The Four Wars of the AI Stack (Dec 2023 Audio Recap)

**Alessio** (0:06)
Hey, everyone. Welcome to the Latent Space Podcast. This is Alessio, Partner, and CTO in Residence Invitable Partners, and today I'm joined just by my co-host, Swix, for a new podcast format.

**Swix** (0:17)
Yeah, it's a bit uncomfortable because we have to just stare into each other's eyes, lovingly, but in our end-of-year survey last year, a lot of listeners were asking us for more one-on-one time, more opinions from the both of us as hosts on what's going on in AI. You know, both of us are very actively involved, and I don't think this year will be any different. This year, there's lots more excitement to come.
And we're trying to grow Latent Space in terms of the types of formats and the amount of value that we deliver to our subscribers. So one thing that we've been trying, experimenting with, is this monthly recap that I started doing around August of last year, where I basically just take the notable news items of the month, and then I sort them and categorize them according to some order that makes sense and write them down in the newsletter.
And this last December recap was particularly exciting because it seemed like it popped off in a number of areas, particularly with the AI breakdown. Our friend NLW featured it on his podcast. And I figured we can just kind of go over that as a way of setting the stage for 2024, but also recapping what happened in 2023

**Alessio** (1:24)
And people always ask me if December is like a slow month, but I think you almost broke Subsec with how many links we had in the thing.

**Swix** (1:31)
No, we actually did. So a lot of people commented to me about the formatting issues within the newsletter that I sent out. And I know that they are there, but I couldn't fix it because Subsec was broken by us with how long it was.

**Alessio** (1:42)
But we had this kind of like four main buckets called the Four Wars of the AI Stack, data quality, and I guess like data quantity as well, in a way. The GPU rich versus poor, which we have a whole episode about with Dylan Patel, multimodality.
We're actually recording tomorrow with Luma Labs about their new 3D model. So we went from text to image to 3D video. I wonder what's next. And then-

**Swix** (2:07)
We're gonna release Hugging Face as well, as I guess I've been thinking about calling it multimodality 101, because the first modality beyond text that you should really pay attention to is vision.

**Alessio** (2:16)
Right. Yeah, and then the rag ops were, I think that's a-

**Swix** (2:20)
I don't know what to call it. I don't know if you would have called it anything else. This is my-

**Alessio** (2:23)
The tooling were, I don't know. But I think beginning of last year, that was like kind of the hottest space because there wasn't much open source model work. And I think over the last maybe like four or five months, everybody's so focused on fine-tuning Lama 2 and like a DPO to improve these models, max trial and all these things. And people forgot about our friends at Length Chain, Lama Index, and some of the things that were maybe top of mind. VectorDBs, you know, it seemed like everybody was releasing a VectorDB early in the year.

**Swix** (2:52)
Yeah, I think that I- I'll be very surprised if any new VectorDBs come out this year, with one exception, which is something I'm keeping my eye on, which is TurboPuffer. I don't know if you've seen them going around.
Yeah, all the smart people seem to be adopting TurboPuffer as the first serverless VectorDB.

**Alessio** (3:07)
Yeah, no, and we're going to have definitely Jeff and Antoine on the podcast at some point. I know they're going to be-
they're going to be fun.

**Swix** (3:16)
I should also mention, the reason I selected these four wars was a process of elimination of wars that I think ended up not mattering. So for those who don't know, inside of my writing, I often include footnotes that are in themselves, just essays in footnotes.
And so I think it's also notable, the things that people thought were hot, that were less hot than expected. So it was agents, definitely less hot than at the start of 2023 And then this one is very controversial, non-selection by me, I think. Open source AI is not a battle in the sense that I don't think there's anyone against open source AI. Everyone is like on one side. There's no like opposing side apart from regulators.
But in my mind, when I think about like for engineers, engineers are all universally in favor of open source models. So there's no battle here. Everyone just wants it to improve. So it's not like interesting to write about. We just want more open source.