These are the "good and decent people" who didn't survive the Reagan presidency.

Perhaps. But I also suspect that LLMs are better on coding tasks because underlying concepts and logic are more tightly coupled to programming languages than to natural language, & programming languages have smaller vocabularies and stricter syntax. That gives language models much better footholds.

We’re seeing some skillfulness in coding assistants, which are remarkable better now than they were ~9 months ago. But I strongly suspect that the substantive differences between computer science and, say, legal or medical practice will mean that we won’t be seeing skills develop at the same level.

I think the distinction there is that a “mostly close” explanation and some case studies may be enough to get another human brain to understand the pattern, but LLMs don’t learn the same way we do.

I suspect that those kinds of advances will only come with something other than transformer models.

Yep, it’s analogous to the shift from expert systems to machine learning. If you’ve got enough data, let the algorithm teach itself.

Which has worked fine for some tasks (e.g., classification), but I haven’t yet seen any evidence that transformer models are learning professional practice skills.

The classic characteristic of implicit learning is that the person can’t actually articulate WHY something looks right or wrong to them — it just does.

If they can’t explain it, it’s really hard to develop an algorithm or QC check for it.

The most efficient/effective use of the LLMs in the near term may be to augment the human practitioners by providing context, cites to the record, flagging unusual language, etc. rather than directly assisting with drafting.

Which may end up running into one of the same problems that plagued expert systems a few decades back: implicit learning plays a role in professional practice. Capturing enough domain expertise to make those bespoke checks work may be extremely difficult.

It generally isn’t. Though it may look presentable at first glance, which just makes it more dangerous.

And it’s especially pernicious since the errors tend to be subtle and require close reading. The fundamental job of an LLM is to create plausible-looking text, so while they don’t often suggest rocks as a pizza topping any more, they’ll get the legal drafting wrong in all kinds of non-obvious ways.

Exactly. It takes longer to clean up an LLM draft than it does for a skilled practitioner to draft it from scratch. Similar to the old story with machine translation (cleaning up a machine draft took longer than a skilled translator starting from scratch).

“call your congresspeople and beg them not to sentence trans people to death” is a regular feature of our politics now and it fucking sucks

I have to constantly remind people that it is a large LANGUAGE model, not a large legal model or large abstract reasoning model. It’s well suited to several information retrieval tasks, but not for drafting contracts and briefs for you.

I, too, work on AI in legal, and am a recovering litigator. There are still quite a few areas where I will not build a system that lets LLMs generate anything close to work product, or to provide summaries for legal practitioners.

Better to have it provide direct quotes from docs and pin cites.

If you think the system doesn't seem interested in rewarding white guys at all that seems to be an issue of your media diet and not a reflection of reality

the full speech is something to see & is one of the reasons it is good to have a trans member of congress

youtu.be/5LqVMCW4e0c?...