00:00:00:00 - 00:00:30:03
Eden Shochat
Think of what happens instead of selling keywords, like Google AdWords, you start selling user embeddings, right? Who the user is, what the user is. Entire industries, from adtech to media, just get disrupted. There's a deeper layer, right? Once you work on the embedding level, you actually—it's like machine code. You can actually make the models do things that they didn't consider, which is both exciting and unsettling.

00:00:30:03 - 00:00:31:03
Eden Shochat
Da da da da.

00:00:31:03 - 00:00:36:14
Unknown
Da da da da da da da da da da da da da da da da da da da da.

00:00:36:14 - 00:00:51:10
Unknown
Oh oh.

00:00:51:13 - 00:01:20:22
Eden Shochat
So in order to explore that, I'm joined with Gilad Lotan, Head of AI Data Science and Analytics in BuzzFeed, which feels fitting because very few people actually know how to turn buzzwords into actual signal better than Gilad. I've actually known Gilad for 30 years now and was excited to see some of the early Twitter clustering work that even today you wouldn't actually see in most companies.

00:01:21:00 - 00:01:24:22
Eden Shochat
And so it is a great pleasure to have you.

00:01:25:02 - 00:01:27:14
Eden Shochat
So you like welcome.

00:01:27:16 - 00:01:31:07
Gilad Lotan
Thank you, Eden. It's great to be here and it's always so nice to hear.

00:01:31:07 - 00:01:33:21
Eden Shochat
So are you ready for a quick round of questions?

00:01:34:11 - 00:01:36:04
Gilad Lotan
Sure. Sure.

00:01:36:06 - 00:01:37:19
Eden Shochat
So, AGI, when?

00:01:37:21 - 00:01:42:01
Gilad Lotan
AGI when, oh my gosh, I—okay, I'm supposed to be quick.

00:01:42:01 - 00:01:53:17
Gilad Lotan
I go back and forth on this. Sometimes I feel like it's in two years and sometimes I'm like, we are crazy to think that it's coming anytime soon.

00:01:53:17 - 00:01:57:02
Gilad Lotan
I know, maybe ten years, 20 years, ten years.

00:01:57:04 - 00:02:05:21
Eden Shochat
You're very pessimistic, that's for sure. But then who? That's a bigger question in my mind. Like, if China, U.S.…

00:02:05:23 - 00:02:09:11
Gilad Lotan
I think China. I think China. Time to learn Mandarin.

00:02:09:11 - 00:02:31:11
Eden Shochat
Yeah, that's the crazy part, right? What we're seeing is that China is publishing articles, U.S. has stopped publishing articles. China is actually speeding up. It's going to be pretty brutal geopolitically, but unfortunately, I think I agree with you. Unless the U.S. regroups.

00:02:31:13 - 00:02:35:21
Gilad Lotan
It's going to be—it’s going to take us back.

00:02:35:23 - 00:02:42:16
Eden Shochat
To figure this out as, in this new world, OpenAI or Google? I think it's Google. You think it's Google?

00:02:42:18 - 00:03:10:04
Gilad Lotan
It's Google. You know, it almost—it feels like they've had kind of a comeback the last few months. They've launched Gemini. They've launched some really, really amazing capabilities. Definitely on the image side, with their new Gemini models. I think also their labs are doing—it’s pretty performant and they're subsidizing this to produce on their stack.

00:03:10:04 - 00:03:22:18
Gilad Lotan
So it's very cheap to use. And yeah, they already have the advertising business. So I do think—I don't think OpenAI. I think as OpenAI, I don't care about the advertising business.

00:03:22:20 - 00:03:35:18
Eden Shochat
Well, they are starting to go into recruiting, which is pretty interesting because it's a swarm of cats, right? It's like matching between people. But let me—so maybe Google, but what's the winning ad unit in the AI world?

00:03:35:20 - 00:04:01:13
Gilad Lotan
I mean, I think it is created by a service, not by a human. The ad itself, I think it is tailored to a person. So the ad that you see is different from the ad that I see—maybe the image, maybe the text—but I think it's hyper-personalized and automated content that's hyper-targeted.

00:04:01:15 - 00:04:12:10
Gilad Lotan
I think that is the world we're heading to. And then the question is, whether impressions will matter or will everything be about clicks? I'm not sure. Yeah.

00:04:12:12 - 00:04:32:08
Eden Shochat
Well, even more than clicks. Right? I would argue that it's actions now that you can actually track. And if it's agents, that's interesting. So what feature are you most missing from your favorite—out of them, you mentioned Gemini innovating finally. But what are you most missing? What have you recently started using?

00:04:32:08 - 00:04:33:17
Eden Shochat
Let's start with that.

00:04:33:19 - 00:05:05:11
Gilad Lotan
We've actually started using Replicate pretty heavily, and it’s effectively a marketplace that makes it easier to use different models. They host, they handle the inference, they make it very easy, and pretty cheap. And so as a content company that's rapidly building, we don't have to host our own version of Stable Diffusion and just deal with uptime and inference.

00:05:05:11 - 00:05:29:02
Gilad Lotan
We don't have to do any of that. We can say, okay, we want to try v0? Let's just use this endpoint from Replicate. We want to try Flux? We can use that. So it's very easy to test and try and shift between models. In a way we are moving much faster than we could ever have moved because we're using this platform.

00:05:29:04 - 00:05:35:18
Eden Shochat
And what are you most missing? If you had a magic wand, you could get anything you want tomorrow?

00:05:35:18 - 00:06:09:19
Gilad Lotan
Yeah, I think the ability to fine-tune—but truly fine-tune. And also, maybe it's fine-tuning plus evals. I'm really missing evals, because I feel like everything we do, or a lot of what we do, ends up being evaluated by vibes. You really want to ground your evaluation. There are various approaches and unit tests or prop tests we could run, but it's still—it takes a lot of time.

00:06:09:21 - 00:06:28:00
Gilad Lotan
And also fine-tuning has not—I'd say on the text side, text has not—we haven't been successful yet with regard to tone or text or different tasks that we’d like to try out.

00:06:28:02 - 00:06:46:06
Eden Shochat
We should have a whole episode about evals. That's another one of the most underappreciated things: just knowing what works, what doesn’t, and how to evaluate a new model when it comes out—what’s changed, and what broke.

00:06:46:06 - 00:06:58:01
Gilad Lotan
Yeah, I don’t know. Are we making progress? I know it clicks. Okay. But is this riskier? Yeah, no idea. There’s so much uncertainty when you launch new products.

00:06:58:03 - 00:06:59:21
Eden Shochat
Yeah. Not totally.

00:06:59:21 - 00:07:02:06
Eden Shochat
This is awesome. Okay, I love this.

‍