Hey everybody, Richard Tromans here again, Artificial Lawyer TV, doing a special interview today with Steve Hasker, the CEO of Thomson Reuters. Thanks for being with us, Steve. It's really appreciated.
Hi, Richard. Yeah, good to be here.
So we're recording this a few days actually before the big announcement which will accompany this video. But let's get straight into it. So Thomson, which Artificial Lawyer, as you know, wrote a big piece about a couple of months ago, tell us more about it. What is Thomson, and why is it such a big deal for Thomson Reuters?
Yeah, thanks, Richard. Thomson, or Thomson-1, as we call the first production-ready version, is a large language model that we've built in-house here at Thomson Reuters. Essentially what we did was, in August 2024, we came across a startup which was centered in and around Cambridge University in the UK. It was a group of former Google DeepMind scientists who were building a large language model for legal. We got to speaking with them and pretty quickly reached the conclusion from both sides that SafeSign ought to become part of TR. And so we bought the company and brought the scientists across. For about 20 months, we've been investing heavily in the development of the Thomson LLM, large language model, with a view that we wanted to create a vertical-specific large language model—one that was purpose-built for the legal profession. Of course, we have a couple of advantages in doing this: one is access to tremendous amounts of content and editorial expertise through Westlaw and Practical Law in all the jurisdictions in which we operate.
The second is tremendous subject matter expertise across many, many legal practices in many jurisdictions. So roll the tape forward 20 months or almost two years, and we put out the first production-ready version last week, and we'll talk a lot about this next week at ILTACON. Essentially, the evaluation suggests—and we've done a very robust set of comparisons to the leading iterations of the frontier models from all of the names that you will recognize—Thomson-1 is outperforming most of those models for general domain tasks. Set aside the legal task for a second; for general domain tasks, it's performing at or above the very latest cutting-edge models from the frontier companies. And then for legal tasks, as you'd expect given our content and our expertise, it's right up there as well. That's with only a fraction of our legal content applied. So we think this is a pretty interesting breakthrough, Richard, for the industry. It provides us and, I think more importantly, our customers with all sorts of options going forward. So we're very excited about showing the industry the model and the results of the model, and then continuing conversations as to how we might bring it to bear for the benefit of the market and of our customers.
Fantastic. So let's dig into this. Right, first of all, Joel Hron was telling me about this a few weeks ago: effectively, these are open-source, open-weight models that you have trained, that you have post-trained. Can you say which models you were working on?
So we started with a couple. I won't go into the specifics of which ones because we have evaluated many and incorporated the open weights of a number of them. But essentially, well known to you, Richard, is that there are some very, very powerful, very thoughtfully constructed open-weight models available for the research community to then build on and post-train on. And so we're able to start with one or more of those open-weight models and then apply all of our expertise and some of our content to create what we think is a really compelling offer.
Gotcha. And when you say the performance is better, this is in terms of just pure accuracy, it's in terms of relevance, it's across a basket of different criteria?
Exactly. If you go back to Joel and Richard Schwartz's blog post earlier in August, they basically go through a series of criteria: it's accuracy, it's speed of resolution of different issues, it's around information retrieval and analysis. Essentially, we look at the performance and the output of the Thomson-1 model relative to, as I said, the cutting-edge frontier models, and we love the results, particularly in its first iteration.
Gotcha. In terms of just one last little technical question, did you have to bring in your own GPUs? Did you rent your own GPUs? Did you buy your own GPUs? Was that not necessary?
Yeah. Given our size and scale, we think we have an advantage in terms of our access to GPUs relative to our customers or relative to many of our competitors. And so we're able to rent and, in a sense, ring-fence our own GPUs to create the model. So we've done that within the cloud environment.
Gotcha. Okay, makes sense. So looking forward, let's look towards benefits. Hopefully incrementally better accuracy, better relevancy, and so forth. And of course, if you're doing legal research—which is one of the key pillars of both Thomson Reuters and the other major legal research companies—obviously accuracy is paramount, so that's a clear win as well. Does this then follow through into your other sort of productivity tools? These days, Thomson Reuters has gone way beyond litigation research. You're into transactional work, into a whole bunch of different things. Does Thomson spread into these areas as well?
Yeah. We think it's an essential addition to our AI stack. If you think about, at the front end, CoCounsel, which is our legal AI assistant, supported by Westlaw Advantage, supported by Practical Law in its AI form, this essentially is all the way back to that source model. We think it's going to give us a couple of different options, and these are the conversations we're having with our customers. The first is, starting next week, we will port across our tabular analysis, our bulk document analysis function, onto the Thomson model. We see that as having a couple of clear advantages: the first is speed, so the latency is significantly reduced relative to accessing one of the third-party frontier models; the second is accuracy, given that the Thomson model was created for this profession and specialized set of tasks; and the third is a cost advantage. We've all read about token maxing and tokenomics.
I was going to ask you about that, but tell us about the token side.
Yeah. Essentially, because it's trained on our own content and the Reuters news file and a whole series of other relevant content sets, to put it in layman's terms, it is much more efficiently constructed than a large language model that's scouring the web and going through Wikipedia or Reddit and all sorts of other sources. This one is much more tailored and specific, so it ends up being more efficient. If we can get the capacity utilization of our GPUs up to anywhere near 100%, we think relative to accessing a leading frontier model, we're going to have a significant cost advantage, which of course we'll be able to pass on to our customers.
So the first iteration of this, Richard, will be to run tabular analysis within CoCounsel on the Thomson model. If that's successful, we'll then move more and more of the functions of CoCounsel across, and we could even eventually move Westlaw Advantage and some of the functionality, including that which we're building, which is Brief Builder, which is a very, very exciting innovation in the pipeline at TR. We could move those things across to the Thomson model. That's the first avenue we're pursuing and discussing with our customers. The second one is—you've written about and are well aware of this idea of sovereign AI.
A big subject—and ensuring that law firms and general counsels protect their IP, or their alpha, as some people call it. Essentially, the conversations we're having with some of our most sophisticated customers is the opportunity to put the Thomson model in behind their firewall and have them co-mingle it with their information, their IP, potentially run an instance of CoCounsel on the top of that, and really create a sovereign AI solution with and for them that they control, and that they are 100% assured that they're not sharing their IP with a frontier model and they're not sharing their IP with a startup which is a reskinned frontier model.
To say the least, that is a topic that is becoming near and dear to the hearts of firm chairs and managing partners and CTOs. That's a second, separate, but very interesting potential opportunity for us.
Yeah, I mean, that opens many, many, many doors, doesn't it? You'd have to allocate space for that law firm effectively to migrate some of their own data to co-mingle, as it were, with your data. How would that work?
Let's assume that within the public cloud environment, they can create an environment where we can put an instance of the Thomson model into that environment that they run and control. We constantly update that model, so it's a one-way feed of information to ensure the model stays current. They then add their IP—whether that's access to their document management system and any other information repositories that they have—and they can run applications in that instance that's created and managed and reserved specifically for that firm. Again, the attraction of that is elimination of the risk of leakage of their IP to a frontier model or a startup, or heaven forbid a competitor, or a competitor to a client who's benefiting from that IP because it's leaking out in some way, shape, or form. So that's the attraction to the firms.
And also so much simpler. A certain very famous law firm, which we both know the name of, is going down a more DIY route. I imagine that this would be way more economical to work with TR on this than try to build the whole thing yourself in your backyard.
Yeah. There are very few firms that have the sort of size and scale, financial inclination, or at the partnership group level the inclination to go down a DIY path.
But there is the interest in sovereign AI, as I said. So I think you've hit the nail on the head: this is one way in which they can meet both of those needs. They're not creating a completely bespoke DIY instance that they may or may not be able to maintain in the fullness of time, but they're accessing TR's.
Yeah. You put your finger on a really good point there, which is maintenance. There's a great big extravaganza, firework display, everyone's excited, we rolled it out. Great. Now you got to run it for the next 20 years.
Yeah. We've seen this, Richard, on the tax and accounting side of our business. As some people may be aware, we're one of the largest providers of tax calculation engines in the world. Every once in a while, one of the Big Four—and this has been the case over decades—will come to us and say, "Okay, we'd like some content or we'd like some software IP, but we're going to build our own solution." Where we can, we'll support that. But almost invariably, what happens is a year or two or three later, that same firm comes back and says, "Hey, would you like to buy this from us? We've developed this application. It's working to varying degrees. Would you like to buy it?" When we ask why, they always say, "Well, turns out the building is one thing; it's the maintenance which is another."
The practice of law and the idea of running a law firm is, at its essence in my view, assembling great people to meet or exceed the needs of clients around particular matters. In our view, the technology needs to serve that rather than usurp that or replace it. If you keep that in mind as the chair of a firm or the group of managing partners, then you're laser-focused on your people. You're laser-focused on having the technology serve your people, and you're not laser-focused on building out bespoke instances of software and technology that you may or may not in a year or two's time be able to maintain. I think it's particularly the case at the moment, given how fast things are moving.
There's a landscape that you and I understand today, or at least we hope we do, in terms of the frontier models, the agentic harnesses, the application layer, and so on and so forth. But that's going to change, and it's going to change rapidly. Six months, it'll be different; six months after that, it'll be different. If you're managing a law firm and you're trying to meet or exceed the needs of your clients, worrying about how fast the technology is moving is a whole separate problem that you may or may not have the wherewithal to get across.
Yeah, indeed. And also, this brings up another point which I'm continually writing about, which is data. Looking at this in particular in the context of the February stock market shock, where Anthropic released a whole bunch of Claude skills and a whole bunch of other people went, "Oh my god, oh my god, this is the end of the tech world." Of course it wasn't, and I argued in an article that a lot of the responses were illogical because surely companies that have a lot of data are actually the ones, if anyone is going to get through this, who are in a good position. I just wondered how you saw things playing out.
The wild swings we've seen in terms of the perceptions around the value of legal content, in one sense I would say we've seen it before, in that Google Scholar back in history was going to commoditize all the legal content that was out there. And then we've had over time various new entrants, some of which have been very well funded, sort of try but largely fail to make a real impression in terms of depth and breadth of primary and secondary law content. I think this is no different. What I would add is this term that we coined, "fiduciary-grade AI," which we use to describe AI that is providing content and solutions to professions that cannot afford to be wrong, like lawyers, like tax preparers or auditors. Essentially, those professions require deterministic answers, or as close to deterministic as possible.
The frontier models themselves provide probabilistic answers, and so the risk of hallucination for the practice of law—set aside the business of law, it may well be that these horizontal tools can be very effective in supporting lawyers in terms of running their business—but when it comes to the practice of law, high-stakes litigation, high-stakes transactions, the output needs to be very accurate firstly. Secondly, it needs to be verifiable, validatable, if that's a word. Essentially, it cannot be a black box. The agent needs to describe step by step what it's doing and, equally importantly, the citations, the sources of information that it's accessing in order to reach a particular conclusion before going on to the next step. That's what we do for a living with our content, our expertise, our data privacy, and our customer support mechanisms: we provide fiduciary-grade AI, which is different than that which a frontier model can provide.
