Data Outpost is the next evolution of Small Data SF: still human, still builder-led, but sharper and more future-facing for the moment when AI starts depending on the data systems underneath it.
The room is for founders, engineers, product leaders, researchers, operators, and data architects making machine intelligence useful inside actual companies.
Julia Silge, Senior Engineering Manager, will present on “The SQL Agent Harness We Didn’t Mean to Build.”
Read the abstract here.
SQL is one of the oldest languages still in daily use for data work, and it is now one of the first things any new AI data product promises to write for you. The promise is often natural language to SQL, and it rests on the appealing assumption that the model is what will make the query correct. In practice, what comes back depends far less on the model itself than on the harness around it, meaning what the model can see, verify, and run.
Over the past year, my team has been working on SQL tools for entirely human reasons, so that a person can browse a schema, get appropriate completions, run a query in an interactive console, and inspect what comes back in a data explorer. Pointing an AI assistant at that same environment turned up something we didn’t plan; what we built was already the harness.
In this talk, hear about what a good harness for SQL authoring actually requires, why the investments that make SQL comfortable for people turn out to be the same ones that make it tractable for an assistant, and which questions have turned out to be genuinely hard. How much capability can an assistant be handed before one more tool makes it worse at the ones it already has? We’ll explore these points through the SQL and data connections work in Positron, a next generation IDE for data science in Python, R, and SQL from the team behind RStudio, and consider why adding an AI mode is so often the wrong shape for this kind of work.
