NCI: Winners & Losers In The AI Race

Ram Ahluwalia & Michael Parekh

Thank you for Listening this Episode!

Support our podcast by spreading the word to new listeners. We deeply appreciate your support!

Episode Description

In this episode Ram discusses with Michael Parekh (former Goldman Sachs Partner), on the winner and losers in the AI race. They dive deep and dissect the strategies of tech giants (hyperscalers), chipmakers (semis), and even governments, all pouring in billions for AI dominance. Is it a war of wallets (capex) or operational brilliance (opex)? Plus, we cut through the hype surrounding small AI and edge AI – are they the future or a passing fad? Tune in to find out.

Episode Transcript

Speaker 1 [00:00:02] Humanoids. Full service FSD robots. Small AI. Winners and losers in Max seven data centers. Geopolitics. I am so thrilled to be joined by Michael Prec, who is, really an intellectual tour de force, retired Goldman Sachs partner. Michael helped bring many of the famous internet companies we know from history to the public markets. When he was at Goldman Sachs, your contemporary, of course, of Mary Meeker and Henry Blodget, who are, you know, Morgan Stanley and Merrill, the time. So I met Michael. That's right. We met at a conference a few weeks ago. And I stalked Michael. We had dinner for four hours. I walked Michael back to his hotel. The next morning breakfast, I sat down with Michael again. And Michael has gone deeper on I in semiconductors and what this means for social change than anyone I know. And we spend a lot of time on this. I've interviewed quite a few guests around AI and, you know, social change on our podcast limit non consensus investing. So it's a real treat. Thank you Michael for joining us today a lot to get into. 

Speaker 2 [00:01:19] It's a pleasure to be here. And you're a glutton for punishment. So back. 

Speaker 3 [00:01:23] For more. 

Speaker 1 [00:01:25] Know we enjoy very much. All right so first off any reactions to the OpenAI demo. Some other topics we want to get into, but there's up for the audience will be, you know, meta llama three, grok. Small AI Qualcomm. What does this mean for geopolitics? Humanoids. What are they coming? Are they real or is it just fluff and hype? What are the value unlocks we might see, in the future? Who's going to win? Search. Can Google maintain their pole position? There's a lot of great stuff to get into. So let's start off. Go ahead, Michael. 

Speaker 2 [00:02:01] Sure. I mean, in terms of the OpenAI announcements yesterday, they obviously wanted to, play a little bit with Google's, I o conference today and, you know, get their get that little shout out before they did a good job, they announced, for. 

Speaker 3 [00:02:16] Oh. 

Speaker 2 [00:02:17] For optimal, I think and, it's pretty impressive. It's, it's really optimized for. Voice and multimodal. It's very natural sounding. If you watch the movie her, it's, absolutely a facsimile of that in terms of the whether it's a female persona or a male persona, it's very natural sounding. Voice changes. You can interrupt it. It's really fascinating exercise in terms of how fast they've been able to innovate on the technology. It's not just the large language models, but the text to speech, the intonations, the ability to see things and interpret things off camera. It's, it's a tour de force in terms of building on a lot of capabilities on top of their core models. And very important, bring down the latency, very natural, real time, like a person, etc.. I would encourage everyone to watch the demo. It's like 20 minutes, 20 to 25 minutes. It's really, really good. And it's a harbinger of things to come. Everyone's going to have to play for these table stakes very, very quickly. 

Speaker 1 [00:03:27] They raise the game. You know, they were never multimodal to begin with. Google was they close that gap. They demoed a translation feature and Duolingo stock price dropped as they were demoing that. So we're already seeing the disruption of size play out, you know, not just Adobe and what's been impacted there by, Midjourney. So we're eager to see, what happens today. With Google. You know, one of the comments you made, Michael, when we spoke, I asked you about AGI. When will we see it? And you made a distinction between AI reasoning with a big an AI reasoning with a smaller. Do you want to elaborate on that? 

Speaker 2 [00:04:10] Sure. I mean, the first of all, in terms of the quest for AGI, this is a very potent question up and down, across the spectrum from people who are optimists about AI to pessimists to tumors, etc. regulators are very, focused on it. The founders of these companies are very focused on it. And the debate is, ongoing. It, the projections anywhere from the next 2 or 3 years to ten years. I did a piece this Sunday on my Substack that people can look at in terms of the range of opinions on this, but in terms of the reasoning piece that, we were discussing when I talked about it as a bigger or smaller, the reasoning piece is important because fundamentally, the way that pure large language or small language models work, it's these are probabilistic tables that they're doing billions of calculations a second and running probabilities on the next token. The next part of my word that I'm about to speak and giving you the best guess. They don't. These models are not sentient. They're they're we shouldn't treat them as humans. They're wonderful computation machines. And yet. And they cannot yet reason. So we we think of them as artificial. We call it artificial intelligence. The reasoning piece is, is, is being built. It's, what you need to do is to. Connect the probabilistic to the deterministic side of things, meaning you have to give it additional elements, hints, programs, data to help it understand the world factually, and then use the, probabilistic elements to help you reason better. So very complicated. I'm sorry, I, I'm getting a little bit too wordy there, but when I say the big and the smaller, the smaller just refers to being able to make straightforward. Straightforward guesstimates of what they're seeing, and the models are getting very good at that. The bigger the bear you're expecting, the human is expecting these systems to work across applications, domains, etc. requires a lot of connecting of all of those environments so that the AI systems can cut across all of that and then use all of the information it garners from all of that to reason better. And so that is going to take time. It's it's, as much a function of, building the computer science underneath as well as connecting everything. And that just takes time. 

Speaker 3 [00:06:50] Right. 

Speaker 2 [00:06:51] Across systems. 

Speaker 1 [00:06:52] Right now. Today we're using the the garage framework retrieval augmentation. 

Speaker 3 [00:06:56] Generation. 

Speaker 1 [00:06:57] To call the appropriate sub functions like a calculator to simulate reasoning or actually to teach AI the laws of physics. You know, there's a great, anecdote by Yann LeCun talking about how a two year old absorbs a couple of terabytes of data in a couple minutes as compared to what? I mean, Jensen's got a five year outlook on AGI. And, you know, it's questions. I define these things, but I expect we'll be on the sooner timeframe. What to find out and revisit that. How do you think about OpenAI and their competitive position. Right. They're owned by Microsoft substantially. It's got a financial interest. Microsoft's in the enterprise market also. OpenAI has got complexity, sort of perplexity, competing, competing on one side of it, they've got Google coming after them from behind. And they've also got Meta Llama three, which is open source and also partnering with grok. Then you've got Cahiers out there, you've got anthropic out there. I mean, how do you think about OpenAI's competitive position? 

Speaker 2 [00:08:04] Don't forget the Chinese learn models. There is at least 4 to 6 of them that are coming out strong over there. 

Speaker 3 [00:08:11] All right. 

Speaker 2 [00:08:12] No, you're totally right. Look, I mean, if this is a multi-billion dollar, race across so many companies and systems and sovereigns are getting into the equation, OpenAI is in the lead. They are in the lead. And as you said, with GPT four, they just announced GPT four. As we talked about 5.0, they did confirm, somewhat that that will be out later this year. So that was a new development that is expected to be probably ten x or more, depending on which variation of the model they release. Yes, Microsoft is a partner. They own about 49%. It's a complicated governance structure, non profit. 

Speaker 3 [00:08:53] Margins which is okay. 

Speaker 2 [00:08:54] But denying 49% of the profits. So that it's not a there's no board seat other than observe receipts etc.. So there's a there's there's there's a lot of common elements that are different than just directly owning a company. So OpenAI is a bit of a very independent company, and they're growing like a weed. And they're phenomenal in product innovation. Right. So it's not just the GPT 4 to 5. It's all the additional things that they've been announcing and working on. I've written about at least seven that I can count, and probably more that we don't know about products and things that are coming in the next, literally next, three to 3 to 6 months. 

Speaker 3 [00:09:34] A. 

Speaker 1 [00:09:34] Rapid product from OpenAI. So. 

Speaker 3 [00:09:39] You're just that's just. 

Speaker 2 [00:09:41] Yeah, this is just OpenAI. So they're they're supposed to, come out relatively soon with their AI for search. So competing more directly with Google on that front. They will they've got a voice product that they've been showing off. It's not yet released. Saw a text or video, which is which is a pretty phenomenal product, although the Chinese have caught up with that. But this is a very fast rate. So anthropic, which is the second company in the, in the space and very, very well capitalized with 4 billion from Amazon and a couple of billion from Google. You've got as you mentioned, Google itself is is very potent here. They'll have big announcement tomorrow. So yeah. 

Speaker 1 [00:10:25] So let's let's talk about meta. All right. So a meta lemma three. 

Speaker 3 [00:10:30] Yeah it was then I want to yeah. Goodrum. Go ahead. 

Speaker 1 [00:10:38] Go ahead. Yeah. 

Speaker 2 [00:10:40] You know, you were saying about murder. 

Speaker 1 [00:10:45] Oh, sorry. My video cut out here. I think we get a lot of people watching the live stream here, saying, you know, unmet. You know, you know, you mentioned also when we spoke, I think it's a great insight around how the training inference loops have compressed dramatically and meta has a competitive advantage, giving the sheer volume. You want to double click on that. And what that means for the edges that I might have now in in the future. 

Speaker 2 [00:11:08] Sure. Mark Zuckerberg is very focused on being a very big player here. He he's garnered probably the second largest, collection of the Nvidia's H100 chips, over 350,000 of these things at $30,000 or more a clip. And he's using it aggressively to, roll out Lamar three, which is their open source model in several sizes. He rolled out a couple of months ago, not just Lambda three, but meta AI, which is their equivalent of the chat bot that is now on top of WhatsApp and, Facebook and Instagram. That basically allows, 3 billion people, depending on how they're rolling it out across the world, 3 billion people to bang away on this thing. And because when people start to use these models, that starts the inference loop that's on, my tech stack chart that I show all the time on my, on my Substack, the inference loops, the, the prompt. And that is very valuable because what that does is it provides additional information to the underlying models to make the results a lot better. And and the more these things are used, the better the model gets as a result of that of that loop. And meta, because they're exposing their meta I AMA three model to over 3 billion people. They're they're aggressive in the race. The negative of that is that you've got to use a lot more processing and compute. You need a lot of processors, and that's expensive right now. And that hopefully over time we'll get cheaper. So meta is, is is very much front and center ahead in the race. OpenAI, as we said technically is in there is is ahead with the best models. They have about 100 million monthly users relative to, let's say, 3,000,000,000 meter users. Being able to use this app, let's assume 10% of them are using meters. Search meter AI every day. As a just a guesstimate. You know, that's, that's probably 2 to 3 times the use of, of, OpenAI models. You've got Apple, reportedly from just, concluding a deal with. OpenAI. It's not officially announced. It'll be in June. So for Siri to be powered with OpenAI, that's the that's the leaked, report. So that's 2 billion people on Apple, devices potentially over the next year using OpenAI services. So the race for users is on in the billions. And then, of course, you don't forget Google Gemini. They're going to be, out talking about, their products tomorrow and Google Gemini reaches billions and they have their own huge processing with their tensor processing units, their TPU architecture, which competes with the Nvidia stuff. And so they have billions of users using Gemini models. And their front ends out front and center is rolling out their large language models on Gemini, Gemini for search. So yeah, in the next 12 to 18 months, expect billions of people, mainstream users for the first time. Just not just hearing about AI, but trying AI every day. And that's the big change in model. 

Speaker 1 [00:14:23] And when we think about that. 

Speaker 3 [00:14:26] I sort of. 

Speaker 1 [00:14:27] Define a moat. You know, the couple of things you want to look for. One is data, right? Google's got a ton of data, and YouTube has got a ton of data generated from users. OpenAI. Well, they everyone has some basic generic data, but OpenAI, maybe they're tapping YouTube. Who knows. So data is one. Two is distribution. You've talked about that. Three is compute. The race is on to get compute. But fourth is like networks. And meta has a network. And I don't believe that is appreciated how we think about AI it. Apple has distribution distribution and a network I believe are different things. Is that a distinction that's that's relevant here? 

Speaker 2 [00:15:06] No. That network is a very key piece because what you're intimating there is that people are talking to each other, connecting around topics and discussions and so on. And inasmuch as AI gets into the mix of that with these loops that I talk about in my writings and so on. It's it again, all of that helps in the reinforcement learning curves and what's called Rag, the reinforcement augmentation. These are all buzzwords. I'm sorry to go techie here, but all these are signals to the model to keep getting better and so on. So whether it's humans talking to the models, whether it's machine queries, talking to other queries into into these loops, all of that makes the models better. One of the things, you know, back in the 90s, right when we took Yahoo public in 95, etc., we at Goldman AI, we talked about we started talking about things like B2B and B2C, which is people connecting, you know, these systems connecting businesses to consumers and consumers to consumers. One of the things with AI that we need to be watching as a metric is just beginning is what I call m to m machine to machine these agents talking to each other. When you talk about networks, oftentimes increasingly as we get into these things called smart agencies are more buzzword smart agents a generic workflows. 

Speaker 1 [00:16:27] These are autonomous agents that can pursue a goal as opposed to prompt response. 

Speaker 2 [00:16:32] Exactly. So you know right now when ChatGPT if unless I type in something or ask if something with a with a voice, it doesn't do anything. It's like a search thing. But now let's say, you know, my favorite example is I want to buy a mattress. Not tomorrow. In three months. I set up an agent, look for shop for mattresses for me. And every week it just goes out and looks from things and just puts it into my bucket wherever it is. And then shows me the stuff that's. And it's then. So my machine is talking to somebody else's machine or a whole bunch of machines and coming back and aggregating that data. That's a different type of network that you were talking about, right? So in other words, in that sense, Amazon could have a network, right? Because now my query, my smart agent on Amazon, could be querying the millions of third party sellers of things on, on, on their machines and their agents. So networks are very important. I think you would you're making a distinction between networks and distribution. Networks are very key. We're just at the beginning of the beginning of these networks of agents starting to talk to each other. And then the question from an investor point of view is, well, where's the where's the beef? Where's the money? And look, there are three ways to make money out there on the internet advertising transactions, subscriptions, three likes to the store. And when you when you've got so far on the internet, we've got call it three 4 billion or the 8 billion people on the internet through smartphones and so on. So far it's been mostly people, people talking people to businesses. But now let's say there's an average of ten agents for one of us and there are 4 billion of us. So in about 3 or 4 years, there's 40 billion agents talking to each other. And that can all be monetized with advertising and subscriptions and transactions. 

Speaker 3 [00:18:28] What's the term then? 

Speaker 1 [00:18:29] How do you define the Tam? 

Speaker 2 [00:18:31] So tam tam is humanity. I love the word tam because it's it's everything in. 

Speaker 3 [00:18:37] The market for the audience to.