2026-09-04 –, Auditorium
Most of us now talk to our computers like we did in the 60s, through a terminal. The difference is what is on the other side. Along the way, we stopped writing code by hand.
This talk is about what that costs and what it buys. Treating AI systems with awe hides what they can't do and hands us back plausible slop. Named precisely, as a tool that drafts, reviews, searches, or automates a specific chore, it becomes something that can be tested, questioned, measured, and optimised.
Most of us have adopted AI tools without the surrounding guidance on how to coerce models into making something good and useful. The speed of development grew exponentially, to the point where we now interact with computers like we did in the 1960s, through a terminal. This talk is about how to concretely take your everyday prompt and make it better, from slop to a workflow that can be measured and improved far beyond your local environment. We'll cover gold standard practices such as context, guardrails, optimisation, and evaluation frameworks, the things that let you take advantage of AI tooling to the maximum, rather than being tied to reviewing every line it produces, increasing both speed and trustworthiness.
AI, agent optimisation
Laís Carvalho works on growth and developer content for Logfire, the observability platform from the team behind Pydantic, writing and building across the Pydantic AI ecosystem. In March, she organised and moderated a panel on open source in the age of AI with Guido van Rossum.
On the community side, she's a PSF Fellow, a EuroPython Society Fellow, and a 2024 PSF Community Service Award recipient, with a long history organising Python community events. She also builds with the same tools she writes about, which is the only reason she's qualified to tell you most of your AI output is slop. She'd know. She's produced plenty of it.