The more this site thinks about law firms training open weight models and renting their own GPUs, the less it seems to make sense. Here’s why.
Latham has got hold of its own GPUs to train open weight LLMs, meanwhile Kirkland is working with Palantir on its ‘own’ AI system, although as AL noted, it’s hiring staff with GPU experience, therefore it’s likely that they are going down the open weights route as well.
But why? Latham mentions in last week’s FT article that some client data is very sensitive so they need some of their AI capability to be on prem, or close to that. It also noted ‘consumption costs’, i.e. tokens. And Kirkland mentioned at the time of their announcement the importance of having their own thing, rather than relying on systems used by many other firms.
But, do these reasons stack up? Let’s explore some of the reasons why a law firm may or may not want to train its own open weights model.
Cost
Let’s do the big one first. A multi-billion $ law firm’s token costs are notable, but not something that would change the game for them. Legal tech companies are trying hard to develop routing systems to reduce reliance on the most expensive models. And nothing is stopping them from tapping a cheaper model directly either.
Then you have the cost of hiring staff that can work on GPUs, then running the project – which will not be simple, as data must be chosen carefully, many tests made, those results then merged with the outputs of other AI systems, and there’s the cost of the GPUs as well. On balance, is there really a significant cost saving that could not be found just by using less expensive models?
And then, how much of the work of the lawyers at this firm needs a special application of AI for them? E.g. you want to do some legal research? Use LexisNexis, TR and Clio. You have a licence already. Plus, TR has its own open model already now, Thomson (which for them makes sense as their business is selling access to curated data). Your token costs are not changing for that – there is just a licence fee. And how often in a transaction do token costs become such an issue that you or the client really gets upset? Very rarely – and much also depends on how those costs are shared out.
Secrecy
Do law firms hold information that is secret? Yes. Does that information have to go into any LLM at all? No.
There is no reason why some very special information needs to find its way into any LLM, whether you control it, or someone else.
You hold this information in your DMS, for example. Can it be added in later to any project a law firm is working on using AI? Probably. Can it be searched internally without using an LLM hosted by someone else? Yes. You don’t need to rent your own GPUs to do a search of your super-secret files.
And how much secret info is there? Will that really be used to train an open model? And if so, why? How much of this sensitive data is needed? And on that point, if a client’s data is so wildly important then why are you co-mingling it with training data from other clients in an LLM, (even if it’s one you ‘control’)?
Or is the idea just to have an LLM that you run locally that has no special training and you just want to have that on prem barrier? Well, you can already have on prem LLMs – there are companies that offer this, and you can have a small model too running locally. Again, no need for you to go out and rent GPUs.
There is really no need to buy GPUs, or rent them, if this is all about secrecy. In fact, the best way to keep a secret secure is to keep it off-line in general. So, if this is all about secrecy, then it’s not that compelling as a reason.
Helps the Clients
Does training your own model help the clients? Possibly. If you can tap your data in a better way than before and that improved insight creates a marginally better outcome for the clients, then yes. But…..there are plenty of legal tech companies, from DeepJudge, to Aloi, to DraftWise, to many others, which are helping law firms dig deeper into their data and past documents, and to tap what we can call the judgment layer.
There is no need to build your own LLM to do this.
Plus, as we have seen, in most cases, improved efficiency around the production of legal work – and surfacing past data is part of this – is not leading to lower bills for the clients, or work completing faster.
Improved efficiency is being used to reduce the impact on the firm of unbillable work, or low billable work, and sometimes when the client really is under pressure to move super-fast, e.g. race through a DD exercise on a fixed fee with lots of AI thrown at it. But the latter use is not the norm, or even when it does occur, the rest of the deal is priced as usual on the hour and so there is no net cost saving for the client.
So, clients likely won’t see any cost savings from law firms building their own open weight LLMs, and if that is the case, then why do it?
AI Sovereignty
Law firms are understandably proud of their ‘data heritage’ – i.e. the accumulated knowledge of decades of work. In fact what is a law firm?
- The lawyers who are there today. Their knowledge, their ability to work, their insights and genius.
- The record of past matters – if that data can be organised and tapped and surfaced in a useful way.
- The client relationships – as without those there is no need for a law firm, as there are no clients to pay for your services.
But, does that data heritage vanish if you don’t have your own AI sovereignty to go with it? No. It’s still there.
Does using a big legal AI platform lead to that IP leaking away? No. They are not training on the very specific wording of certain special clauses, they’re trying to figure out how to improve the entire application layer to help lawyers.
Moreover, the real power in all of legal AI are the frontier models – and they do not care about a specific clause here or there either – they’re trying to get language understanding and sustain agentic work to a level of reliability that makes AI totally trustworthy for knowledge-based industries.
If OpenAI can see that X bank uses Y wording in a certain contract does it change the world for them? No, not at all. What OpenAI and all the other model makers want is to refine how genAI works to the point that it’s capable of performing tasks as well as the most experienced humans. And that will be based on overall architectural changes, improvements to the chips, new ways to post-train the models before release, and other ‘macro’ factors.
In short, your sovereignty as the owner of useful legal knowledge is not under threat from AI. What is under threat is that at some point in the future the general ability to do any, and all, legal work will be so commoditized by AI tools that have become wholly reliable, that the legal market changes forever.
But….even then, X law firm will still hold Y client relationships and Z knowledge.
Marketing
A senior person from a major legal tech company contacted this site over the weekend to note that Big Law is now echoing ‘The Innovators’ – AL’s serialised story about a law firm grappling with AI, which includes the issue of marketing. (See more here.)
The news about Kirkland and Latham is there for the clients to see. It has a marketing effect.
This comes as Big Law is facing a confusing time. Its entire business model is based on getting things right for a very large fee. Efficiency has never been part of the deal. And that was OK before, because there was no realistic alternative.
AI changes this because it is all about efficiency. And that is a problem for anyone who sells time.
Sending the message that you are using AI in new and exciting ways, which perhaps might help the clients….although perhaps may not, in order to make the clients feel you’re ‘doing something’ is a trend we will see more and more of.
Law firms don’t want clients to:
- Do more legal work themselves by tapping AI on their own.
- Send work to the NewMods that offer to do the same work, very quickly, much more affordably.
- Or, see other Big Law firms using AI to raise their game and offer the clients the same work as they do, but perhaps also more affordably.
In short, Big Law is fighting against an incoming tide of commoditization.
Doing things that seem as if you are putting a wall around what is special, of protecting what the clients have given you, and also showing that you are at the cutting edge, may stave off – for a few months, maybe years – this incoming tide. But, it won’t be a lasting solution.
Conclusion
Building your own LLMs does have some positives, but on balance – especially in the context of what’s available already in the market – those positives don’t seem that significant.
That said, if two firms are doing this, then more will do it.
Richard Tromans, Founder, Artificial Lawyer.
—
It will soon be time for the Legal Innovators conferences in London and in New York, both this November and taking place over two days!
Come and join us in London this November 4th + 5th at Legal Innovators. The conference at the intersection of legal AI and the business of law. Day One: law firms, Day Two: inhouse.
Legal Innovators UK – London, Nov 4 and 5
And,
Legal Innovators New York – Nov 17 and 18
See you at Legal Innovators New York – Nov 17 and 18 – join us at the intersection of legal AI and the business of law in New York for the most important legal innovation event of the year! Day One: law firms, Day Two: inhouse. Top speakers from across the legal AI world, leading law firms, and major companies. All for one purpose: to describe and explore the frontier of legal technology. See you there!
UK and New York event links are here:
Discover more from Artificial Lawyer
Subscribe to get the latest posts sent to your email.