Notes from AI Engineer Miami
I am grateful whenever there is an opportunity to pause and reflect with other humans who are facing similar challenges, to compare notes and trade tips. My favorite conferences are small and single track. AI Engineer Miami was a very nice event, and being on the banks of the Miami river was a real treat. I especially appreciate that the organizers provide a livestream on YouTube, making the content available for everyone regardless of ability to pay for a conference. The livestream recordings are linked from the website.
This was a note I shared internally at work, and is affecting how I invest effort in keeping own skills up to date. The ability to write a basic agent loop being just as fundamental as understanding a linked list or a primary key was a sobering statement.
this was the most thought provoking talk from yesterday: https://ghuntley.com/real/ (blog form) https://www.youtube.com/live/6IxSbMhT7v4?si=oQ7-k7to4HP9HuWH&t=16576 (timestamped to start of talk from the livestream). It covers how the very best coding agents run at $10/hour, predictions for org charts being compressed so that a director level management role has individual contributors reporting to them, but those ICs are AI agent managers, and a claim that just like understanding primary keys and linked lists was table stakes for a software development job, now being able to implement an AI agent loop against an LLM is the bare minimum level of knowledge for using agents to produce software.
I also really appreciated learning about Fresh, a terminal editor that is much more approachable than vim https://getfresh.dev/. Another neat tool discovery was termDRAW, a program that allows you to draw editable diagrams, UI mocks, and text graphics without leaving the terminal. The abundance of tiny special purpose tools is delightful.
Another interesting thing from the conference was the talk analyzing different AI SDKs and libraries. If knowing how to implement an AI agent loop is so crucial, choosing a good AI SDK as a starting point seems important too. These were lumped into three generations. 1st gen SDKs were the OpenAI API and BAML. 2nd gen SDK is best represented by the Vercel AI SDK. 3rd gen SDKs were OpenCode SDK and Pi.dev. Afterwards I found another talk online that went into more detail about using Pi as an agent harness: A love letter to Pi.
Quality still matters, more than ever. A real highlight was Nnenna Ndukwe presenting on How to Embed AI Code Quality Gates in Your SDLC. The talk addressed the need to "codify development best practices into machine-readable guardrails". Figuring out how to elicit human expertise and taste into agent readable form is something I've been thinking about a lot this year.
I continued my practice of taking time away from the phone and computer to be outdoors and rinse some of the AI out of my brain. There was a Citi Bike station near the hotel and riding the bike over the Venetian Way bridges to Miami Beach to go to the grocery store was a good exercise break.
