LLM Impact Mapping: A 4-Step Thought Exercise with Claude 3.5

Using Langchain + Claude 3.5 to brainstorm about AI's effect on common roles

I'm starting to plan what to do next in the long run, and I’m entertaining the idea of joining a team. I think things might be moving too fast and too much demand for skills for risking time on solo projects.

Three types of play

Started considering which spaces are interesting now. I think this take from Hamilton Helmer is a good one: Three types of play:

1. The tech play: example of semiconductors, you get Intel and etc.

2. The first-tier play: companies that wouldn't exist before semiconductors (Microsoft)

3. The usage play: car companies with tons of semiconductors. They already existed, just got better.

The whole video is worth watching, but heres’ the time-stamp about AI.

I'm personally more interested in the second play. And finding these companies early and identifying which ones are already in a good direction is so hard.

Brainstorm exercise

A good place to start could be looking at common roles and how they will be impacted and can be augmented or even replaced.

A programmer's job is already seriously impacted, customer support even more so. Same for SDRs. Some other roles, like localization, have already been pretty much replaced. These highly impacted roles are the things I'm most interested in.

So, just as a thought exercise toying with Python + Langchain, I wrote a 4-step chain with Claude 3.5 to do following:

1. Brainstorm departments and roles common across any vertical

2. List the most common tasks in those roles

3. Brainstorm ideas for how each could be augmented or automated

4. (Rank them by difficulty and confidence, but this was just for nudging the model)

I only spent a couple of hours and $3 in credits, but the results turned out better than I was expecting.

There's obviously some nonsense and some things that are regular ML models more than LLMs in the output (my fault, I should have optimized the last step of the chain better), but a lot of solid ones that I think haven't happened yet, and some others in the “augmentation” category fully in progress like Figma, Intercom, Cursor, GitHub Copilot, etc. If you squint and look at the roles with the best ideas, you can see Piper, Fin, Harvey, Wisq, Devin, 11x.ai, etc

You can see the whole list here:

https://docs.google.com/spreadsheets/d/1vpXzfUFNhIZJp9CEmiZqGKaVrP_Xx8McMZZlLCxiHNg/edit?gid=0#gid=0

Next step, I guess, is to start collecting companies and maybe have a little agent help me with the research.

—

I haven't had time to think through the whole list yet, but here...Some of the ones that stood out that the ideas are mostly very doable, or already happening: