Large language models are already changing legal practice, allowing lawyers to work faster and offer more to clients, but adoption so far is limited.  Use of LLM tools is much more widespread in software development (they are better at writing code than contracts) and this talk by Andrej Karpathy sheds light on some of the issues for lawyers.   Programming can be automated, but it quickly results in a mountain of code that is very difficult to understand, let alone maintain or amend.  Too much automation in legal work will cause the same problem.  Cursor (a popular code editor that incorporates LLMs) shows a better way forward for legal tech:  partial (and adjustable) autonomy, where the human remains in full control, able to audit the work of the tireless but fallible LLM and go faster.