How Generative AI Will Change the Legal Sector - with Dan Katz (Part 2)
In this second video on the subject, Generative AI + The Law - Professor and legal AI expert Dan Katz talks to Richard Tromans of Artificial Lawyer about generative AI and the legal sector. Dan will also be a keynote speaker at the Legal Innovators Conference in SF in June 7 / 8.
legalinnovator...
#legaltech #legalinnovation #ai #generativeAI #gpt
Areas covered:
Some of the issues we explore:
Privacy and data control - how do this? Much depends on risk appetite. Need a moderation layer and an audit log, can also have differential privacy - so redact aspects that have been imported - strips out identifiers. But may still need people to check this. But that is not free of issues. Fact pattern could reveal things.
ChatGPT’s release led to us all taking part in in the largest data annotation exercise in human history (i.e. ChatGPT).
We will start to see people build many more LLMs - as unit economics is dropping every day. GPU cost is dropping, and also once you make a model you can compress it, which reduces the cost.
Things get cheaper. People are so focused on these problems such as to optimise LLMs. New chips are coming too. (RT - reminds me of Moore’s Law).
Also will see what Bloomberg did - they bought own GPUs and built own propriety LLM. Bard will get better. There is DeepMind, Amazon, Anthropic, others. And Also will be domain models also.
Is this MySpace? I.e. in terms of where we are now and in fact it’s so early it’s very hard to know where things will go? Dan says: ‘A lot of what have now (re. LLMs) will be gone in two years.’
OpenAI got a jump on the market, but this is not over.
M&A (as noted previously) will grow because companies will see that those who entered market first educated the market - but then other waves will follow. We saw this with legal tech. Buyers normalise things, they get more sophisticated.
Would you agree that ChatGPT’s hallucinations made a big mess and that now everyone in the generative AI field is having to clean that up so people trust this technology more?
Dan says that GPT-4 is much better, but people need to stop thinking that ‘zero shot’ approaches will be sufficient. Need a sequential approach to using generative AI. (RT - reminds me of good old-fashioned user training of NLP.)
Can we get to zero hallucinations? Being wrong with absolute confidence you are right is very risky. Dan says: ‘There is no system with a zero-error rate. But you can reduce it, yes.’
Already been done. Can add a retrieval log to make sure.
The answer will be the growth of an LLM ecosystem of tools to help validate the outputs. There will be plugins that can help. You can connect to Wolfram, for example for math. (RT I.e. the solution will be building layers and feedback correction loops into the process, perhaps via other apps.)
Also, Dan notes: ‘The law does not have a zero error rate either.’
So, what will change most? KM will be a key area for change. Due Diligence, less so, but still opportunities to go deeper / further using this tech.
Is doc automation dead? Hard to see how doc automation can survive in the traditional way. How survive? It seems it’s existential. It doesn’t replace entire tool. But if two competitors then customer will go to one with LLM capability. Hence why there will be a lot of M&A.
‘Hard to see how the traditional way of doing doc automation will survive. …..it is existential to a lot of people (not just doc auto),’ says Dan.
As to companies adapting......first many legal tech companies have ignored LLMs, then they’ve said they can do it themselves (e.g. relatively quick and easy APIs to OpenAI), and then they will realise they will have to buy a company in the LLM space.
What happens to ALSPs? Does generative transform them or kill them? Can the ALSP model come of age by utilising LLMs?
Dan says: ALSPs have a big pile of work that is amendable to these tools. Focused on steady stream of similar work and you get economies of scale.
ALSPs used to be body shops, and similar to law firms, but have some automation. Can they make automation first. Private Equity can help here. Help with moving one model into another. ALPSs are in a nice position to get most out of these tools.
And we also discuss Dan’s new LLM-connected legal data project, that he’s doing with Michael Bommarito and others.
Conclusion: it’s another AI summer for legal - (and RT - this time it’s going to last....!)
25 окт 2024
