Swyx and Louis Knight-Webb @ AI Engineer Europe: DevTools, code mode, and the future of AI engineering
Swyx and Louis Knight-Webb discuss how AI Engineer became an industry venue, which AI DevTools retain defensibility, why PLG still wins enterprise demand, and how human writing and live demos preserve signal in an age of generation.
Swyx is the founder of AI Engineer, and Louis Knight-Webb organized the London AI Tinkerers community. They joined Jack Bridger live at AI Engineer Europe in London.
Their conversation ranges from global conference design and code mode to squeezed infrastructure categories, human provenance in writing, model interpretability and the difference between a polished event and one alive enough to surprise its audience.
The Q&A below is experimental — extracted from the episode transcript with AI. If you spot a mistake, let us know at jack@scalingdevtools.com.
How does AI Engineer expand into other countries?
The model is closer to JSConf than a formal franchise. Experienced organizers who have attended an AI Engineer event receive brand assets, marketing channels and support with speakers and sponsors, then run the local event. That creates brand risk, so partner selection depends on trust and operating experience. The network was expanding through Melbourne, Singapore, Miami, Paris and a planned Shanghai event.
How is the AI Engineer community changing?
The job titles and company sizes evolve as AI moves into production, but Swyx cares more that attendees remain open-minded and near the edge of practice. Large-company teams can behave like startups internally. AI engineering stays at the core while each World’s Fair adds adjacent roles—first leaders, then product and design, and now researchers—as the work itself broadens.
Why does industry AI need its own research venue?
Traditional academic formats center papers, yet much frontier work is now closed, embedded in products or credited to enormous model teams. Engineers still need to exchange practical capabilities research and understand training decisions that affect inference. Swyx wants the conference to align reputation, recruiting and community contribution much as GDC gets game teams to explain techniques they otherwise spend the year building privately.
Why does Swyx think code mode will endure?
Letting a model write and execute code gives users a general mechanism for transforming data, checking schedules and producing results that predefined tools cannot anticipate. Swyx used it across the event website and his own speaker planning. Fast-moving product announcements expire quickly; code execution is a more durable principle, and he selects talks for those lasting ideas rather than only what changed that week.
Where is there still room for AI DevTools?
Swyx sees strong demand around sandboxes, reinforcement-learning environments, model labs, gateways and infrastructure that serves a genuinely difficult narrow function. Generic AI frameworks are squeezed because integrations and orchestration can increasingly be generated. At the other end, vertical agents can own a durable relationship by becoming the outsourced AI team for lawyers, doctors or another specific profession.
Should an AI startup begin with enterprise sales or PLG?
Enterprise supplies margin, but Swyx sees an asymmetry: bottom-up product adoption can pull a company into enterprise accounts, whereas an enterprise motion does not automatically create individual use. Developers inside banks and other buyers still ask for the products they already know. That makes product-led growth and visible distribution relevant even when the eventual contract depends on retention, security and procurement features.
How can writing prove that a human is behind it?
Swyx suggests accumulating small signals: selective rule-breaking, imperfections, unexpected but precise vocabulary, humor and awareness of live culture. These were qualities of strong communication before LLMs, not gimmicks invented to defeat them. Provenance may change how a reader feels, but he also resists absolutism: if generated material was genuinely useful, judge the outcome rather than automatically declaring the experience worthless.
How do you stop technical demos becoming too polished?
Match risk to the venue: workshops can be unscripted, while a keynote announcement may be prepared and still happen live. Swyx’s most engaging talk built a small React clone from an empty file, nearly failed and was rescued by an audience member who spotted the bug. Production quality matters, but visible possibility of failure creates attention and authenticity that a fully prerecorded performance cannot.
What should the next conference format optimize?
Swyx is skeptical of conventional hackathons because contestants can cheat and arbitrary time limits obscure how long a product really took. He is considering pitch formats and more deliberate peer learning instead. Networking is also a product problem: apps rarely solve it, but themed discussions and lightly structured groups can create both relevant meetings and the serendipity of encountering somebody you did not know you needed.