Written by Dennis Crouch
As many of you may know, I have been engaged in machine learning research for many years, starting in the 1990s when I was a university student. However, like most people around the world, I am surprised and impressed by the power of his LLM-based GenAI technology and have found ways to use it both in my patent work and in my work as a law professor. I'm trying. We hope this will help you become more effective and efficient.
On the Patently-O side, I have made some refinement of my approach. I've tried asking GenAI tools (such as ChatGPT and Claude) to simply write blog posts for me, but the results lack insight and the tools keep fabricating wrong answers (i.e. hallucinating), so those attempts almost failed. My approach these days is to first draft a blog post myself and then refine it with insights from GenAI. I usually ask for feedback on mistakes in a post or suggestions for what else I might have said. I also use these tools to generate quick ideas when I get stuck as a writer. I write at least one blog post almost every day, but I have about 1,000 unfinished posts, many of which are sitting idle because I don't really know how to put them together. These days, when I'm stuck, I usually ask an AI tool to help me brainstorm (“He gives me 5 ideas for what to write next”). Ideas are often not very good, but sometimes they are good. And most of the time, you'll find that one of them sparks an idea for what to do next. In short, this has allowed us to get more work done, but we are still working on improving the quality.
I also use AI tools to transcribe Federal Circuit oral arguments. This is a little more complicated and requires some Python and API calls to make it work. This tool does a great job with automatic transcription, including being able to catch legal phrases without any problems. However, the current version cannot differentiate between speakers. This summer I hope to complete a project to automate serialization so that the transcripts look like real court records.
In my patent application class, students are busy drafting patent applications. This year, they are using a drafting tool known as Edge, created by his Evan Zimmerman. This tool has a particularly good chat tool for creating a detailed disclosure of your invention. I really enjoy watching my students delve into this. We envision GenAI assisting with many tasks, prior art searches (including detailed comparisons with claims), responding to office actions, and drafting specifications (including multiple predictive examples). They may even insist on drafting.
There are still many hurdles and false starts, but I continue to imagine many possibilities. We'll keep you updated on what's happening with this process over the coming months.
The most interesting recent GenAI project involved recreating some famous photos using GPT-4. I first uploaded the original photo and asked GPT to provide enough details for the artist to recreate the photo. I then fed that explanation back to her GenAI, which produced the following results: It's quite impressive.
I suspect that one reason these copies are so robust is because they are iconic photos that are likely used repeatedly in the training set. I did the same experiment with my photos and the results were not as great, but still pretty good.
We would love to hear your thoughts.
