The AI Daily Brief: Artificial Intelligence News and Analysis - The Four Wars of the AI Stack: The Battles Shaping the Development of AI

Episode Date: January 14, 2024

The wars include the Quality Data Wars, the War of the GPU Rich/Poor, the Multimodality War, the RAG/Ops War Inspired by this excellent framework from our friends at Latent Space https://www.latent.space/p/dec-2023 ABOUT THE AI BREAKDOWN The AI Breakdown helps you understand the most important news and discussions in AI.  Subscribe to The AI Breakdown newsletter: https://theaibreakdown.beehiiv.com/subscribe Subscribe to The AI Breakdown on YouTube: https://www.youtube.com/@TheAIBreakdown Join the community: bit.ly/aibreakdown Learn more: http://breakdown.network/

Today on the AI Breakdown, we're discussing the four wars of the AI stack. The AI breakdown is a daily podcast and video about the most important news and discussions in AI. Go to Breakdown.network for more information about our YouTube, or Discord, and our newsletter. Hello, friends, happy weekend. Today, we are going to have a very fun little episode. And first of all, I have to shout out the crew that came up with the framework for this episode, which are the fine folks over at Layton Space. Laton.Space is a great newsletter. You should absolutely check it out. They are big in the AI engineering space. They're

driving a ton of exciting concepts there. They also have the best AI podcast for engineers, which again is called Layton Space. The host are Alessio and Swix. They've been on this show before. And I absolutely recommend you subscribe, like I said, both to the newsletter and to the podcast, latent. They just came out with a long piece called The Four Wars of the AI Stack. Now, where they took this is a very long an awesome article about all the different elements of it and recent news, and I think that you should go check that out. What I'm going to do is just explain their framework, because I think it becomes a really interesting way to look at, broadly speaking, the big trends in the AI space. So the four wars of the AI stack are, one, the quality data wars, two, the war of the GPU

rich versus pores, three, the multimodality war, and four, the RAG slash ops war. We're going to start with the quality data wars, as I think this one will be very familiar to listeners of this show. This is basically the battle between content platforms on the one hand and the content raiders on the other. In other words, the AI labs who are training their models on all sorts of data, some of which publishers think they shouldn't have access to without paying. So, the belligerence in the quality data wars are, on the one side, the journalists, the writers,

and artists, and on the other side, the AI researchers, the AI startups, and the synthetic data research when it comes to leaders and flag bearers. And again, this is all the latent space framework, which I love. On the journalist-writers' artists' sides, you have the New York Times, stack overflow, Reddit, giddy images, Sarah Silverman, etc., etc., etc. On the AI researcher's side, you have OpenAI, Stability AI, Microsoft, Common Crawl, Axel Springer, and many more. Now, this is a battle that is being fought, in my opinion, on two very different battlefields. One is in the court of public opinion. It is very clear that publishers and writers and artists are taking their case directly to the people. They are being loud about their issues with

how tech treats them. They are playing on previous concerns and anxieties and frustrations with big tech, concerns and frustrations that have grown over the last decade, and they are translating that to this new generation of companies that are focused on AI. Meanwhile, the folks who are on the AI side of this are basically trying to argue something very simple, that treating training AI models on copyrighted materials as fair use is net better for the world. They are trying to argue that any exact competition to those publishers, writers, or artists in the form of complete mimicry is a mistake and something that can be minimized technologically. Now, ultimately, whatever the public decides to think about this, it will be an issue that is actually solved in

the courts. The New York Times lawsuit against OpenAI is, of course, the biggest and highest profile lawsuit so far, but there are many other happening across multiple dimensions of this space. What's more, whatever the decision in that case, it seems very likely to me that it or another case like it makes it all the way up to the Supreme Court. I just don't think that we're going to get resolution on this AI training copyright question without invoking the biggest court in the land. Now, interestingly, as this all happens, as these battles play out in both the courts of public opinion and the actual courts, the combatants are also still trying to work with one another. OpenAI has signed deals with companies like Axel Springer. Apple, it appears, is trying to

spend a bunch of cash to have this be a comparative advantage for their AI. And so there's also the possibility that it resolves itself from a market standpoint, but ultimately, I do think it is going to be a court that determines who wins that particular war. Next up is the war of the GPU-rich versus the GPU-pours. This is about faster, cheaper inference, fine-tuning, and training. The belligerents are on the one side, the GPU-rich clouds, GPU-rich manufacturers, and VC-funded companies, the invidias, the Googles, the Microsofts, the Amazon bedrocks, the fireworks.aIs, the together AIs, the N.E. Scales, and the replicates. On the other side, the belligerents are the GPU-Poor AI engineers, the edge and local compute, the new

model research. The leaders and flagbearers include modular, MLC, TinyCorp, K-Lora, R-slash-Lama, consistency models, Apple slash MLX, Mamba slash Striped hyena, etc. Now this is all about, on the one hand, the attempt to throw ever more computing power at more advanced large models to try to drive new capabilities that way. That is, of course, what you have the Googles of the world going after. On the flip side, is an attempt to do more with less, to create new architectures and new approaches that allow for faster and cheaper inference on less powerful devices. Now, just to be clear, it's not just a division of who has the most money and the most resources. There are reasons, even big money reasons, to be really interested in that more with less side

of the battle. That's why you heard that Apple, and some of its recent research, is on the belligerent side under the GPU poor AI engineers. It's because they're trying to figure out how to run more advanced and more powerful models directly from your iPhone without having to touch the cloud. For them, that is the big prize. Now, this is a battle in which both sides can be winners.

It's not necessarily in either or. It's mostly just a very different framework and a very different perspective on what actually matters going forward. Third, we have the multimodality war, which the latent space folks define as specialist models versus everything models. The belligerents on the one side are text to image startups, text to audio startups, and other modalities. It's the mid journeys, the playgrounds, the 11 labs, the 11 labs, the Higens.

On the other side, our OpenAI and Google DeepMind, more or less. The question is, as we race towards a multimodal future, will it be these everything models like ChatGPT, like Google Gemini, that become the standard bearers? Or will people still use workflows that incorporate lots and lots of different specialist models. There is certainly a lot of exciting innovation on both sides of this. Obviously, OpenAI and Google are spending untold and effectively infinite amounts of resources trying to build God-type models, but the competitors on the other side are no slouch either. Pika Labs 1.0 has been impressing people. HeyGen has incredible avatar creation. Suno.a.I got people actually excited about music generation. I don't know whether in this

space there is one side wins or not, but I'm glad that there seems to be no sure. shortage of resources for those specialist companies to keep pushing the field forward in their particular areas. The final war of the four wars of the AI stack is the RAG slash ops war. Put differently, they say databases versus frameworks versus dev tooling. The belligerents include three different categories, general databases, vector databases, and search slash record systems, LLM frameworks, LLM platforms, and AIEF standards, and MLOps, PromptOps, and LLMOps. Now, this goes way beyond the technical scope of the AI breakdown, but is basically all about how AI actually gets built into the computing and enterprise systems of the future.

There are very different approaches to how people are going about building in this space, and it's not clear if one of these categories will rise above the other. Now, one war that I might add to this stack, which kind of is touched on by that last one, is the enterprise adoption war. And there are lots of different ways to look at the various belligerents in this. You could divide it between, on the one hand, top-down enterprise adoption versus bottom-up enterprise adoption. CEOs deciding that their company needs to use AI

versus individual workers simply bringing it into what they do. Another way to look at that would be big SaaS incumbents versus startups. Companies like Salesforce, Amazon, Google Cloud, Microsoft Cloud, Microsoft Azure, companies that already have the affiliation of the enterprise, are kind of bringing the AI solutions to them as well, whereas in previous generations of technology that might have happened from startups. And then, of course, just within that, even if it is those big SaaS companies, the battles amongst them are also significant. Anyway, like I said at the beginning,

I just love this framework of the four wars of the AI stack. I think it's such a great way to think about these big different dimensions of the AI space and the big trends and battles that are shaping it. So once again, just one reminder. It's the quality data wars, the war of the GPU rich versus the GPU pores, the multimodal war, and the rag slash ops war. One more time, go to latent.space to check out those guys' newsletter and their podcast. Thanks for the great framework and excited to have you guys back on the show soon. Appreciate you listening as always, and until next time, peace.