What would it take for a biglaw firm to hire like 2 engineers for 3 months to stand up this free, open-source version of Harvey/Legora? More secure (local), cheaper, likely just as good. As the creator notes in comments, Harvey has moved away entirely from fine-tuning the models they use because the labs come out with new models every 3 months that are much better than anything Harvey can put together.
Related Posts
ADA roadmap update tomorrow!
Has anyone sold at IQVIA? What’s comp like?
Additional Posts in Big Law
I’d rather quit than revise this agreement.
New to Fishbowl?
Download the Fishbowl app to
unlock all discussions on Fishbowl.
unlock all discussions on Fishbowl.



Enthusiast
First — if a firm were to build a competitor, it wouldn’t be in their interest to license it under an OS license.
But as to the general concept of making an OS Harvey-type competitor, I’m sure there are some OS skills.md that get at what Harvey and others do.
I see Harvey having two moats over OS:
- refined, thoughtful workflows.
- data (precedent, playbooks, case law).
Both can be overcome, but the OS community may make it harder to converge on design given that it would be by committee.
As for the second point, OS isn’t really great for data from a legal standpoint. You need an authoritative source to depend on.
Good point on relying on an OS license. On the data point, I thought Harvey had moved away from RAG and tuning its models. It just relies on frontier models, which are all better than any Harvey-tuned model. Anthropic has come out to say that RAG is now outdated and probably causes models to overindex on information. If the data moat you’re referring to is exclusive deals with Lexus or whoever to access their info, then I agree that’s a real moat.
Community Builder
Its pretty amazing
Coach
While harvey is marketed as a vertical ai saas product, insiders told me that it functions more like a thin Chat GPT/Lexis wrapper. Basically, the tool does a simple API call to lexis or gpt to generate responses, with minor pre and post processing of the prompts and the resulting output
Yup
I’m told by our AI committee that the ABA rules require that if we’re going to put client data into a model, (1) training on client data be turned off and (2) human review for fraud and abuse also be turned off as it’s a breach of confidentiality. For #2, it’s virtually impossible to get a frontier model to agree to this so Cowork is not an option. But somehow Harvey has been able to secure these commitments from its underlying frontier models? Curious if anyone else has heard similar.
Coach
partner 1 does not know much of AI