Week notes — 15th August 2024
Cooperative AI & Agents
Last week I started thinking about how the classic HCI concept of ‘task’ applies to AI agents. The week before I looked at DeepMind’s work on Cooperative AI. I keep on trying to find a structure to shape my thinking in this area — none is forthcoming. I found Google’s announcement of a feature to automatically compile notes from meetings fascinating. I think the interesting thing here is that a strong limit on what AI might do in the workplace is how it gets instructions and understands social context. If it’s listening to all your meetings… then I guess that’s one road block out of your way? Posts I have in mind on these topics:
— A post thinking through what an AI ‘agent’ might do in the context of web development, the business context I know the best
— A post thinking through agents from the perspective of cooperation costs, that is, how AI ‘agents’ might reduce bottlenecks of information flow between people
— Something looking through ‘agent’ type products on the currently on the market
CS224N — Deep learning & Natural Language Processing course
Assignment 5 has been a killer. In contrast to previous assignments, it required me getting stronger at the fundamentals of linear algebra. If you were taking the class at Stanford, you’d have a tutor to guide you through the assignments. Instead I’ve been relying on two resources — other people’s answers that have been posted online, and ChatGPT. Other people’s answers are interesting, in part because some have clearly copied each other, and in part because some of them have the wrong answers, even those that at first glance seem to indicate the person really knows their stuff.
On the other hand, ChatGPT is just phenomenal at not only giving answers but explaining them. I don’t trust either ChatGPT or other people’s answers completely, so there still is the requirement to properly understand the material. It’s taken me quite a while, but I have finally broken the back of the theoretical part of the final assignment on the course.
I’m also starting to feel like I might ‘tune’ the personal project work to be more relevant to what I want to do — probably deploying a modern open source LLM, rather than working with BERT, which is something I’m not likely to ever need to know.
