Five Lessons from Putting AI Into Research Teams
Which tasks to automate, where analysts push back, and how to tell if a workflow is quietly getting worse - five lessons from a talk I gave last month on getting AI into research workflows.
Writing about software development, applied AI, and machine learning.
Which tasks to automate, where analysts push back, and how to tell if a workflow is quietly getting worse - five lessons from a talk I gave last month on getting AI into research workflows.
A tour of two practical Modal deployment patterns: cron jobs, secrets, volumes, static IPs, web apps, background AI jobs, and sandboxes, all in plain Python.
I joined Khe and Brett on the Invest with AI podcast to talk about how investment firms move from individual AI experimentation to a shared, firm-wide capability. These are the seven strategies from our conversation.
I used Claude Code to connect open federal energy datasets into WattsOpen, an interactive site about where the grid has room. It started as a way to test two theses about AI and public data.
A tweet about fine-tuning on frontier model outputs got me thinking about how much of that loop could be automated.
My attempt to describe what an agent is and why it's so incredible, yet simple.
Raw, unedited notes on what it feels like to build software right now.
I had my kids build their own games using Claude Code, and it got me thinking about AI, education, and equity.
People asked me to share the portfolio optimization workflow from my last post. I turned it into a Claude Code skill you can install and try with your own brokerage data.
I fed our brokerage data into Claude Code and iterated over multiple sessions to produce a portfolio optimization plan - phased actions, tax impact analysis, fund recommendations. From messy CSVs to an actionable plan we're now executing.