#150 - GPT Store, new Nvidia chips, DeepMind’s robotics progress, bad uses of AI

And this week Jeremy is not around. He is off talking to politicians, I guess somewhere. So we have a guest host.

pretty technical. So not for everyone, but if you like going deep on, you know, certain topics and, and really getting into the weeds of stuff, then you might want to check it out. And before we get into discussing the news, I just want to have a quick shout out to a bunch of new reviews on the Apple Podcasts. Uh, I guess people heard our appreciation last time, and so we got like six new ones. Uh, which is great. Yeah, we really love seeing it.

A couple of you mentioned that we recorded while we were sick last time, so it's cool to hear that, uh, you know, that inspires you or you think, you know, that makes us committed to this, which I guess we are. And, uh, yeah. Thank you, everyone for reviews that one of you mentioned, uh, longer segments on arts and entertainment with AI. And actually, we'll have some new stories on that this week. So that would be, I guess, uh, nice for you to hear.

All righty. Kicking off our first section, tools and apps with OpenAI's custom GPT store is now open for business. So this is I guess would be a new story of the week. Is OpenAI has their store for custom chat bots? Uh, this is after the GPT uh builder program, which was announced in November, was kind of launched. And there's now been 3 million bots created by users. So originally this store was supposed to launch actually earlier but has now launched.

So essentially, instead of just chatting with one ChatGPT, you can now chat with all these various GPT guys, right? Customized versions of ChatGPT so that users on the platform can create. And this is now available to users of ChatGPT plus enterprise users and to this new tier, uh, subscription that we'll cover next after this.

There's also a reporting system and kind of seeing a little bit around the Twitterverse. I've seen at least a few people who have made custom GPT that ultimately got taken down and seem pretty unhappy about it, so I'm not entirely sure what those I'll. I'll be honest, I haven't looked too deeply into the guidelines just yet, so I'm not entirely sure if they're being applied consistently, but it's, um, interesting to see the response right now.

So these customized GPT will diverge from kind of main ChatGPT over time, potentially as people use them, which would mean that this will be a true repository of, I guess, you know, thousands or hundreds of thousands of different chat bots trained on different data and different interactions. So, uh, yeah, seems like probably a big deal. Really?

So I'm curious if the GPT store is going to really change the game in that regard.

in that direction very much. And already when you go to the store page, you can kind of browse for, um, different applications for image generation, for writing, productivity, programing, education, etc. etc. etc. so if you're a fan of ChatGPT now, you might want to go and look if there's a customized version for your needs I guess, or create one for yourself. And onto the second branch of OpenAI developments of this week.

The next story is on how OpenAI has released this new way to subscribe to chat. GPT aimed at small teams, and this is, uh, which had GPT team here. That is kind of in between a single user and enterprise. Seemingly. This is, uh, workspace for teams of up to 149 people, and it introduces admin tools for team management and, uh, you know, all the usual access to all the tools and also has a guarantee that this will not be, uh, using your data to train similar to the enterprise tier.

And this is priced at $30 per user per month, uh, for multi building or 25 per user per month. So a bit more than the standard ChatGPT Pro if you are a single user. But yeah, it's interesting that I guess we're expanding their offerings to now cover small businesses seemingly.

And I think that for many of the things that businesses want to do, there just aren't a lot of good solutions out there. And there's still a lot of research problems that need to be solved, it feels like. But it does seem that OpenAI is still targeting this market in a pretty important way.

of storage. It runs on rabbit's own operating system called rabbit OS, which is based on what they call a large action model. This acts as a universal controller for apps that can control music. It can order cards, buy groceries, send messages, and do more through a single user, uh, through a single interface. And again, this is a pretty interesting move. I think that there are a lot of people who are really interested in building more a gigantic products and things based off of large language

models right now. So it's very interesting to see rabbit actually kind of come out there with a device that is looking to serve this sort of need.

We just announced it, and I guess a lot of people on the Twitterverse and elsewhere got hyped about it. The initial 10,000 units sold out already, so we'll see. Yeah, it's just announced. I don't know when they'll even come out, but people seem pretty hyped.

different personas. These include fictional characters and historical figures. You might have seen that meta recently launched the sort of thing as well, and their messenger app, so this seems to be the sort of experience that a lot of companies, especially in the social space, seem to want to build around. Splash launched a free Alexa skill. This enables users to create songs using their voice. They can choose a musical genre, add lyrics, and either rap or sing along.

Perhaps good if you're kind of interested in creating music, but you have no capacity for actually composing stuff. Like me. Volley has introduced a generative, AI powered 20 questions game. This uses AI to interact with users by asking questions, providing hints, and explaining yes or no questions.

section. Google is working on an advanced version of Bard that you will pay for. And that story. It's supposedly going to be called Bard Advanced, and this will be something you pay for through Google One. This will presumably be powered by Gemini Ultra, this, uh, version of Gemini, their flagship model that's akin to ChatGPT that, uh, is yet to be out. So yeah, not too surprising here. I guess. Something we probably all expected, but, uh, will be very interesting to see when this does come out.

If it will measure up to GPD four and other, I guess paid tier chat bots.

So the question I guess for Google is will they be able to deliver something that is enough of an improvement over everything else and is distributed in the right way so that they can recover some of that market share from the competitors? I think that's a really big question for them right now.

Relatively cheap relative to the high end GPUs people use for AI training and things like that. And these will have tensor cores for running generative AI applications. So this is kind of moving away from the business enterprise level of GPUs that cost tens of thousands of dollars each towards more of a consumer bent. And as we've covered, AMD and Intel both have had some, uh, of their own hardware announcements aimed more at runtime, not training at

inference. And so with these announcements, Nvidia is, uh, running into that category as well.

I think as soon as GPT three came out, for instance, we already saw people experimenting with using it to generate dialogs for characters. Um, and also like the ship can be used for tasks like generating images on Photoshop's Firefly generator, like removing backgrounds and video calls. Um, and Nvidia is also developing tools for integrating generative AI into games. So this is something that's going to be pretty huge going forward. And I'm not at all surprised to see Nvidia jumping onto that.

And I think a lot of ways right now, this isn't just happening in games, but in the case of valve, this is aiming to increase transparency around the use of AI and games, protect against the risks of AI generated content, and allow customers to make informed decisions about purchasing AI based games. Um, these rules are coming after a lot of developers complained that valve was rejecting game submissions containing AI generated assets due to copyright concerns.

And so again, this is just a case where having that transparency is going to be pretty important. There are a lot of things people will run into when they're developing AI generated media, where they come into conflict with things like copyright and having some knowledge about what's going on, that there is a generative AI system being used, and probably some of the technical details of that system are going to be pretty important to both understanding and mitigating these concerns.

And in many cases, uh, because it's not like you can necessarily tell. So yeah, interesting to see in the gaming space, which is of course huge and asset generation being a huge kind of application of AI. Steam, for those who don't know, I guess we should probably mention, is a major marketplace for video games. So if you want to buy a game on a PC or Mac or any kind of non, uh, console, you would usually go for steam. Uh, so it's, it's a huge deal for them to allow it.

It's kind of like, I don't know, YouTube allowing AI in their videos, so to speak.

Those are some pretty big names in the investment landscape, so this is like a pretty serious fund raise. The round is valuing the company at $520 million. Post-money. That's a lot of money. Uh, so perplexity was founded in August 2022 by a couple of engineers with backgrounds and AI distributed systems, search engines, databases, basically all the stuff you need to put together to create a search engine like this.

And perplexity offers a chat bot like interface that allows users to ask questions in natural language and respond to the summary containing source citations, which is again a really powerful alternative to something like Google. When ChatGPT came out, people were really excited about the fact that you don't have to go to Google and then look at like the top ten lines to figure out the information you were looking for. You could just have it delivered to you.

And so when you develop a system that is capable of doing a lot of the job of that search for you to deliver you the information and deliver it correctly with. Links and sources and all of that. You've got something really powerful, and it lowers the amount of work a user has to do to find what they're looking for.

They claim to have 10 million active users and they're now valued at over 500 million. So we'll have to see. But, uh, yeah, if you're looking to try out a tool, then perplexity might be something you're interested in. And next story is about self-driving cars and about how Waymo will start testing robotaxis on Phenix highways. Waymo has been testing and actually running commercial services in several cities for a while now. They've been in Phenix for a very long time.

I've been in San Francisco for a while, but this has always been in the city itself, in the streets where I've not been allowed to ride on highways. So this will be kind of expanding that to allow the cars to use that. And, uh, yeah, once they start testing, presumably, you know, and sometime after it will expand to allow that in the commercial commercial offering as well.

I think this was sophomore year or so, and we were driving home from Sky Harbor Airport, and I just saw random Waymo's on the street, and I'm pretty sure that was one of the first times I'd ever seen a self-driving car actually going on the road. So it's pretty interesting, exciting, I think, to see how far they've come in that regard. So I'm definitely very curious to see how things work for them going forward.

They are launching generative AI by AI stock via text image platform that they are going to allow you to create stock photos with. This is building on the previous AI image generation tool, but is designed for individual users rather than enterprise solutions, and this was done in collaboration with Nvidia. Trained with a Picasso model I learned from Getty's library and uh, also AI stock, stock photo library.

So yeah, it's, uh, expanding, I guess the range, uh, of users that can create stock photos beyond just big businesses to small and medium sized businesses.

This allows robots to gather training data in new environment and multitask. The goal is it can simultaneously direct multiple robots to carry out diverse tasks in a range of settings. It's been tested in real world evaluations over several months, and demand is integrated a robot constitution into auto art to ensure that the robots follow specific safety rules, including, you guessed it, Isaac Asimov's Three Laws of Robotics.

This is, uh, pretty fun. There's also another system they developed called the Self Adaptive robust attention for robotics transformers to improve the efficiency of robotic transformer models. And you might have heard of the robotics transformer projects are DeepMind has come out with recently. I think that the development of foundation models really offered some very interesting research directions for grounding the outputs of these models.

So, for example, putting a language model together with something like a robot arm and grounding the language model, suggestions for achieving a task like, I want you to move this block from this part of the room to another part of the room based off of what the robot arm could actually do. So lots of hard problems like this to explore, and it seems like DeepMind is going down that route again.

I guess we're starting to highlight the growing amount of work we have in this direction and the, I guess, growing capabilities of a models for robotics specifically. Right. So these are foundation models as, as part of that other title, uh, they say embodied foundation models, models that are trained to really control robots. And the numbers here are pretty impressive.

So they say that they had authority proposing instructions to over 20 robots across multiple buildings and collecting 77,000 real robot episodes via both teleoperation at autonomous robot policies, and that that allows them to collect a lot of data and therefore train these models, uh, to, you know, continually, I guess, expand the amount of data and control more and more robot robots allowed all over the place.

So it seems like a pretty exciting time for biotics in terms of getting these models that allow you to do low level control. You know, moving a robot to pick stuff up and then move around your environment. Now we also have these higher level control things, like our team that are orchestrating and making decisions on what different robots should do and kind of doing the high

level decision making. So once we get these foundation models trained, and DeepMind seems to be very much pushing on that front, but then you could get to a point of general purpose robotics. And, uh, within the next year or two, that is seeming more and more plausible given the pretty rapid advancements in that space.

And that inference time, that inference actually squares with the length of the context you gave the transformer the computations. It has to do so if you give it a hundred words, then you can think of the the computation that it takes to be 100 squared. Not the best explanation in the world, but it looks something like this. And that's not the most computationally feasible thing to do, especially when you scale to super long context lines.

So one thing researchers are doing right now is they're trying to figure out how to a work on transformers so as to mitigate that issue. But also there is a line of research exploring something called state space models, which offer linear time inference with respect to the context land. Again, that's much less expensive here. And they also had this pretty efficient training process via a hardware aware design. Uh, state space models are pretty complicated math wise.

They're inspired by a lot of things like control theory, but basically there have been a number of deep space state space models introduced. Um, and they are currently especially with the recent state space model called Mamba, apparently challenging the dominance of Transformers. So people are really looking into this research area. And despite the fact that all of the big architectures today are based on transformers, it's kind of another important line of research.

This paper, Emily Mamba, combines the recent model model with something called Mixture of experts, which is a class of techniques that allow drastically increasing the number of parameters in a model without much impact to the operations, the number of operations required for the models, inference and training. Basically, again, making that model a lot more powerful without having to substantially increase the computational

cost of running that model. So this paper basically combines these two techniques and claims that to unlock the potential of state space models for scaling, they should be combined with this mixture of

experts technique. And so they showcase it on the recent model and find that it outperforms both the original number model and transformers, with mixture of experts achieving the same performance as Mamba and much fewer training steps while preserving inference performance gains of mama against the transformer.

are better together? Yeah. That's like an exciting, uh, finding, given where we cover again, again, that it costs, you know, crazy amounts of money to train these models. Millions of dollars. And here Mamba is making it so the model itself is more efficient. Mixture of experts makes it so you can train with half the kind of computation to get the same performance,

seemingly. And that means that we could unlock a lot of efficiency and potentially that would enable a lot of scaling, which would mean that we can make our models even better. So yeah exciting times in the I. You know, architecture technical detail space for a while, for a couple years it was all transformers all the time. And nothing seemed to really, you know, cross that threshold of being good enough to replace Transformers. But you might be getting there.

So pixel delta is basically a delta, a next step on that that introduces some extra tricks into the process of the selection consistency model and and control. Net uh, just combining some existing concepts and that significantly, uh, speeds things up. It produces high quality images and just a few computation steps. So that means that it takes only half a second for generating a thousand by a thousand pixel image. Uh, that's a seven fold speed up.

And it is also meant to train within a single day on high end GPUs. So yeah, it's following up on a lot of progress also in this space of make it being able to generate images quickly. And, uh, yeah, for a lot of these businesses and applications that have generative, uh, texture image capabilities, soon we might be able to see images being generated in under a second, just super quick.

And now you can do that with 3D scenes. And as I've said before on this podcast, I think 3D generation and editing of 3D is going to be a big trend, an area of improvement throughout this year.

And so again, the relationship between chain of thought effectiveness and the length of reasoning steps and prompts isn't super well understood. So in this paper, researchers conducted experiments to explore that relationship I just mentioned. They manipulated the length of reasoning steps and train of thought demonstrations while keeping other factors constant. Again, this sort of length of reasoning stops things.

Uh, has a bit to do with how complex the issue you're trying to get your language model to reason through might be, and they found that lengthening the reasoning steps and prompts without adding new information actually significantly improves our reasoning abilities across multiple datasets. Uh, possibly because there's some more context or something like

this added to that. And they also found that shortening the reasoning steps, even while preserving key information, considerably reduces the model's reasoning abilities.

It is allowed, and we've covered quite a few papers here on this whole prompt engineering genre of research where you're like, okay, how do I alter my prompts and get the model to be, uh, more accurate chain of not being one of the big ones. So this is pretty useful in terms of understanding how to use chain of thought. Well, and uh, we've, you've covered quite a few papers that integration of art in data generation and sort of factual factuality, uh, checking and then various things like that.

So it's it's pretty significant to, I guess, understand a little bit better how to use chain of thought properly. And one paper, which we aren't going to go too far into because we have already covered this topic of mixed raw. But I think it is worth highlighting that, uh, the company Mistral has released the full white paper, so originally released a model. Make draw eight by seven b, which we covered back when it came out to this mixture of experts.

Uh, variant of on a chat bot that was very good and is very popular with people building on top of this open source, uh, release of the model. So we've released, uh, a paper now that you can go look through for a whole bunch of results. Uh, you know, lots of analysis on training and performance and so on that, uh, um, yeah. Provides more details that kind of corroborate what we already know, which is mixture model is quite good.

And using a mixture of experts seems to really improve training and accuracy, uh, in pretty impressive ways.

An interesting name there. It offers a variety of scenarios, including a brothel staffed by underage girls, which is raising concerns about AI powered child pornographic roleplay. The uncensored AI economy includes a lot of sites like chubby AI, and this was initially spurred by OpenAI and later accelerated by Mat as a release of its open source llama tool. Again, technologies that can be used for good, but they're always dual

use. You release something open source that anybody can use and and things like this are just going to happen. Um, and experts are warning again that these activities and papers real world dangers for minors and raise questions about legal and ethical boundaries, as well as tech companies accountability for uncensored AI. Um, top AI is actually, for context here, an uncensored clone of character AI, which Andre mentioned earlier.

It allows again users to engage in roleplay scenarios of AI characters. And as we mentioned, some of these involve child pornography. Not a fan.

can kind of stop you. Whereas with, uh, models where open source models are released, you can basically take them and, uh, do whatever you want. Uh, to some extent with them. So yeah, not surprising, I suppose. But we do live in a world now where these models are in the wild and have been in wild for a little while.

So having a lot of this uncensored AI, so to speak, a lot of people are kind of putting a lot of effort into having models that you can use to do whatever you want, even, let's say, role play with underage. Uh, what what time should I use here? Uh, underage, uh, sex workers. Let's go have sure that, uh, is, uh, just happening. And if you're interested, I guess in the details or at least.

And, uh, real world example of this already happening job is, uh, discussed in a decent amount of detail in this story as, uh, as a prominent example. And apparently it's making a lot of money. It's, uh, generated over 1 million in revenue since launching this chat service, so. Mhm. Yeah. We'll we'll have to see if meta does try to kind of fight uh, uses of its open source models that are this problematic.

So that's again kind of highlighting the research area of there is this back and forth where companies like OpenAI, like anthropic, like meta, are training their models not just to predict the next word, but adding these techniques on top of them in order to ensure that they. Output, things that are reasonable, that meet certain guidelines that they're going to want. And this does mean that users using ChatGPT might have a more frustrating experience.

But it also guards again to use cases like these. And so there is this back and forth between the training techniques used to make these models have safer, more aligned outputs with principles. But then people, on the other hand, who are inevitably going to come and figure out ways to get around that.

But as an organization you could also just not say anything potentially. Right. You could just build that service off of any open source chat bot. And meta may not be able to go after you necessarily. So there's a lot of kind of dimensions to open source. And should you open source and whatever about whatever. We're not getting into it. You're just kind of highlighting something that is now happening because of revelation of open source.

Um, as we all know, a lot of the data from the internet that ChatGPT that image generators like stable diffusion are trained on is covered by copyright. And just recently, the New York Times sued OpenAI and Microsoft, accusing them of unlawful use of its work to create their product.

In fact, as this was going on, a lot of users went into ChatGPT and realized that yes, it was actually quite easy to get ChatGPT to just verbatim spit out quotations and actually pretty extended quotations of multiple paragraphs of New York Times articles just by prompting it in the right ways.

OpenAI pretty quickly covered this. Up I went, and a couple hours after seeing some of these posts to try it myself, and pretty obviously OpenAI had sort of patched this up, but it's definitely pretty concerning. And interesting that you were just able to get these verbatim repetitions.

And opening a has been making the case that it is a case of fair use to use copyrighted material to train a model. Uh, and this is kind of an instance where they submitted this, I guess, uh, argument to the House of Lords. So, uh, we'll yeah, it'll be very interesting and impactful to see where this goes legally, I guess, will be a big fight over the question of whether data is fair use when it is used for training a model. You know, times obviously you're saying that it isn't, but, uh, it'll

it's still an open question. And it is kind of a very important question that will probably be addressed, uh, in sometime this year.

Um, and again, in cases like this, notice might be questions, you might have different opinions, but at least it's kind of interesting to see that they appear to be willing to work with people. As to the details of what that actually looks like, it's it's hard to say very much right now.

to host nations. And you've seen several news stories already of cases where in the actual cases, lawyers, uh, cited kind of fake precedents or fake information because of chat bots. So, uh, this is among the first news stories I've seen where there's an official kind of policy of use chat bots to maybe write your rulings, but do not use them for research.

And I think that these safeguards and limitations and that use, uh, at least just the rules around it are pretty important, especially when you're talking about, well, discourse statements people might make that could be legally binding in some way. And so I think it's actually a pretty big question for the law about for judges, maybe for companies when they issue statements that are legally binding. And maybe some of those statements are AI generated, what that

looks like, what are the parameters? How do you deal with that?

of mistakes. So they found that between 69% to 88% of responses to these legal queries can, uh, result in hallucinations in state of the art limbs. And also they highlighted that real limbs often lack self-awareness about the errors and can reinforce incorrect, uh, information, which is, uh, yeah, something that if you use an alarm, you might have found that sometimes if you point out things wrong, it would just say that, uh, it is in

fact, correct and so on. So very much ties into that previous story in terms of given that we know that it seems that current limbs make a lot of mistakes when you ask legal questions, and I guess would do analysis and research, it might be a good idea to restrict their use for this, uh, case, uh, just as we heard England did.

But again, if you spent any time with the language model, you might ask that these sorts of trick questions where included in the prompt, you get it, you lead it towards the wrong answer, and the model is very likely to take you up on that and just kind of go along with what you said. So the study found that alums are susceptible to counter factual bias, or the tendency to assume that a factual premise in a query is true, even if it's flatly long wrong.

Like if you ask it, why was peanut butter invented in 2020 or something? I can't promise you that Trump is going to work out, but when you make statements like this, when you query it, when you ask things that implicitly assume or just say something that is totally wrong, the model more often than not, just kind of goes along with you.

Uh, unless it's, I don't know, a super well known or something, but in general it should be very careful.

This list was used in a court exhibit and a lawsuit against companies Midjourney, stability, I didn't, Art and runway AI again, all very big names accused of misusing copyrighted works to train their AI systems. Many artists have accused Midjourney specifically of stealing their work without permission, and a spreadsheet listing almost 16,000 more artist names as proposed additions to a MIT journey style list was also shared on social media.

This is again highlighted artist's frustrations with the lack of regulation around AI generated art. Questions have been raised about the fairness of profiting from mass produced images when the AI models that create them are trained on and imitate styles created by real life artists. Again, if you're somebody out there who has put out a lot of paintings in your name, then somebody could prompt an image generator to ask for an image in your style.

But it's the image they want to see for you as an artist that might significantly impact your livelihood. Um, the document that contains all list was publicly accessible until it went viral, but an archive of the spreadsheet does remain available online.

It's very much still. And this kind of spreadsheet that highlights how thousands of artists, uh, have been sourced, so to speak, in terms of the data used to train the models for many of these people, kind of is adding fuel to the fire, it seems.

It's just really, really hard to get all the data. You might want to make these models perform well, and imposing rules like getting consent from every artist whose work is still under copyright is really, really hard. So then the questions. There aren't a lot of good answers here. Or do you just not train the model on them? In that case, you don't get the good model. Or do you see? Permission. That doesn't scale very well.

It's, um. We're definitely not at a good point where we have a great resolution on things like this.

So apparently there's a genre of true crime content for using the AI to create disturbing depictions of victims. Uh, and yeah, now there's an explicit policy that says that you're not allowed to do that. It's. Yeah, pretty insane. But just another example of the sort of stuff we're gonna get from AI when it's in a wild.

disgusting. And again, YouTube is imposing some, some, uh, some actual recourse here. So violation of this updated policy results in a strike that removes the offending content and temporarily restricts to users activities on the platform, and penalties increase for further violations within 90 days, which would potentially lead to the removal of an entire channel.

So this, uh, was released and, uh, yeah, met a lot of criticism, including from George Carlin. Starter.

material. But again, these are these are thorny questions. Is this the sort of thing that people obviously want to consume? These things? There's a market for it, but is it a good thing to have? Do we want this sort of thing? Hard to say.

And these terms include informed consent for the use of AI to create digital voice replicas and, uh, requirements for save storage of digital assets. So very much yeah. Expanding on video that uh, was already in place, uh, that dealt with digital replicas and AI versions of actors.

Um, and so again, this is going to be another back and forth that we're all going to have to pay attention to going forward.

Uh, free early Mickey Mouse cartoons entered the public domain via black and white ones, and a lot of people immediately started training on it and generating data to, in 1928, design to, uh, yeah, mess around and make funny things with AI, Mickey Mouse, AI 1928 Mickey Mouse.

Excel doesn't make these images 100% legal because, as we mentioned earlier, the base model still incorporates copyrighted work in its training data. Uh, but again, this is, uh, more fun, interesting use of these things.

As always, you can find the articles we discussed here today and subscribe to our weekly newsletter with similar ones that last week in that I thank you, Daniel, for guest co-hosting.