The alpha, twice a week
That was 3 min.
Get the best new moments in your inbox every Wednesday and Saturday. It's free.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
Training Data · 3 min · Mar 11, 2025
Harvey tests multiple models instead of relying on one
Winston Weinberg
Winston Weinberg says model leaders change and different models suit different tasks.
In this episode
Full analysis- 5:26Harvey builds specialized workflows, then unifies them3 min
- 9:09Legal AI needs process data and expert evaluationPro pick 4 min
- 25:25Prestige and tailored demos help firms trust new AI3 min
- 16:01Harvey is moving from seat sales to selling legal work6 min
- 29:51Law firm review layers change AI’s minimum quality bar2 min
- 32:13Routine legal work may shift to fixed fees as advice gains value3 min
- 35:09Legal rules limit access, but some states are testing changes3 min
- 41:27Harvey tests multiple models instead of relying on one3 min
“we don't just use one model for the entire system”
Winston Weinbergat 43:01 · Training Data
Takeaways
- Different components of a legal workflow can use different models.
- Harvey continuously tests models from multiple providers.
- Customers’ use of other AI tools shapes their expectations.
About this segment
Winston Weinberg says model leaders change and different models suit different tasks. Harvey tests providers and combines specialized components to manage that shifting landscape.
Winston Weinberg is cofounder and CEO of Harvey.
Co-founded and leads Harvey, after working as an associate in securities and antitrust litigation at O'Melveny.
Sonya Huang, Host · Partner, Sequoia Capital
Works with OpenAI and Anthropic; co-wrote Sequoia's essay on generative AI.
Pat Grady, Host · Co-steward, Sequoia Capital
Joined Sequoia in 2007; backed Snowflake, Zoom, ServiceNow and OpenAI.
From “How AI Breakout Harvey is Transforming Legal Services, with CEO Winston Weinberg”, Mar 11, 2025. Watch the full episode
Thanks. We'll look at and fix what's off.
Transcript
Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.
- 41:30
what has it been like and how has it evolved over time as it's gone from kind of undiscovered to being the center of the universe in many ways. Yeah. I mean, so I think the thing that we have kept up that has been really awesome for the engineers at Harvey, but also at OpenAI, I think, is we have always been working on things that I think a lot of companies aren't. And so what I mean by this is OpenAI would give us models all the time and say it performs way better on all of their benchmarks. And we would respond and say, sorry, but it doesn't on ours. Like it's actually not better for us. And I do think that the reason our relationship has been so strong is we keep saying, OK, here's a model that can do X, Y, Z. I would like it to do X, Y, Z and the rest of the alphabet, too. Right. And that's what we're trying to do. And so I think we've actually
- 42:24
helped them a lot, too, with at least like applied use cases and how they think about post-training and how they think about what are some of the things that companies are really pushing to try to do right um and so i think that has been really really good for the relationship overall do you have any hot takes on microsoft open ai um yeah um one of the hardest things about being an application layer company is you have to bet on model providers um and it is not like we build our system so that you could kind of just pull out one model and replace it for another but you could do pieces of that right so like we don't just use one model for the entire system you'll build a piece that's really good at this and then a piece that's really good at this and then you'll chain them all together right um but the reality is
- 43:13
you know who's in the lead changes so often um and different models are good at different things right and you have another problem too where all of your customers or most of them i said that partners probably aren't using a bunch of ai tools but the associates for sure are and the associates are using all these different AI tools, right? And so what I'm trying to say is we, I think, have done a good job where we work with all the model providers and we are constantly testing out models with all of them.
- 43:42
And I think that Microsoft, which has, you know... such a massive customer base, they have to figure that out too, right? They have to figure out, oh, wow, we have customers that think that Claude is better at certain things. What do we do about that, right? I mean, they can't really do anything about that. But they're like, you know, Mistral or whatever it is, right? And so I think that that relationship is evolving as it naturally would. If the opportunity for Harvey is to revolutionize a trillion-dollar industry and provide better access to justice, what is the threat?