Might re-make the first tool I ever built with an LLM: a note cleaner-upper. I’m always taking messy notes while working, reading, talking to someone, etc. Then have to go back, structure them, and clean them up. Need a little tool to do it for me.

In cases where I can condense all the context I need my LLM to have for a given user into ~10-20 pages, I'm having trouble seeing the benefit of RAG anymore. Claude is great at pulling info from that amount of context. Yes you pay for more tokens in context, but vector DBs are

Evaluating job opportunities by what could make my LinkedIn as confusing as possible.

Crazy upside potential for people who can do mundane app development but are deep enough in AI to see novel applications.

It's hard to evaluate when an agent is factually right. But sometimes a good enough proxy is evaluating whether the tools it used made sense.

I don’t want to be a “software engineer” or a “product manager.” I want to experiment with AI’s capabilities and build useful tools with them.

Open ai engineering problem: agentic data analysis inside saas tools. On the fly reporting. A year ago I was sure every b2b saas would have an interface acting like a simple embedded Julius AI . These tools are full of customer data, but usually have limited user facing

NYC Ship

New York bodega system is unmatched.

Pro-sumer AI tools has to be most exciting product space right now.

What's the best open source model or API available for video generation?

If you're struggling to get the model to do something reliably with a prompt, stop and ask: could a random human off the street do this task given these instructions? Just caught myself with this one.

I think there’s a lot of products here. Specialized coaches for big life decisions.

High confidence take: some of the most simple tasks for LLMs (summarization, re-writing) are extremely valuable for all sorts of niches and are still under productized.

I would love an LLM that disagreed with me more.

Searching for prompt alpha

Crazy how quickly we adapt to new technology / abilities. Used to take like a week to figure out a new code base. Now if it takes longer than a day it’s frustrating.

I'm en route to v0.0.1 with hypothesis, then it'll be onto v0.0.2 with data.