The Most Important Events in Open vs. Closed AI

Today on the AI Daily Brief, I am joined by Tiana Baker Taylor, co founder of Venice, to discuss the biggest topics in the battle between open and closed AI over the past month. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. To join the conversation, follow the Discord link in our show. Notes. All right, Tiana, welcome back to the AI Daily Brief. How are you doing?

Nlw I'm doing great. How are you?

Good. So we had this concept to on sort of like a roughly monthly cadence, come together and kind of talk about the most significant events with a loose lens, let's call it around the idea of open versus closed AI. So it's not strictly just like a technology review of open source or anything like that, but you kind of see, and I think that, you know, our listeners will hear across the set of conversations we have today that one of the undercurrents really is a power question when it comes to AI. You know, is it going to be, you know, the, the big tech companies that have power, you know, can open source disrupt that? And then where does the government fit? So there's a lot of news that has happened over the last month or so that fits that and, and so we're going to kind of run through a set of different categories and I think where we wanted to start is actually with a technological update this month we had, you know, it seems like two open source models, or opener models, let's call it catch up, the most we've seen when it comes to the state of the art. And so those two are Llama's 3.1405 B, which is the first version of that model that we've gotten. And Mistral Large 2, which kind of went a little bit under the radar. It happened just after Llama, but has been getting more and more buzz, at least I've seen from developers. Was this what you had expected with this Llama release? You know, there had been a lot of speculation, you know, that, that maybe, you know, Zuckerberg has certainly sort of planted the flag that he didn't want to be, you know, just, just good for open source anymore. But what did you think when you saw, you know, the sort of information start to come out around the performance and capabilities of llama 3.1405 b. Yeah.

Well, I mean, if you look at the metrics, you know, straight out the gate, seven out of 15, I think, are exceeding the major closed models that we've evaluated previously, then you're waiting to see what the user experience is like, I think that we're used to working with smaller models certainly in a generative AI environment and sometimes speed becomes just very tangible user metric. And so a larger model is obviously going to be smaller and there's a lot of education around what these larger models are going to be best designed to be used for. Right. Where a 7B model will give you quite a quick response. Maybe this is going to be more helpful for other types of use cases. But I think that the response is also in admiration of what Meta continues to do. So they're setting the bar high to deliver state of the art models. They're making them completely open source. And I think if we look at the comparison between Mistral's large too and Meta, I mean there's competitive metrics, but the difference in the licensing and the ability for commercial use is a pretty significant one. Right. So while Mistral came out after Metas, you would expect that there would have been, you know, a little bit more hype around it, but there isn't the ability for as many people to use it and implement it with some of the licensing restrictions. And so from an open source perspective, the whole idea is that you can use this as a base to iterate. And I think that that was a business decision that Mistral made, but I think it is in fact impacting kind of the uptake and excitement. Even though the model is very performant.

Yeah, I mean, so, so for, for those who haven't been paying close attention, basically the big difference with Mistral Large 2 is that it's for a non commercial usage that it's open. Right. So that's the biggest difference. Do you think that this is an example of basically big tech in the form of Meta actually leveraging its market position to be able to do, to make a different type of financial decision than Mistral can? You know, I mean it's. Mistral is, you know, well financed, but it's certainly a lot smaller. It doesn't have the coffers that something like Meta does. And they do need to figure out they've been that they have to figure out some sustainable model where, you know, they can continue to advance and make as much open as they can while also making money, whereas Meta can kind of throw that out the window. And you know, especially with Zuckerberg sort of, you know, the legendary control that he has over that company, even relative to Wall street, he just is playing a totally different game when it comes to what he does and doesn't get to give away.

Yeah, well, we're going to talk a little bit later about kind of little tech versus big tech, right. And what some of those challenges are. And I think that this could be one of those examples where, you know, is little tech able to compete in the same way as big tech now? You know, transparency and open source and the ability to be able to make informed decisions about the product is obviously there for both. Right. The weights are published, the context windows are the same. Like, they're pretty, they're pretty similar. Mistral's been pretty aggressive up until now, right. So they've raised a couple of pretty large rounds, they have launched specific kind of task models and there's been a lot of recent discussion around, you know, whether or not generative AI is now approaching or is in the trough of despair of the innovation hype cycle, basically meaning that, you know, all of the money that's pouring in is going to stop because it's difficult to monetize. Right. We haven't seen anybody build a huge revenue generating business from these LLMs. So I don't necessarily think that that is the case. I don't think we're in the trough of despair. But I think that all new technology goes through this really kind of unfair expectation that you should be able to demonstrate application and revenue straight away. If we had those expectations of the Internet, when the Internet came into our lives, whatever, 20 odd years ago, then we didn't. And so it was allowed to grow and become part of E commerce and be part of research and all of the different use cases that have come to light over that period of time. So I think some of it is a bit unfair expectation and some of it equally is, yeah, Meta's big and they've got some money to burn. But I also think that there is an ideology there that potentially maybe stems from Meta's legacy, that it has gone through quite a lot of scrutiny for in fact, having been very closed about their business practices. And so maybe as they move into AI, they've made a conscious decision to be open about it. I don't know.

Yeah, I mean, Zuckerberg is fairly open about his motivation, where a lot of these beliefs came from. You know, like it was very. It's very clear if you take him at face value, that the experience of dealing with Apple specifically really made him think differently about open and closed ecosystems. You know, one can be cynical about that given that he was, you know, sort of in a similar position relative to, you know, Facebook app developers previously. But I think that, you know, it does feel to me, if we take it sort of non cynically that there, there was a, you know, a Road to Damascus conversion experience that he had which is certainly, you know, changed him. Now he's also very smart and very strategic and this was an opening that he had available to him. But whatever the case, it has certainly created a different opportunity.

Well, commercially we, we at Venice were able to make the 405 be available the day that it was published and our users were using it several hours after, you know, it hit the news that it was out. We didn't have that possibility with the Mistral model. And so, you know, we are a consumer facing brand and some of this is just, you know, awareness. Brand awareness. Right. So if you're, if you're, you're, you may be cutting off your nose to spite your face if you're trying to build a moat around what you're doing, especially when you're still little.

Yeah, well, so let's, let's actually this is a perfect segue into the sort of second big theme that we wanted to talk about, which is this idea of big tech eating everything. And you know, it's interesting because to the extent that this meta move is them leveraging their big tech standing and their, you know, revenue coffers to make things more open, I think there's probably a lot of people who, you know, might be uncomfortable with the power that they exert, but are, you know, at least at this moment, you know, pleased with how they are exerting it. Whereas we're also seeing just a wave of consolidation more broadly where it really appears that when it comes to the foundation model game, there is only a very, very small handful of, of companies that are going to be able to compete or seem willing to sort of spend, you know, what it takes to compete.

Yeah.

So the, one of the most, you know, kind of clear parts of this story over the last six months was inflection going over to Microsoft. But now we've also had this, this acquisition or non acquisition, whatever, whatever it actually is of, of Character AI by, by Google. So you know, let's talk a little bit about, about that. You know, what, what your perception of that deal is and what it reflects in terms of this question of big tech eating everything.

Yeah. Okay, so this one is, this one's a little odd to me. So essentially for those who haven't been keeping up to speed, you, it was reported last week that Google came to a non exclusive licensing agreement with an AI chatbot called Character AI which basically allows you to create Very specific Personas that engage with those Personas and also acquire its LLM technology. And so the founders of character AI are previous Googlers. So in this deal the company's co founders come back to Google along with all of their staff and they essentially paid somewhere in the neighborhood of two and a half billion for this. But the deal isn't an acquisition. Now they haven't described why it's not an acquisition, but nonetheless the investors are being bought out and a licensing agreement will exist for the use of character AI's technology and the co founders go back to Google. So I'm not quite sure how that's not an acquisition, but understanding that Google is facing scrutiny and other areas for potential antitrust behavior, I could see why they would like to characterize this character AI acquisition in this way, no pun intended. But it's odd. I mean, it's odd from a market perspective. I don't know what the strategy is here, but it does make me wonder when I look at the technology and the use case today specifically for character AI. It doesn't sit neatly in my mind as part of a Google suite of products. And potentially the use of character AI leans more toward what people tend to use private type AI activities for, not necessarily a Google type of environment which is kind of open for everyone to inspect inside Google. So it's, it's an odd match to me and I don't understand how it's on acquisition.

Yeah, I think that a lot of people have that, that question of the actual mechanics of IT and how these companies think that they're going to avoid scrutiny with this. You know, you and I were talking a little bit before about, you know, I was kind of wondering if it's not so much a consideration of avoiding scrutiny but being able to actually win the case when it, when it eventually comes. But you know, one of the dimensions.

That I think is to kill it.

Right? Yeah, I mean, yeah. So one of the dimensions of this that, that I think is, is interesting and a little bit under discussed is there is a, there is a real strong interest right now, particularly from media to, to have this sort of new narrative of AI entering the trough of despair. Right?

Yeah.

And you see it, it's not just media. There's also like for the first time since ChatGPT launched, Wall street is having, you know, sort of more, more questions. And one of the things that is that, that that has led to with both the inflection acquisition or non acquisition and the Character AI. Non acquisition. Acquisition is an argument that it's just about the difficulty of competing in the, in the sort of, you know, the foundation model space, right. The cost, the high cost of, of compute versus ability to make revenue. And I think one of the things that gets lost is that both of these companies were trying very novel consumer interactive experiences that did not have precedent, right? Like PI wasn't just a copycat of ChatGPT. It was making a bet on a specific type of interactive experience between humans and AI, which we have no reason to, you know, believe or we have no precedent to think is like was certainly going to be a thing. Designing consumer software is extraordinarily difficult, right? These character AI interactions similar and you know, it seems like character AI has done more to validate that this is a type of use case, at least for some people that they're really interested in. But that still doesn't mean that they sort of, you know, can monetize. Well, I think one of the things that's going to be really interesting to see is, you know, in the case of inflection, basically the whole team went to Microsoft. I mean it was pretty inarguably like, I mean they left such a skeleton crew behind. Whereas with this Google acquisition It's only the 30 or there were about 130 people at Character AI. Of those around 30 are coming over to Google, around 100 are sticking around. The 30 that are moving over to Google are the people who are involved in training and fine tuning their custom models. And so it makes sense why at least from a talent perspective, Google was interested in that cohort of people. But the plan as expressed so far from character AI is to instead rely on open source models as opposed to their own custom. I think it's going to be actually really interesting to see if in this case the custom models, the fine tuned models were a requirement of the success or if they actually just hit on a type of user experience where they could build a product around it, where the models that they have access to through open source are going to be totally sufficient for that purpose. And it's really more of a, a product question like I could actually see, whereas Inflection, to me it seemed like it was instantly dead. Basically the pie product, at least for now. Character AI I feel like really does have a chance to still create a going concern if they change the economics of how it's run. So I don't know, I'm less pessimistic for character AI although that's a whole separate conversation than the, the sort of big tech eating everything thing.

Well, I think there is definitely a use case for that type of user experience. There are lots of character AI competitors. The type of engagement is very personal with these type of chatbots. And I don't necessarily see an alignment with a super curated and fine tuned model with a lot of safety and filtering applied to a model that Google would be comfortable putting out into the market. Just take a look at Gemini and a model that would need to support character AI. Right. I don't in my mind see a great match there. So yeah, I don't know what their intention around kind of using only open source for that is going to be, but one thing that might not have registered in the us I'm based in the uk, is that in the UK their merger is now under investigation with the Competition and Markets Authority here in the UK and it's being evaluated for potentially having too much control over large language model activity. So it hasn't been scrutinized in the us but the UK is looking at that merger.

So perfect, perfect segue once again to sort of bullet three, which is antitrust scrutiny. Right. So I think it would be great to hear a little bit more about sort of what you're seeing there because it does seem like the UK is being a little bit more active with some of these cases. But also we did have, even outside of the AI space, Google faced a major decision against them in the us particularly around their interaction with, with other companies like Apple and their sort of, you know, ability to make themselves the default browser. You know, what has this month shown us about the state of antitrust, you know, broadly speaking, and perhaps that in AI.

Well, I think what was interesting, I mean some of the outcomes that we've now seen once the decision has been made public that essentially the law was broken, section 2 of the Sherman Acts, that Google maintained a monopoly over search and advertising. And I think we all kind of thought and felt like, sure, they have a monopoly, but maybe they're just delivering the best experience. And that's why. But within this ruling it is claimed that Google spent billions of dollars to create an illegal monopoly to become the world's default search engine. Apparently spending more than $26 billion in 2021, in 2021 alone to companies such as Apple to become their default search engine on devices. So you know, that is potentially anti competitive. And going back to the questions around meta, are they just big enough that they can afford the fine or give it away for free? That's quite a lot of money to be spending. So. So in addition to that there was claims made that Google had had a history of deleting internal communications and setting all their chats to automatically delete. And these were messages that might have not served them well had they gone to trial. So this is really kind of looking at business practices, but when you put that in the context of again, these large companies having a lot of control over information and now that the capability around AI combined with that information I think is really compelling. Right. Which to me only makes the stronger case for open source and decentralized AI because you can put regulations in place, you could put rules in place, but ultimately if you have decentralized open source AI that kind of regulates the market in itself, you don't need governments to come in and, and tell these companies what they can or can't do because the market will decide. So as opposed to these kind of, it sounds like potentially anti competitive practices that may have taken place.

Yeah. So I think one of the things that's so interesting about this to me is right now there are, you know, there, there are a couple different buckets of antitrust scrutiny when it comes to AI. One is, you know, I think in particular the deal between Microsoft and OpenAI. There's a lot of scrutiny around that. The inflection thing sort of adds another dimension to the Microsoft story. But then the other area is Nvidia. And Nvidia seems difficult because on the one hand they do absolutely control a ridiculous portion of the sort of market share. But at the same time there's a sense that it's because their product has been differentiated, they just are literally that far ahead.

Yeah.

And you know, it's interesting especially to compare to the Google decision because a lot of, you know, to your point, a lot of the things surrounding Google, it wasn't just that their search engine is dominant, it was specific business practices that made it that way, whereas I think that did it.

So that's the.

Well that's, that's the question. Right, Yeah, I mean, but you know, so I think that Nvidia, it sounds like the. So we haven't gotten any sort of real new things happening within video other than that they seem to be gearing up for this. Right. They didn't have an office in Washington. They now do. They didn't have any policy people on staff. They now do. It's things like that where it feels like they sense an inevitability around this.

From many fronts too. So just to give a sense of. Outside of the U.S. the French competition authority is investigating Nvidia for A reason similar to what you were describing. So they're concerned about the sector's dependence on Nvidia's CUDA chip programming software. So not just the chips, but the software, because it's the only one that's 100% compatible with the GPUs that have become essential for accelerated computing. Right. So they're not concerned about access to the chips, they're concerned about the programming itself. There is an investigation by the DOJ on the run AI acquisition. And so that's one thing. Doj is also investigating them on a separate matter related to their business practices, again specifically asking about whether they create conditions to limit access to their chips based on the purchase of other products or commitments to not buy other products from consumer or from other competitors. So there's a whole bunch of things happening with Nvidia and around the world. The European Commission and the UK put out a joint statement in July with the DOJ and the FTC that they were looking at concerns about a few companies having all the necessary resources to compete in this space. So, I mean, one, they're big, right? And so that's going to attract scrutiny. But equally, I think that they're, it sounds more like the business practices themselves are what are under scrutiny, not just that they're big, right? And I think there is a chicken and the egg discussion around did I get to be the world's largest search engine because maybe I did some things to ensure that that happened or am I big like I'm now Nvidia and I'm trying to maintain market share and, and maybe employing some of these like, which is it, did they get big because they did something that they shouldn't have done? Or now they big and they're getting scrutiny because they're big.

Well, and so this gets to, let's use this as a bridge into sort of our final set of, of conversations, which is, you know, political developments, broadly speaking. But I think that the, the specific lens that, that it may be most interesting is this idea of big tech versus little tech. So obviously this has come up. You know, this was a, Teresa and Horowitz folks are, you know, they sort of wrote about this, they kind of, you know, put some, put some language in context around it. For them, it was specifically in relation to the US Presidential election cycle. But then you also have these sort of it playing out in other ways as well. You had Lina Khan from the FTC showing up at Y Combinator and talking about the importance of open source. Right. And she's been one of the sort of loudest, you know, antitrust, you know, kind of advocates or opponents in some ways for, for big tech. And so that's an interesting time.

She spoke at Y Combinator. She spoke last year, November I think, not specifically on AI, but on this kind of big tech little tech challenge.

Yeah. And so, so how, how would you frame this for people who haven't been keeping up with this particular conversation?

So essentially the idea is that is it possible that conditions have allowed the biggest technology companies to gain an advantage in AI? And she makes the case that if you control the raw materials, then you control the market and then you have the ability to exclude smaller companies and thwart innovation. Essentially she made the quote that I have here is that open weight models can reduce costs for developers so that they can focus their capital on products and services rather than expensive model trading. And they can free up venture capitalists to pursue promising new applications of models rather than starting at square one with model development, which we know is the most expensive part of LLM revenue generating activities. So I think that she made it very clear that she supports open source weight models rather than AI models that kind of claim to be open source but don't make their weights available, which I think is fair enough. And obviously an example of that is Meta's 3.1 where the weights are made available. So you know, again, she's been speaking about this for a while and I think she gets a general, you know, good response from the industry. But I think it's not as simple as little tech and big tech in my mind. So to me the argument is like, okay, so little tech equals more competition, which antitrust regulators, which she is believe supports consumer protection. Sure, I get that. But does little tech not ever aspire to be big tech when they grow up or hired right firm's not looking to succeed and exit. I would argue that they are. And to me little tech isn't an ideology. Right. It just happens to be a moment in time. And so to kind of lean into the narrative that little tech good, big tech bad to me doesn't resonate. I mean, I think there's fair market activity going back to, you know, the other things that we've been talking about. If you're, you know, engaging in anti competitive business practices, whether you're big or small, that doesn't create a fair and open market. Right. Where consumers and businesses can make informed decisions. And I think being big or small does not prevent you from conducting yourself in a appropriate and transparent way.

Yeah, I tend to agree I also think that there's all sorts of downstream implications for this that make the sort of philosophical piece of this fall through a little bit. Right. It's like, you know, little tech exists because there's capital that wants to go into little tech to fund these experiments and iteration. But if there's no M and A market on the other side of that, if there's no IPO market, you know, which right now it's M and A is the vast majority of exits for, for startups. Right. To get to that point. But if that market freezes up because you know, I mean, basically I think that there's a. It's worrying to see all these companies do this weird non acquisition acquisition things. You know, the people who get hurt in that scenario are the 100 character AI employees who weren't part of that, you know, crew going over to Google. And you know, if it becomes normalized for founders to take these big paydays and just drop their companies, it's going to have dramatic implications for who's willing and not willing to, you know, to be a part of startups, to fund startups, you know. So I think that there are, you know, it's not just that it's a spectrum, which I think is a key point, it's also that it's all part of an ecosystem, you know.

Absolutely. Well, and the other concern that investors start to have, I mean right now there are investors that are concerned after all of the what are seen to be kind of FTC crackdowns on M and A because they're kind of cutting off a key piece to the startup ecosystem by allowing these companies to kind of exit, right. And be acquired. If there's no path, then you have a whole bunch of companies that potentially won't actually realize not just the company's potential, but maybe the technology's potential. Right. There has to be a path for growth and that doesn't necessarily mean an exit or an acquisition. Some people build a product and they intend to, you know, deliver it and into a publicly traded company, right? But to again say little tech kind of good, we support little tech and big tech bad, that implies that you don't ever expect those little tech companies to actually perform and be revenue generating and that doesn't make any sense. So somewhere in there there is, there's a tension that needs to be resolved.

Absolutely. I do think that if you're looking for sort of the optimistic take on this, the encapsulation of little tech as a thing, as, you know, even if it is A state in time or sort of a type of category that should be encouraged and protected. In some ways there's probably positive manifestations of that. You know, as long as they don't sort of fall into these, you know, unreal binaries that, that sort of, you know, don't, don't recognize the reality of the situation. Right.

There's, if you look at other industries like fintech, for examp, you know I used to work for a bank, I worked in payments and there was a period in time 15 years ago where it was a big deal like these fintechs are coming for our lunch. Right. It was very contentious and large banks were doing everything they could to either invest or acquire some of these fintech companies to either plug in their, their systems into theirs and kind of hyperscale their innovation timeline internally within the bank. But more often than not they were buying them to kill them. Right. So over time I think we have seen that the fintech ecosystem has become very known for doing, you know, a specific activity. Well, like Stripe does payments really, really well. They can be very efficient and they can cater to different types of markets that maybe a large transaction bank isn't going to cater for. Right. Subscription businesses for example. So there is different ways to kind of slice the pie. And you know, Stripe isn't going to go after trade finance activities that one of these large transaction makes may undertake. But that whole thing, you know, that friction again that I'm talking about probably took 12 years to shake itself out. And now we don't talk about, you know, these little startups coming to like eat banks lunch. It is a well understood imperative that we have these fintech companies that perform these functions really well, very efficiently and economically and it forces larger banks to just do better.

Super interesting conversation. We could, we could wander down that path forever. But to wrap up, I want to kind of actually now flip you around and look forward. You know, we are, we're in August, we're kind of in the, the last quiet month of summer, quiet as it can be. What are you watching out for? You know, either with the sort of the, the open versus closed question in specific or just AI in general.

Well, the one thing we didn't have time to get to I think was to talk about some of the, I don't know if it's you know, regulators remorse, but there has been some commentary made over the last couple of weeks around the AI act in Europe becoming, you know, just showing itself to be far too stringent. And one of the demonstrations of that is you know, Meta has announced that they are not going to roll out their new models to, to Europe, period. Right. Not to businesses, not to people, not going to have access to it. And so I think that is a really. And then you've got the author of the AI act, you know, giving a interview to Bloomberg three days ago. I think it was basically saying, yeah, this, this is, this isn't going to work.

I've got the quote, they've got the quote here. The regulatory bar maybe has been set too high. There may be companies in Europe that could just say there just isn't enough legal certainty in the AI act to proceed.

Like Mistral.

Yeah.

French based company. Right. So, I mean, we didn't, we didn't get into that. But is there a reason why, you know, they're, they're trying to maybe maintain a little tech position? Because once they get to a certain point, then a lot of these provisions within the AI act kick in and become very complicated for them. So I think that we're looking at now a presidential election looming. We potentially might have some shifts from a Biden administration to another administration and their views on AI. How are these kind of political trade winds going to affect that? But I think what we are starting to see is countries that have taken a view on this and maybe have taken a view too soon are already starting to say, oh, this maybe it's not going to turn out the way that we expected it to. And I just don't understand why Europeans aren't a little bit more concerned that they're literally being potentially cut off from the next major innovation that, you know, humanity is going to experience. It's crazy.

Yeah, it's, it's wild. Well, you know, maybe next month, depending on what happens in California, we'll have another context to discuss something very, very similar to that. But for now, that was a great conversation. So appreciate you, you hanging out and spending some time.

I am. Love it. I love chatting to you and it's been a blast.