Brainstorming Session

Summary of Brainstorming Session

January 20, 2024

Introduction

Our team recently had an ideas brainstorming session where we talked about initial ideas. The raw document with our ideas can be found here, but this document will serve to provide a bit more understanding/segmentation of the ideas neatly.

Approach

Outside of the usual team bonding/warm-up activities (i.e., O'Brien did a great job of running 20 questions), we focused a lot on thinking about problems as opposed to solutions. One thing that we feel is a big trap is the notion of thinking about emerging technologies and trying to apply those to problems, rather than working backwards from what irritated us. Furthermore, we tried to be as diverse as possible during idea generation. As the document and this page will detail, naturally, a lot of the ideas do relate back to our shared experiences in ML and Systems. Nonetheless, we tried to think about all aspects of our life, from our interests in sports to daily events.

Themes

Below is a non-exhaustive list of themes that we came up, in addition to some added notes about elements such as the size of the opportunity, founder-market fit, etc.

AI Agents

LLMs have been all the rage in the last couple of months, and there has been extensive effort in building a lot of tooling for them (i.e., vector databases, observability tools, prompt management) as well as large-scale research teams building them from the ground up (i.e., OpenAI, Anthropic, Mistral). The natural progression of a lot of this is shifting from language to multimodality, as well as allowing the model to actually interact and perform actions in the environment. These are AI agents.

There's a lot of working being done in this field, from companies striving to build these agents (i.e., Adept, MultiOn) to academic contributions on how to make these agents emulate humans (i.e., Generative Agents, which O'Brien was a co-author of). While the technical challenge might be difficult, we think there may be some opportunities to build some specialized agents that can be adapted for a specific workflow:

• AI law assistant –- find law cases? Patents? See CaseText.
• AI Teachers -- AI music teachers, video game teachers, Leetcode, etc.
• ML model showing you how to do tricks in video games -- “Video game tutor” – data from good players, no marketplace, aligns with v-tuber/terminally online thesis
• Gen agents for A/B testing -- Programmatic A/B testing – showcase ideas to try out based on user data
• AI-powered gardening helper
• Terms and Conditions Summarizer