Thomson Reuters already has dozens of genAI experts, why did it just buy 10-person legal-specific LLM developer Safe Sign? Joel Hron, the giant company’s CTO told Artificial Lawyer the reason: because despite assumptions, legal-specific LLMs – with the right training data – work really well.
First the deal. As reported by AL yesterday, TR has bought the small, UK-based Legal LLM or L-LLM developer, bringing with it former A&O Shearman trainee lawyer Alexander Kardos-Nyheim and AI expert Dr. Jonathan R. Schwarz, plus a team of other engineers and legal experts.
All well and good, but as Hron told this site, the legal tech and publishing giant already has around ‘190 people in TR Labs (their internal tech team), of which the legal portion is between one-third to about one-half’. Plus, globally across the entire company there are other genAI experts who are not formally inside the Labs group.
So, why the deal? Hron, who is based in Zug, Switzerland, explained that there is one overriding reason: ‘What’s different is that this is a legal LLM. There is a misconception that [general] LLMs are good because they’ve seen all the data in the world.’
And this is now a widely held view in the legal world, i.e. that at the start of the genAI era in late 2022 there was a debate about legal specific vs general, and general won. Now, companies such as Thomson Reuters are looking at things differently.
Interestingly, across the Swiss border in France, Paris-based Equall, another L-LLM development company, has released its new SaulLM-141B and 54B models, specifically designed for the legal domain (see AL story).
Hron added that he’d been very impressed with what Safe Sign had achieved so far by working primarily on legal data.
‘Their approach to model training and fine tuning is quite unique. They are very data efficient.
‘When it comes to LLMs the data is extremely important. So too the quality of that data. So, this is not a ‘kitchen sink’ approach, it’s very targeted and they can focus on where [general LLMs] don’t perform well.’
‘We can apply this [L-LLM] to our estate of content and data and then scale it.’ he added.
And TR certainly has one of the largest stores of legal data on the planet. I.e. if you have a team that has focused on legal data training for genAI outputs and they have – according to Hron – achieved very good results on just the data they had at hand as a small startup, then imagine what can be achieved when applied to the terabytes of legal data that TR holds?
Moreover, TR employs a large number of legal sector experts who can further help when it comes to the useful application of the L-LLM.
In short, this may just give TR an edge in the legal genAI war that is currently raging across the sector as companies battle for market share and to win the hearts and minds of the legal community.
And at the heart of this battle is accuracy. Completeness and relevance of answers also matters, but at the core of this is simply getting the response right. This additional L-LLM approach may help here.
‘At the end of the day this will [in combination with what they have already] be a better solution. Accuracy is one of the most important things that we think about and we want to be as accurate as we can be,’ Hron added.
So, how does it fit with what they have already? They use OpenAI, Anthropic and Gemini LLMs, for example. Aside from just using the L-LLM, or at least the expanded one they will be building based on TR’s trove of data, can it be used alongside other LLMs?
Hron noted that it is indeed possible to use multiple LLMs on the same specific task.
‘You can use use agentic flows to use many models together,’ he explained. ‘And we remain agnostic on models. We use any and all LLMs to get the best accuracy already, and this will be no different.’
Finally, how did they get together? Hron noted that TR’s Venture arm does a lot of market scanning and found out about Safe Sign, which has been keeping a low profile. They got talking about how to work together and soon enough this moved to a deal.
So, where does this go to now? Here’s a couple of thoughts:
- If an L-LLM can be trained on the incredible data store of TR, with experts in genAI and legal domain experts all helping out, then they may be able to make really significant strides in terms of outputs and overcome any nagging doubts about accuracy, especially following the now infamous Stanford study into genAI tools.
- If L-LLMs can outperform general LLMs that have seen answer refinement and system-prompting to improve legal domain results, then this suggests that others may well need to explore this approach as well. E.g. perhaps you need to go and say hello to Equall, or perhaps 273 Ventures, among others.
- Plus, that when one considers that genAI was not even really a popular subject in legal tech until late 2022, the speed at which companies are moving ahead, testing, developing, learning, and then pioneering onwards, is extraordinary.
- And also that those companies with huge data troves, e.g. TR, LexisNexis, vLex and a few others with smaller sets, may have an advantage when it comes to training up L-LLMs. And if that’s the case, then this may have a significant impact on the ongoing genAI war.
Interesting times.