Moontower #300

My wife Yinh pinged Khe to see if he’d host a seminar for people interested in getting more mileage out of AI. He agreed (and texted me separately with “I’m sure you know this already, but your wife is a force of nature”…Oh, I’m quite aware).

She emails 60 friends and associates, inviting them to this session. She got 40 affirmatives by the next time she checked her email. As she told me by text, “everyone feels like a boomer”.

I’m not going to spend time speculating on AI futures but it’s hard to not realize that sources of value are shifting beneath our feet. AI could have written a lot of my technical posts. It wouldn’t have been in my voice, and insofar as my written voice has some entertainment or organic attraction, it’s not the same quality. But it would convey much of the instrumental information content and at far lower cost.

To paraphrase a tweet I can’t find anymore (I gave Grok 2 chances but no dice) from a lifelong software developer: I used to pride myself on solving tricky problems and finding elegant solutions and watching LLMs do this instantly has rattled my sense of self-worth.

A lot of value-added work in the past is commoditized. Without speculating on futures, I only invite you to prompt yourself, “how much of the work I do is truly unique?”

There’s an ongoing whole layer of AI-enabled automation creep that we’ll embrace, I’m not talking about that.

Aside: The latest one I’ve appreciated is the Gemini summaries that sit on top of long email threads:

I’m talking about the increasing automation of formerly complex tasks that we actually hang our hats on. My voice is the organic part of the writing. But what’s that even worth relative to the content? A doctor’s job is a mix of inference and bedside manner. What’s the relative value of these things? If you are in sales, how much of your cycle has to do with trust and a firm handshake versus data persuasion and storytelling. What will AI lower the cost of, not just for you, but for competitors? It’s the Innovator’s Dilemma in homine.

You squeeze one part of the balloon and the air bulges into another part. It’s like the “commoditize the complement” thinking in product strategy.

This article by Jon Matzner is a good example of updating your models hard in the face of re-structured inputs:

The Only Cost I Want to Go Up

Where does the value go? What are your sunk costs? What is a trapped asset or idea that is now tenable? These types of questions were always relevant, but the speed at which they are coming at us in a wide-open frontier is only precedented at major inflections of general-purpose technology. A relatively recent example on the timeline of innovation is Netflix knowing long before it became a streaming-first company that streaming was the future, but broadband needed to catch up. The AI inflection is a confusing sandstorm rapidly covering old monuments while revealing new spaces to build. Anyone paying attention can feel it. Hence, the FOMO response to Yinh’s email.

A couple weeks ago, in work is going to feel very different by next Christmas, I said I’d share some of the things I’m doing with AI. For example:

I can write a description of a bug in Linear or Jira (who am I kidding — I upload a screenshot with a blurb and have AI write a detailed bug spec complete with testing protocol) and assign it to…”Claude bot”. A dev approves the change. Push to prod.

Tools I use

1. Claude Code with Opus 4.5 on a Max plan ($100/m)
2. Gemini Pro ($16/m)
3. ChatGPT Plus ($20/m. This is really Yinh’s but I use it sometimes. Don’t tell.)
4. NotebookLM when I want an audio summary of lengthy text. Sometimes I will give it a bunch of notes and have it generate a podcast conversation about them which can help me find connections or through-lines that outlining doesn’t. Something about hearing vs hearing I guess. I could get to the same place sparring with a text-based LLM but this way I can do it while walking or in a less taxing way. Or sometimes because I need to just let some material wash over me.
5. A service (in stealth at the moment) that connects to any data source including the moontower options database that’s hyper-tuned for inference. An example I’m messing with now is something I call VCR or Variance Contribution Ratio. What percentage of the realized variance in a list of assets in some window came from N% of days. You can ask it a question like that and it will come back with all kinds of ideas, charts, insights. It’s an intern that skips the photosynthesis→food→brain power+senses drivetrain altogether and just runs on sending electricty through a recipe trained loop that our ancestors have run through many times before. Progress. Suicide. Both. I don’t know, I’m just an ape with a lever.

Comments

• I was using Terragon as an agent that can talk to my github repos as well as deploy to vercel but Claude has replaced all this workflow. Similarly, Claude has replaced my usage of Copilot in VSCode.
• I find the native MSFT Copilot in Excel to be trash. I haven’t used Claude for Excel yet. I have found the Claude Chrome extension in the browser to be helpful with Google sheets.
• More broadly, Claude and Gemini browser tools are still clunky in some contexts (“port this substack post to my wordpress site”) and useful in others especially within your email. “Give me a list of all the people who reached out asking about XYZ” or “go find the loan agreement document with entity X and compute the accrued interest according to its provisions through today”.

Silo’d projects I’m currently using AI for

As a matter of procedure, I use Projects folders within Claude to compartmentalize initiatives. You can upload lots of files to the project for context and then have a limitless number of chats within the project itself all dealing with different aspects. I like to tell Claude to consolidate what I’ve accomplished or learned from individual sessions and then I feed that back into a living document in the files section so its always handy to the bot for future chats. These documents also include feedback about directions Claude took that I didn’t like. I don’t know what reinforcement learning is mathematically but the process feels like manual training.

Here’s a few things I’m working on:

Investment Idea Generation

Agent reads investment letter email, generate a bullish/bearish template that considers the author’s sentiment, timing, etc, which can automatically be fed another agent connected to moontower.ai infra to see how it accords with option surfaces. Another agent then presents the output. In an advanced form this is becoming part of the product itself, capable of generating watchlists and so forth. (I’m more interested in the pipeline at the moment. The business end of it would be more complex since there’s author IP involved).

Status: very early…need to get email agent running

Mooncoin Option Trading game [skunkworks name for now]

This is less agentic and more LLM conversation about gameflow. However, Claude Code was integral to helping me style.