Latent Space: The AI Engineer Podcast
00:27:20 · Alessio Fanelli, Jacob Lauritzen, Shawn Wang
[00:27:20] S2: I would have assumed you guys signed some sort of agreement with the law firm, and they're like, oh yeah, [00:27:24] S0: I wish, yeah, I wish, no, but the thing is, like, a lot of the data that they have is not even theirs, right, it's their clients, and it's super confidential, so we can't look at it. And so, have an evalist team that's, like, trying to approximate what we think everyone [00:27:36] S1: is doing [00:27:36] S0: and, like, build that into eval sets. Wow. So that's just, like, a super annoying kink of the business. [00:27:41] S1: Can you do paraphrase? For example, ChatGPT has like a slash feedback thing that directly fully sets its full trajectory. But obviously, that would... That is not possible for you guys. But you can use an element to paraphrase, like, from this real thing Yep. Make a synthetic thing that is, like, close enough, but change all the names, change all the details, but, like, still...
00:27:41 · Shawn Wang, Jacob Lauritzen, Alessio Fanelli
[00:27:41] S1: Can you do paraphrase? For example, ChatGPT has like a slash feedback thing that directly fully sets its full trajectory. But obviously, that would... That is not possible for you guys. But you can use an element to paraphrase, like, from this real thing Yep. Make a synthetic thing that is, like, close enough, but change all the names, change all the details, but, like, still... [00:28:03] S0: Yeah. Yeah. We can't... There's... I mean, there's always a range, right, of, like, if you if you paraphrase all the way down to, three values and you pick one of them, then obviously that's fine. If you if you just remove the names, then that's maybe maybe not fine. Yeah. It's like, we are generally pick the the most safe way. Yeah. Because, like, we we restrict ourselves as most possible because we, you know, there should be no risk at at all. Yeah. But there are things where it's like, okay, we know that the failure modes are probably these 10 or 20, so we can classify it into one of those, and that's a helpful signal. We can classify it into the types of tasks and the practice areas also, and then, like, we have an idea. Something that we're always trying to to figure out how to do in a smart way. [00:28:42] S2: Also, like, on the privacy front, I guess, I'm assuming even two years ago, it must have been, like, insanely difficult to get the law firms to [00:28:49] S0: be okay with this. Yes. [00:28:49] S2: But, like Yep. How do they think about their data being sent over to Anthropic or Yeah. Yeah. It must be, like, a huge thing.
00:28:49 · Alessio Fanelli, Jacob Lauritzen, Shawn Wang
[00:28:49] S2: But, like Yep. How do they think about their data being sent over to Anthropic or Yeah. Yeah. It must be, like, a huge thing. [00:28:56] S0: All of our infrastructure is on Azure Yeah. Primarily for this reason, because, like, when we started, law firms were like, AWS? No. Like, [00:29:04] S1: no way. Law firms don't trust AWS? [00:29:07] S0: Yeah. Yeah. It's different now. No. But there was like Amazon, Google, we can't trust them within our data, so basically, we had to host everything on Microsoft because they already have three sixty five. [00:29:16] S2: Oh, so they're okay with Microsoft? [00:29:17] S0: Exactly. Because they've already done, like, whatever security review, etcetera, for all of their documents and all of their Word and whatever, so Microsoft is an approved vendor in their mind, and so we had to host everything on Azure, and that's changed over the past few years. So, now people understand, okay, we need to get on cloud, we need to do all this kind of stuff. There was also a lot more people talking about on prem back then. That's also disappeared completely. People understand that, like, you should just do it in the cloud, it's and it's safe, and it's probably safer than what you're doing. Yeah. So that's changed [00:29:43] S1: a lot. [00:29:43] S2: But I
00:29:43 · Alessio Fanelli, Shawn Wang, Jacob Lauritzen
[00:29:43] S2: But I [00:29:44] S1: was And by the way, it's only been three years, so, like, it's actually, like... Yeah. It's a... Yeah. Yeah. It's a shoot shift. It's it's it's great. I mean, good for you guys, but yeah. I I don't expect the, like, you know, the legal firm, legal industry has somehow ignored this for the last, like, ten years. Yeah. Yeah. I think [00:29:58] S0: we needed the AI kicking the butt for everyone to understand that we need to change Okay. Our ways a little Okay. [00:30:04] S2: So also then on, like, open source models now are getting really good. So, how are you thinking about that with the... [00:30:10] S0: Yeah. We're just thinking about, to to your point, just bring our clients along on, like, the Pareto frontier. Yeah. And, like, the the map of all the models and what they're good at, and their latency, and their cost, and their governance, and and and their regional availability, how do we just optimize across all of that? So, some open weights models are great for certain tasks, and then we'll use them there. Some will post train, and and some will just use the frontier models. [00:30:36] S1: Yeah. You call... You call it model arbitrage. Yeah. You have different favors. You you did some research on this where you were firmly anthropic at the start of the year. Yeah. Grok is improving a lot. Yeah. You wanna give an example of a model that wasn't ranked highly overall, but, like, you wanna shout it out as surprisingly good?
00:30:36 · Shawn Wang, Jacob Lauritzen, Alessio Fanelli
[00:30:36] S1: Yeah. You call... You call it model arbitrage. Yeah. You have different favors. You you did some research on this where you were firmly anthropic at the start of the year. Yeah. Grok is improving a lot. Yeah. You wanna give an example of a model that wasn't ranked highly overall, but, like, you wanna shout it out as surprisingly good? [00:30:55] S0: Yeah. I think not overall good. Gemini, one of the early Geminis is, yeah, extremely good for certain types of transcription and speaker attribution. So we use that for trans... Transcription actually. [00:31:07] S1: Oh, so you have a [00:31:09] S2: Yeah. What you gonna transcribe? [00:31:09] S0: Like a granola? We don't have... We actually probably will build, like, a meeting thing Yeah. But this is just, like, talk to speech, speech to text, text to speech, just like in in any other app that has this. And we do actually also transcribe files. If you upload hearings or whatever, we'll then transcribe them, put them into text, and so we can do all of our analysis on it. ElevenLabs. ElevenLabs, for DPA reasons. Same thing as like the data processing agreement, I should say. So like, our data processing agreement has very few vendors on it. [00:31:37] S1: Yeah. You just need to keep [00:31:38] S0: it Exactly. Need to keep it super tight on where any data goes. And so
00:31:38 · Jacob Lauritzen, Shawn Wang, Alessio Fanelli
[00:31:38] S0: it Exactly. Need to keep it super tight on where any data goes. And so [00:31:43] S1: I think, no, that that is one of the benefits of, like, a Gemini and OpenAI where they're like, you you come to us, we have most of the things that you need, and that is actually very valuable. Even if it's like... It doesn't have to be, like, state of the art. It's always just, like, mostly good enough for all tasks Yep. Then who cares? [00:31:59] S0: Yep. Yeah. No. Absolutely. I should actually shout out also Microsoft's transcription model. They made one Transcribe? Yeah. Okay. It's great. [00:32:08] S1: Okay. [00:32:08] S0: It's better than the old Gemini. [00:32:10] S1: Yeah. Yeah. Yeah. They they they did some good results. I don't know... Don't even know. Like, I think, like, usually, all of them are, like, top two, top three, top four in terms of, like, the the arena scores, and that's good enough. [00:32:20] S0: Yeah. Exactly. And it's like, you have all the governance and it's super scalable, and it's like, then that's probably good [00:32:24] S2: enough. Yeah. I was gonna say, like, at this point, are pretty much the speech to text models, like, good enough to where you're like, oh, one word here or there, if it's off, [00:32:32] S1: that's No? [00:32:33] S0: No? No. [00:32:33] S1: For sure. Like, I've done this a lot.