting bit of news with open source as to this topic.

Basically, you'll be able to query perplexity and answer, complex questions. Interesting to see perplexity starting to do these sorts of partnerships and expanding beyond just your own platform, providing AI services to other platforms. Have you, used perplexity at all? John?

That's pretty much the story. And I guess for people who love with discord in a way of generating images, another tool you can use now.

The interesting thing I think about focusing on UI, like perplexity does, is that it gives you the capacity over time to be improving. If I went to Back to Perplexity today, I might say, wow, you guys have really done a lot more here and it is so much more valuable to me. Whereas, you know, with a kind of discord experience, that's not going to change.

You know, we've mentioned it before. It seems like a no brainer. And I'm sure they'll get it going. Pretty seriously. And there are some announcements, you know, presumably coming in the next few months regarding this huge story.

on the iPhone. And this similarly seems to me this kind of situation where surely Apple would prefer not to be dependent upon Microsoft or ChatGPT for this technology, but it's expensive, and there's very few talent on the planet that can get you to the frontier. Like just a handful of labs around the world are on generative AI, and so it makes a lot of sense, pragmatically, I'm sure.

However, with things like scrapping the Apple Car, which I was super excited about, I'd I'd love to even just see what that was going to look like. But things like that project being scrapped so that resources can be diverted to generative AI. You. I wouldn't be surprised if we're not too far away a year or two, maybe from Apple being able to do these kinds of generative AI projects on their own.

And I would bet that they will have a huge focus on security and, yeah, reliability, probably more so than being right off the cutting edge of capability. I imagine Apple will be a little bit more conservative in trying to make sure that they have. Yeah, the most secure, the most reliable outputs.

So as always, be mindful that there's a whole ecosystem over in China of people competing and trying to create ChatGPT models and on to applications

and business. And the first story is open AI and Stack Overflow partner to bring more technical knowledge into ChatGPT. Just last week, we talked about how OpenAI has been making a lot of deals with news publications like the Financial Times, and now there's been this announcement of OpenAI and Stack Overflow making a deal. Stack overflow is. Has been kind of a leading website for conversations regarding programing and just technical details.

A lot of people go over to ask questions of like, how do I do this, I have this bug, etc., etc. and as a CSS student from like a decade ago, this used to be a site you went to all the time, right? They have a lot of data, and I'm pretty sure ChatGPT and all these, our models have already been trained on that data. Kind of. Anyway, so there's now a deal for OpenAI to have access to the StackOverflow API and be able to use feedback from, that community to enhance

their models. And StackOverflow will be attributed within that ChatGPT. Interestingly, similar to how ChatGPT will be attributing to things like Financial Times. So yeah, another play by them, another potentially way to be able to overcome the limitation of having, knowledge cut off in these models. Right. That's one of the key limits of GPT is it will always be trained up to a point in history of like April 2024

or whatever. So I think these sorts of plays speak to that where we have integrations of new sources and, you know, these kind of forums regarding different question and answer topics. These tools can still be useful for things that come out after retraining data.

But then the internet allowed us, of course, to be able to do Google searches and find information originally more so in the documentation, kind of the official documentation for a given programing language. But then StackOverflow became invaluable. It became the default, where I would get some stacktrace and grab the error at the end of that trace, paste it into Google. It takes me to the Stack Overflow, and then I can just copy whatever the solution is without even having to think about it.

And to underline the importance of Stack Overflow. It's interesting how. Gemini from Google Cloud, also inked a deal with Stack Overflow in February. And you can see why if you want to be getting great coding suggestions, Stack Overflow is got to be the best place on the internet by far. That is 90% of the time or more where that stack trace, copy paste into Google takes me. And now it is really interesting.

That is an example of something we were talking earlier in the episode about tasks like my voice being simulated for, an episode of mine once a month where that didn't quite meet the mark. Generative AI for me. Now, Claude three opus is where I go every single time when I have an error in my code, because it is so brilliant.

In the same way that if you're familiar with doing that in GPT four or I haven't used, GitHub Copilot myself, and I'm sure it's the same kind of situation where you get great feedback right away. And the thing that is different about doing that with a web search is that it is contextualized exactly to your code, so it can just rewrite your code for you. You don't have to change any variable names or anything.

I think it is worth highlighting that an additional story that came out after this ad was that Stack Overflow has actually been banning users who have, in a sense, rebelled over this partnership. Some users have started trying to delete their answers to prevent them from being used to train GPT three. And, yeah, Stack Overflow has tried to prevent that by banning users.

So as we've seen often with these sorts of deals, to some extent with Reddit, not everyone is happy to have their work, their, you know, contributions be used to train AI models.

Model collapse, where if we start losing high quality real sources of data like Stack Overflow, if people stop contributing to Stack Overflow because they can just do everything in generative AI tools, does that mean that we're going to have a big, negative impact on the quality of generative AI models in the in the future, such as they completely collapse? Andre, I don't know what you think about this, but I actually I'm I'm relaxed about this.

It seems to me like the AI systems, their ability to evaluate their own outputs. For example, I just said that my favorite generative AI tool of this time is called three opus, and that uses reinforcement learning from AI feedback as opposed to from human feedback. And that seems to be part of how it's been able to outperform GPT four and Gemini. And so, I don't know, it seems to me like we're we're figuring

these things out. And we the the folks at the Frontier Labs, seem to have really clever ways of distilling down to the highest quality training data one way or another, whether it's AI generated, human generated. I don't know if model collapse I'm relaxed about, but what do you think, Andre.

And the bigger problem, I guess, is the end of data that is being approached, and where there is kind of an open question of, you know, how much, how much further can we push your systems and onto the next story? New Microsoft AI model may challenge GPT four and Google Gemini. So Microsoft is building their own large scale language model that apparently is codenamed my. One. And this one will have 500 billion parameters approximately.

So this is much bigger than what Microsoft has been releasing with Fi, kind of smaller large language models. This is more on par with things like GPT before. And I guess it it is interesting because we know Microsoft has a partnership with ChatGPT. They have been integrating OpenAI into Azure. ChatGPT. Rather into Azure. They brought on, Mustafa Suleyman from infection, where they trained a large language model, and it seems like he will be heading this development.

He also was, of course, at DeepMind. So it seems like a way to kind of de-risk things from the Microsoft side, where we don't have to rely on OpenAI to have their own large language model. They will have it and of course, will be able to compete with Google and Meta as one of the players that is able to develop these kinds of systems.

to. If you want to be at the frontier, you've got to have access to a ridiculous amount of cutting edge compute and inflection. Just might not have had that. And so they figured, you know what? Let's go over to Microsoft and do it there if they've got tons of compute. And the interesting thing you already highlighted

this. I'm, I doing the thing that that, reviewer explicitly asked us not to do, but I maybe with a little bit of a twist or a little bit of additional commentary, which is that it makes perfect sense to me that Microsoft would be trying to do this stuff on their own without requiring OpenAI, because they haven't acquired OpenAI. So they've taken a huge stake. But this isn't their IP. OpenAI can make deals with Apple like we just talked about earlier in this episode.

And so Microsoft would, you know, you could imagine a scenario where if somehow Microsoft had created GPT 3.5 and released a chat GPT tool, something of that quality before OpenAI had done it, that would have been an even better outcome for them than needing to invest in an an outside party and get the models from them. So, yeah, you could yeah, you can totally understand why they would be doing this kind of thing.

pack there. But wave now has a billion of dollars, so they'll keep working on it.

They have some autonomous taxi rides and. Las Vegas and deliveries in Santa Monica. Those will be halted. So, yeah, seemed like a pretty significant setback. They, did get an investment of 1 billion as well from Hyundai. So they will be continuing to be active and working on this. And perhaps this is kind of a strategic thing. Hard to say. But regardless, another player in the space trying for the same thing and I guess delaying their plan to try and launch until 2026.

We've got to make sure that this is really buttoned up, or let's see what happens with Waymo and the regulatory situation before, you know, we take more risk in this kind of situation. I don't know exactly what's going on, but you could imagine something like that.

Some of them are not necessarily working on frontier AI models. So moonshot, for instance, focuses on digital assistants to summarize text for students and office workers. Minimax is targeting Viv, targeting the gaming market with anime themed characters point 1 or 0 one that I, of course, has been developing AI models. So yeah, they are big players in China. We may not hear about them a lot in US media, but, definitely interesting to kind of keep it in mind that this is happening in China.

And now on to the projects and open source section. And the first one is, Paper Prometheus. Do an open source language model, specialize in evaluating other language models. This is, a collaboration between six different institutions. So Kaist, AI and MIT, various universities, algae research. And as per the title of the paper, this is related to evaluating language models. So they claim that they figure out a way to closely mirror human and GPT four judgments.

Basically, you can have, high correlations. And unlike Confucius, one which were released last year, they focused in that previous model on basically giving a score one out of five here. This one is greatly expanded. So it can do pairwise ranking, which is what this, arena does. It lets you compare two different answers and pick which one you like better. And as with Prometheus one, there is a lot of infrastructure that goes on this way.

Just highlight how you kind of have user defined evaluation criteria. That is how they had scores 1 to 5. And in this paper we go a lot into how they expanded it. So now support also pairwise rankings. Have they also have an expanded dataset with a lot of examples of pairwise ranking and also verbal feedback, of outputs. And as per a title, once again they released some models, code and data, all open source.

So pretty promising effort I think, to help with that evaluation challenge where you can go beyond benchmarks, use these sorts of labs to evaluate, you know, which I think is a pretty, standard approach that people are doing.

So the same, faculty member at Berkeley, Professor Julia Gonzalez, who was, one of the faculty members behind that, chat bot arena also in March of last year, in March of 2023, was one of the faculty on the release of vacuna, the, the model that was a fine tuned llama model and it out competed llama and what they did for their evaluation, they said at that time it was like only partly scientific, where they were using GPT four to compare side by side,

let's say, how does our vicuna do against llama? How does it do against GPT 3.5? And so you have these head to head tests similar to what humans are evaluating in chat bot arena. You allowed GT4 to do it, for the vicuna evaluations and for us at our company, Nebula.

That is what we do most of the time because having people either your own customers who you might annoy if you have them, kind of doing evaluations, if you bug them with that kind of thing all the time, or even if you try to internally say like, you know, let's ask our business development team or our product team to be evaluating this model that we built and compare it against GPT four or compare two different versions of our model.

But this gets it is so unscalable to ask humans to evaluate your models, because there are so many situations as a data science team on a daily basis, even as we're checkpointing a model, as we're training it, we want to know, are we overtraining and are we starting to overfit? Potentially. Or have we reached a point where this model that we've trained is, is is mature? And so these kinds of computational pairwise evaluations, to be able to compare two models head to head is so important and is.

Is the future of this space, if it isn't already today. And so I really appreciate, folks like the researchers behind Prometheus to developing the specialized model for evaluating other language models. And you can bet that our company will be using this right away.

The reason they say this is an economical and efficient model is that, those many, many parameters of a neural net, only 21 billion parameters are activated for each token. And so when you look at a graph, we have in this paper of optimal activated parameters versus performance. Now they claim kind of a top spot in terms of four of

a number of parameters. You get really good performance on the ML you benchmark, which is kind of one of the leading ones that people still use to claim very really good. So it's roughly on par with Lama 370 B while using way fewer parameters. And there's various details on how they do that. We use an multi-head latent attention mixture of experts. They train on 8 billion tokens, with various tweaks, and they actually released like a 30 page paper summarizing a lot of

details. So to me, a pretty exciting paper, just because a lot of this is still a dark art, and this has a lot of details on how it works, and I do release the model checkpoints on GitHub.

And, another interesting thing here is that the chart is you're saying, you know, the ML, you asterisk already there in some world where ML you is a perfect, indicator of how great a model is. It is interesting to me that for deep V2, they have done exactly the same chart. Which you can see in the paper. It's figure one in the paper that's linked to this episode. They, they created the exact same chart as Mistral did for the mixture of eight by 20 to be release.

And the point that Mistral was making was, hey, look at us like we occupy this top left corner of this chart where you have activated parameters on one axis and performance on ML, you on the other axis. And these two mixture models eight by seven B and eight by 22 B occupy the top left corner and then deep seeded deep Secret comes along and says, F you guys, we're going to occupy an even further corner in the

top left. And so yeah, it's yeah, it's interesting to see how people position themselves relative to each other. It does make a lot of sense. It would be nice if we lived in a world where I don't know, you was 100% reliable. Because then this kind of release would mean that you have a no brainer to go for if you want to have the best performance with an open source model simultaneously without the cost of running A37 DB in production.

the lightning round. First up, Open Voice V2 evolving multilingual voice cloning with enhanced style control and cross lingual capabilities. I think we're gonna probably want to go quick here, so not too much to add to that. This is coming from MIT and Michelle that I and Shanghai University and, yeah, it's, really nice model to be able to do voice cloning in multiple languages.

And as always with these new stories, it's the best. And they do release it under a super open license, Apache 2.0 for both research and commercial use. So, yeah, lots of progress still being made on lamb specifically for code. And next we got Huggingface launches a robot open source robotics code library. So Huggingface, made some news earlier this year, just a couple months ago by hiring from a Tesla.

And the first fruit of that effort is now coming out with this open source lab robot package that they say is basically meant to be sort of like where Transformers package Transformers is a package that is used very, very widely for doing work in natural language processing with, with transformer architecture. And they want to position this as that type

of package. But for a biotics, where it's a toolkit that is a comprehensive platform that has, you know, support for getting data sets, for simulators, for different types of robots, for, training models, for having pre-trained models, just lots of stuff. And, yeah, it's someone who has done work in robotics during my PhD. I think it's pretty exciting. We really don't have this sort of unified, kind of primary package. What people build on there is kind of a

mishmash of different ones. People have sort of tried to do this over years and hugging face, getting into that ring and releasing this, I think, has a pretty strong chance of supercharging AI research for a lot of organizations.

So Peter Abeles company where they released RFM one Robotics Foundation Model one, and the specific idea behind that, similar to the robot, is to be bringing LMS into the physical world. So autonomous vehicles being another place that you could imagine LMS and AI being able to make a huge difference in the physical world with robotics, that there's eventually going to be an infinite amount of, of real world capability infused by AI. It's more expensive, as you guys say, on the show.

Time and again, hardware is hard. So it happens a bit more slowly than software. But when you're working with actual real world physical things as opposed to just bits, there's, you know, a lot of potential there. And I'm super excited.

And they say what this is meant to compliment u MoU is, multiple choice, question. And here they have like a golden answer. That is the ideal answer. And they have an evaluator rigor core that gives you a rating from five. So kind of similar to Prometheus in a way, but for specifically multimodal AI rather than a just chat bots and yeah, exciting. Again, lots of work being done on evaluation because of just something that is just needed.

And now on to research and advancements. And we start of course, I think we have a big news of the past week and that is AlphaFold three. As is so often the case, the major research news in the media, and I think kind of fairly also just, a mock the community is coming from DeepMind and this one is AlphaFold three, the, next generation of er, AlphaFold models. So in the past we focused on, protein synthesis,

and analysis. Now they expand it to other capabilities within the kind of the biomolecular, interaction analysis. So this is going way beyond my knowledge base. But just reading out of the abstract, it is a model that is capable of structure, prediction on complexes, including proteins. And you click acids, small molecules, ions and modified residues. They say that it significantly improves accuracy over many previous specialized tools.

So far, greater accuracy on protein ligand interactions and protein nucleic acid extractions. Again, I don't really know what that means, but probably pretty cool. And the sort of exciting bit from an AI researcher perspective is they do this by kind of simplifying their approach. So in the paper it passed in AlphaFold two. It was a pretty complicated architecture with some very kind of custom built parts for the task.

It wasn't the sort of general purpose model, as we see with foundation models, typically it's just kind of a transformer have nothing specifically built for the task, and that's why they're able to scale and do lots of things. So in the paper, we're going to how we essentially have simplified the architecture, trained it on many tasks. And despite removing these things that are, you know, meant to make, the model do better at a specific task, we're still able to train it up to be about as good.

One more detail. They do say they will not be open sourcing this unlike we did with AlphaFold. This is done in collaboration with, the kind of more commercial, arm of alphabet Isomorphic Labs that is in partnership with DeepMind. So it will be interesting to see if they try to commercialize it. Whatever. But yeah, exciting news once again from the mind doing research on AI for science.

exit. And that leads from GPT two to GPT three to GPT four. More and more emergent capabilities. And so they're kind of taking that engineering focused approach of let's take this attention mechanism and scale it up really, really large.

See what happens. Google DeepMind might have been surprised by how effective that scaling approach was, that engineering approach, because what they were doing with things like AlphaGo, through to AlphaZero, they were systematically creating they're trying to make their way towards artificial general intelligence.

So an algorithm that has all of the learning capabilities of, human by starting narrow like a single game, like go and then saying, okay, let's see what we can do to have an algorithm that can not only beat the world's best go players, but also the world's best chess players and the world's best shogi players. And then a year later, let's add in Atari video games alongside that. And so their approach was to, you know, chip away generally.

Generally, generally adding more and more capabilities until eventually you reach this AGI algorithm that can do anything. And so that seems to be the same kind of approach that we're doing here, where with AlphaFold, AlphaFold two, the kind of the press headline at that time was that they solved protein folding, at least for some types of proteins.

And so similarly here, DeepMind has taken that same kind of approach that we saw from them, going from a single game to multiple games to games in completely different kinds of modalities. You know, adding in board games and video games here, they started with proteins, absolutely crushed that and then said, okay, well, let's see what we can do with other kinds of sequential molecules. So proteins are made up of strings of amino acids.

And here they've said well let's see what we can do with strings of nucleotides which form DNA or RNA. And then that gives you so much more training data, a lot more diversity of data. And just as we saw with the shift from, say, AlphaGo to AlphaZero.

AlphaGo model. And similarly, here with AlphaFold three, we're seeing the same kind of thing where AlphaFold three is outperforming AlphaFold two, which was already the state of the art at protein folding by taking into account other kinds of information, by being able to model other kinds of data like DNA and RNA.

And then in this case, unlike, you know, there's no obvious, at least immediately, like maybe as you get closer to AGI, then there's is some kind of way that you can think about how being good at chess would make you better at playing pong on Atari or something. But interestingly here with this innovation, there is built in these interactions because in a biological system, DNA is interacting with RNA, which is interacting with protein.

And so building right there you have yeah, this generalization providing quite a bit more utility. And yeah it's really exciting to see.

Transformers. LSTMs are a form of recurrent neural nets, and they were sort of the main thing being used in natural language, basically until transformers for a couple of years and they achieved a lot of cool things. But architecturally, they were sort of abandoned because they are kind of tricky to train and tricky to scale up compared to transformers. They have some of these inherent limitations

architecturally. And so this paper, zdm is presenting a variation and extension of LSTMs that aims to fix that. In a sense, it introduces some pretty fancy new, modules. So just to get into a little detail, they have memory cells of LSTMs that have memory mixing and exponential gating. They also have ML, LSTMs that, add the ability to do more parallel training. A lot of kind of nitty gritty details where this, extension method would take probably half an hour to get through.

So unfortunately, we won't be getting into a lot of details. But the end result of these architectural improvements is that at least in the experiments on this paper, they show pretty comparable characteristics to transformers in training. So we're training at 2.7 billion parameters on a bunch of tokens.

They show kind of V golden outcome, which is you can have a scaling law that shows that as you go from not super many parameters to a ton of parameters, you have a smooth decrease in perplexity on next token prediction. And in fact, when I compare the characteristics of ACM to things like Mamba and Lama and other architectures of neural nets that do language modeling, in this paper they have a best, performance at every sort of scale of model.

Very much related to what we've been seeing with Mamba as far as exploring alternatives to Transformers that incorporate some of your ideas from recurrent neural nets, to Transformers. Attention. All of that, and potentially a big deal. I mean, the results are pretty impressive. And this is coming from one of the pioneers of natural language processing, neural nets. So generate a lot of discussion and will be very interesting to see if people continue to build on it.

And so that was Yann LeCun, Yoshua Bengio and others at AT&T Bell Labs at that time, who created this one net five architecture that was commercialized for the US Postal Service in order to be able to recognize handwritten digits, for zip codes. So postal codes in the US and so be able to add some automation to routing of mail. And that was kind of the peak of neural networks for a long time. We then went into an AI winter, where to get funding for neural network research.

You pretty much had to be in Canada. No other federal government was, supporting that kind of research. Until the AlexNet moment out of again, Elliot's. Let's give her Jeff Hinton, working on that AlexNet paper and Alex Shevchuk, of course, the third author on that paper and the namesake of AlexNet, that in 2012 brought kind of deep learning into this, into acceptance that not only is it super powerful, once you scale up compute wise, but it's also highly generalizable.

And so anyway, so a little bit of a history lesson there to say that this kind of architecture, the Lstm dates back to before that last AI winter, that last deep learning winter. And so this kind of approach generally from Hulk writer and Schmidhuber has been around for decades. And it's interesting that they have been able to with this paper make changes because. If you tried to use an Lstm alone before we had transformer architectures, in that Dasani at all. Paper attention is all you need.

Before that, LSTMs were kind of the best way, the best known way of trying to link many words together in a sentence or in a paragraph, and be able to make linkages between those to be able to have some kind of attention. The transformer completely blew the Lstm out of the water such that I'm not aware of commercial applications of it today, but hell creator Schmidhuber, other researchers have continued to push along on this Lstm idea, and it's interesting to see that we're now getting to a

point, at least according to this paper. And again, as you already stated earlier this episode, as usual, it is the best on the benchmarks that they that they just selected. And so it's hard to know exactly where it stands. But, yeah, I mean, even if it's performing comparably to Transformers or to, to mama, these other kinds of state space models that are coming out now, it is exciting because

it would be great. And it wouldn't be surprising if in the coming months, in the coming years, we're able to get architectures beyond the transformer that allow us to have that same level of attention capabilities that the transformer has without the really heavy compute.

But still, when you deploying these things at scale, when you're trying to optimize the cost of compute, something like this may not necessarily be the best, but at the same time, contrary transformers, they this kind of architecture has constant memory complexity as opposed to for dramatic increase in usage with respect to the size of input.

And that is similar to things like Mamba that just have these better scaling characteristics in terms of how much compute you need as you try to input, you know, a crazy amount of text, so will remain to be seen whether this is like a new mamba and everyone starts working on this and go crazy, but definitely exciting in terms of still seeing potential successors to Transformers, at least in some context.

And so yeah, trying to have our cake and eat it to not use too much memory, be highly parallelizable, get the same kind of attention that we can get with Transformers. But, you know, there's so many labs trying to do this kind of thing, and it is so important in order to make AI be something that is broadly accessible and not super expensive, and not requiring more and more and more nuclear power plants to be set up, to run AI centers. So, yeah, really cool.

But this is exciting, in part because it's coming from a pedigree of original inventors of science until Lightning Round. And the first paper is story diffusion consistent self-attention for long range image and video generation. They present a semantic motion predictor that in particular helps with consistent video generation. And they also have examples of kind of multiple image

generator shows. They kind of present an idea of a comic book where you have a single character that is consistent across, the frames of a comic book. And this is one of the challenges with, image generation is, you know, if you say generate X, you will have a different outcome every time. And there has been many papers on how to make consistent outputs. For, you know, creating comic books, creating illustrations, whatever.

And so this is presenting a new way to do that and has really nice looking smooth videos. And, these are comic book panels. So again, you can go and check out the link for those example outputs. I wish I had the time to actually edit videos to add all these, example, images of what's happening here, but pretty cool bit of research here. Next up, chain of art empowers Transformers to solve inherently serial

problems. This is more of a theory paper that honestly would take me a while to understand, but the gist is that they show that, kind of theoretically the idea of chain of art, which is, as you mentioned many times, prompting your language model to list out its reasoning steps before providing an answer to a given query. They show that this general approach for a certain category of problems that inherently require

serial computation. So multiple steps of thinking enables transformers to solve those types of tasks that if it were not doing, chain of art, if it were doing, just an immediate output of whatever answer you want, it just would not be able to get it right versus for these types of problems, it could get it right. So for people who enjoy reading dense theory papers, I'm sure this is an exciting one.

And this is based on decades of experiments that him and Amos Turkey did. And so these two thinking systems, there's, system one, which is fast thinking, which is what we're using all the time. I'm using it right now. Words are just coming out of my mouth. And I'm kind of just hoping that the next word that comes out is useful to

some listeners out there. Whereas the slow thinking system to this is where you're sitting with pencil and paper, and you're making sure that, you know, you've got all of your assumptions, right as you work through a mathematical problem or you're writing some computer code. And in in order to make this analogy very easy to understand, as you say, learn how to drive a car. You start off with that slow system to thinking where you're like, okay, foot on the brake, turning the, the key

in the ignition. And now I'm, you know, going into reverse and pressing decelerate. You're having to to do all this very deliberate thinking. Whereas once you have learned how to drive a car, you're listening to the last week in a podcast as you're driving and you're having a conversation with a friend and drinking out of your coffee all at the same time, and so it moves from this, your own brain becomes adept at shifting things from the slow processing to the fast processing chain of thought.

Prompting like this, to me, is like making the most of that fast thinking. You're prompting the generative AI model, which is just predicting next token like you are when you're thinking, when you're speaking in real time or thinking in your mind, you know, in language, in real time. And so chain of thought prompting is like making the most of the fast kind of thinking that humans do.

What I'm excited about, in combination with this kind of chain of thought, prompting and generative AI is approaches like alpha geometry as well as Q star, which are specifically designed to be much more like the deliberative thought that you have when you're doing math, when you're learning how to drive a car, when you're computer programing.

And so that kind of, that kind of slow thinking in combination with fast thinking, including chain of thought, fast thinking could be a really interesting combination towards realizing AGI.

ONS where you have some weights, on edges between neural network units, where the units do some function on a combination of its inputs.

And this paper presents a whole new alternative that basically puts that paradigm on its head, where you have functions on the edges and the nodes just sum up the outputs of those inputs, and they go into a whole lot of detail of kind of, conceptually why this has some advantages in terms of interoperability and, the potential expressivity of this sort of a whole new paradigm for how you build your neural nets.

Again, a bit more theoretical in the sense that probably people will not adopt this at scale, I suspect, but interesting if you like, these sorts of very conceptual papers. And onto to a policy and safety section

that with the first story being. US lawmakers unveil bill to make it easier to restrict exports of AI models. We've been talking a lot about export controls for GPUs and hardware to China in particular. That has been in the US, and now a bipartisan group of lawmakers unveiled a bill that would actually make it easier for the Biden administration to impose export controls on AI models, make it hard for software to be shared, essentially.

And if approved, the measure would remove some roadblocks to regulating export of open source AI. That apparently there's already some precedent for that, ability by the white House. And there's some more details in the story. Interestingly, apparently China, many in China have been building on top of Lama, in particular as the starting point to making their language models. So evidently that would be why they're, the U.S., lawmakers are seeking to do this.

So, yeah, very interesting conceptually to think about actually having export controls for open source models.

I mean, I'm not pretending that it doesn't exist, but it's a shame, because it seems like we're so close to having so many of these emerging technologies like AI, nuclear fusion, quantum computing, being so close to realizing them and having them being able to bring abundance to everyone on the planet.

But these kinds of things, which I totally get it, makes perfect sense in the real world that we live in with geopolitical conflict, but it's just a shame that we can't have everyone getting along and everyone collaborating, on open source, on hardware, instead of having to have these kind of competing groups in one region, having to feel like they are needing to catch up. And it's interesting because it seems like mostly like a futile exercise.

Anyway. So, you know, the U.S restricts access to some hardware or to lama architectures to say China and China, then invest a whole bunch of redundant capital in recapitulating the same kinds of advances separately, and eventually seems like they'll catch up. So it's like, I don't know, it seems like a lot of a lot of wasted potential, a lot of wasted capital and, you know, slowing down the, the capacity for, for everyone all across the planet to be, you know, living a higher quality of life.

But yeah, I mean, I'll take my rose tinted goggles off.

So this kind of notion that these are enemies isn't the full picture, even as there is very much a real, situation where the US law is seeking to basically combat the ability of China to make advancements. And, I think with this kind of law, it's would be surprising if it can actually be effective to actually block the ability of open source models to be used in China. But it does speak to making big moves, basically continuing to add to pressure, as we've seen also with hardware export

controls. Next up, OpenAI's model spec outlines some basic rules for a AI. So OpenAI AI has released the first draft of a proposed framework, which I call Model spec, to propose basically how AI tools should respond. It's a bit like you could say like an ethical framework, almost, where it proposes three general principles AI models should assist developer and and user. This should benefit humanity and in the case of OpenAI, they should reflect well on open AI.

There's also specific rules here, such as following chain of command, comply with laws respecting creators rights, protecting privacy and not generating not safe work content. And. Has been positioned at least online. What I've seen as trying to make it very clear what principles are guiding OpenAI's development. So it was maybe a bit opaque.

And what I've seen positioned is if, ChatGPT doesn't want to reply to you with this model spec, they can say, okay, this was a mistake and the model should have behaved differently, or this was what you want to do according to our model specs. And, yeah, it's very interesting to see, given the impact of OpenAI and just how influential they are now, seeing this released and having more clarity on their internal development processes.

context. So this model spec is kind of clarifying exactly how they want their models to behave and why they might, refuse to do certain things, but do other things aren't of lighting around for story. Robot dogs armed with AI targeting rifles undergo U.S. Marine Special ops evaluation. So the robot dogs are literally the sort of things you may have seen already from Boston Dynamics and, these ones developed by Ghost Robotics specifically.

And they are equipped literally, there's a gun mounted on top of it from defense tech company Onyx Industries. And this is kind of small scale testing at this point. There's two of these things that, they are experimenting with.

This organization did clarify that weaponizing the dogs are just one of the many use cases being evaluated, and that it apparently adheres to all Department of Defense policies concerning autonomous weapons, presumably meaning that you do want to have a human in the loop and not have it be fully autonomous. But regardless, I think pretty notable in that we still don't have fully autonomous systems.

We have some examples of drones being semi-autonomous, but we are still not in an age where robots are being deployed in battlefields and doing a lot of a work of kind of a traditional soldier. And if these sorts of systems of dogs with guns, you know, undergo variation development within several years, it is not kind of out of the question that they will be a large factor in battlefields, in warfare. And that is something that maybe we are forgetting in this moment.

You know, there's a lot of concerns about disinformation, misinformation. But, you know, it's it's just a matter of time until we have kind of scary robot soldiers.

And so the your opponent then says, okay, well, let's create all kinds of jamming to prevent the remote control operator from operating this drone effectively. But then that same drone operator says, okay, well, if I lose the ability to control my drone in enemy territory, what should we do? And 2023 was supposedly the first year where in the Ukraine Russia conflict, you know, is the first place anywhere in the world that we're aware of. Drones started to be able to act on their own after

losing. Connection to their controller. So, you know the jamming is effective. It prevents you from being able to see what your drone is doing. But you still, you know, the drone has some expense, it has some munitions on and it's behind enemy lines. And so it's using it. It's using maybe open source.

It would be using it. So we're just saying like all these kinds of tools that we're developing to try to create a better world for the most part Python code machine vision libraries, PyTorch would be being used, to have machine vision systems work behind enemy lines.

Once you've lost control of that drone and it's, you know, hopefully that machine that like open source based machine vision system is accurately detecting the enemy as opposed to just, you know, someone tending to their turn ups or something. And yeah, it's, obviously a, a minefield. And it seems inevitable that we're going to see more of this, which is, yeah, a scary thing.

And it's not inconceivable that some years from now, most of the, yeah, most of the fighting could be being done by autonomous vehicles as opposed to by humans, making decisions and. Yeah, and obviously lots of concerning things there.

They are releasing a deepfake detector to disinformation researchers. So this is a tool being released to a small group of disinformation researchers meant to detect valley free outputs to distinguish between real images and synthetic images generated by their tool. And OpenAI says that this will have 98.8 accuracy, although it will not be able to detect images produced by other generators like Midjourney instability. So good news to be able to see what images are from here or not.

But this whole area is still kind of unsolved. It seems to be.

But people have been, for all of history, able to generate fake text. And so we're kind of prepared to take with a grain of salt things that people say or things that people write, but images or video, up until the last a year or so, we have been able to, you know, to to have the saying seeing is believing. Whereas in the last year that statement is basically no longer true.

And so it's great that, we are able to detect and yeah, it's interesting that it's kind of fundamentally different because of these kinds of signatures that we can detect in it. Yeah. I don't know if you have more kind of technical knowledge on that. Andre.

And we've been covering there's been more and more announcements of things like meta saying that will add watermarks to their generations, partially based on some pressure from the, executive branch of a U.S government. And we some of policies institute there. So definitely some progress being made. But it's still very hard when you release a model that's open source, like stable diffusion that is for bad actors, easy to mess with and make it so you can't detect images.

So it's not really a solvable problem fully, but at least for things like Dall-E free, you can do these sorts of techniques onto our sections synthetic media

and art. And the first story is audibles. Best of AI voiced audiobooks tops 40,000 titles. So last year, Amazon has announced that self-published offers in the US, who make rare books available in the Kindle store, are would be able to use a new tool in beta to release audiobooks on audible with AI generated narration.

And as per this, news article that has now resulted in 40,000 titles being released on audible and that, you know, as you would expect, has led to mixed, responses, with some audible users being unhappy that there is this onslaught of audiobooks while independent and self-published offers who basically will not be able to afford to hire a voice actor to do a professional narration, benefiting from this and being happy that they can use the tool. The article is pretty detailed.

It goes into some conversations with actual voice actors, and it doesn't seem like jobs are being impacted quite yet. But of course you can see the concern that maybe it would impact that industry soon enough. I wonder, John, I know you wrote a book, Deep Learning Illustrated, and, I don't know if it makes sense to do an audiobook.

Probably you do have illustrations and add book, but I could also imagine as you write things, perhaps a new book, wherever you would consider doing this sort of thing or with your podcasting experience, just do your own narration.

that I took there. But even worse than the images is that I tried to minimize equations. And so, you know, there's not a. There still are quite a few equations like this. There's got to be 100 or more equations in the book. Those would be very difficult to kind of visualize if you're just listening to somebody reading out equations. But even worse than the equations is that the book is full of Python code hmhm.

Not related to this idea in general is that, one of my best friends, his name is Zach Weinberg, and he was a guest on Super Data Science in episode number 646. I brought him on that, episode to kind of be like a layperson, giving their experience of how ChatGPT has impacted them, professionally. And he said to me.

You're so lucky, in a way, to have written a book before the generative AI era, because now, anytime anyone writes a book, there's going to be some skepticism as to how much you actually did yourself. And that'll happen more and more and more and more. And so I guess, yeah, I'm I'm lucky to have squeezed one in there right at the end of the of the human writing era.

Rather, these platforms will just include information to save the tour generated on the given platform. And now TikTok will be building on top of a standard to make it clear whether images are generated or not. And speaking of that, a couple, maybe slightly more fond stories related to AI generated images. The first one is that Katy Perry's fan made AI image is so real, at full the world into thinking she was at the Met Gala.

So yet another example. We've covered some of these before of things going big where they are in fact AI generated by people this many people thought were real, and in this case it was very vague images of Katy Perry being dressed in some very fancy dress at the Met Gala. There are some fun outcomes of this particular one, where apparently Katy Perry's mom even texted her and said, wow, you were at the Met Gala. This is a gorgeous dress.

And Katy Perry then responded saying, that's AI generated beware. And so yeah, and this article goes into some details on how a fan of Katy Perry, who runs a fan Instagram developed dress and how it took actually some work. They used a Microsoft tool. They used a second tool to event post-process it, things like that. But yeah. Now basically don't believe what you see is the lesson that we're getting.

But there were a lot of interactions where first, this supposed Musk told some stories, and eventually even they had a video call with a deepfake of Musk convincing, this person to then give up some money in the scam. And we've covered some stories like this in the past. I keep saying this, but if this is something you need to be aware of, scamming is going to be a big thing with AI. And be careful people.

I've told all of my loved ones that if you ever get a phone call from me asking for money or anything, like they'll call me back, you know, or make sure it really is my number because, you know, you and me are in the same boat, Andre, where there's a lot of we have many orders of magnitude more samples of audio available publicly on the internet than needed to simulate our voices effectively. And it'd be very easy to, to scam our voices. And, yeah, it's definitely scary. Things can happen.

And her likeness was used in generative AI tools, where she's speaking perfect Mandarin with the Kremlin in the background, comically. And there are dozens. She found dozens of accounts using her face. Where, in these videos, she's glorifying China and, you know, complaining that Russian men are drunk and lazy while praising Chinese society and technology.

They these women with names like Natasha and Sophia in their fluent Mandarin, are saying things that like, for a Chinese husband, we'd be delighted to cook and wash your clothes and bury your children. And so it's it seems like it's, a propaganda thing, more so than, than a targeted scam, but, yeah, nevertheless, I don't know. There's something about this story that. It's I don't know. It's like the confluence of so many different things. It's like geopolitical issues, romance, generative AI.

Anyway, it makes maybe kind of a fun story to end today with.