Most teams reach for foundation models by default—even when they’re energetically wasteful and slow. This talk shows how Small Language Models (SLMs) deliver faster response times, offline capability, local-first privacy, and dramatically lower costs using 10–30× fewer resources. You’ll see side-by-side demos of identical tasks run on SLMs vs foundation models, with real latency, resource, and environmental metrics. We’ll walk through practical setups using Chrome’s built-in AI APIs and local model deployment strategies. You’ll leave with a model-selection framework and concrete techniques to implement efficient AI systems that improve performance, user experience, and budget.

Deploying AI in healthcare demands reliability, alignment, and rigorous monitoring. This talk shows an end-to-end approach we use with Phoenix: build and maintain ground truth datasets from production traces, design evaluation rubrics, ensure and measure human alignment, and add observability that actually catches failures. You’ll see where vibe-checking is unavoidable and when an A/B test in production is the right call. We’ll demo a feedback flywheel from development to deployment and continuous improvement. You’ll leave with a practical blueprint for evaluation-driven development of LLMs in healthcare.

With AI speeding up development, security needs to keep pace. This talk introduces the Hackian — an autonomous agent that performs penetration testing without direct code access. You’ll see how it identifies vulnerabilities, probes systems, and reports issues before attackers do. The session demonstrates practical use cases for integrating AI-driven red-teaming into your workflow.

Foundational models are reshaping robotics—but accessibility and cost still limit who can build with them. This talk introduces SmolVLA, a lightweight vision-language-action model for affordable robots, built through open collaboration on Hugging Face’s LeRobot platform. You’ll see how to record and share real-world data, expand the community-driven dataset, and train on it using SmolVLA’s modular architecture. Live demos show how lightweight design enables fast iteration and low-cost training. You’ll leave with practical tools and a roadmap for building open, collaborative foundation models for robotics.

AI is entering its industrial age. Each year brings larger compute clusters, models, and experiments—but scaling progress depends as much on engineering as on research. This talk shows how poolside built its automated Model Factory platform to industrialize model training. You’ll see how world-class systems programming, orchestration, and modular component design turn one-off experiments into scalable pipelines. We’ll unpack real examples of automation and integration across complex systems. You’ll leave with a playbook for building AI infrastructure that compounds productivity and accelerates discovery.

End-to-end learned Physical AI unlocks new robot capabilities—and raises reliability and safety challenges. This talk demos open-source libraries in interpretable Physical AI for continuous and automated safety monitoring of robots. You’ll see human-in-the-loop corrections in action: a remote operator inspects behaviour, injects corrections, and verifies recovery, all inside an auditable control loop. We’ll cover failure detection, telemetry, and policies that keep autonomy responsive without losing oversight. You’ll leave with practical patterns to combine automated monitoring with efficient remote correction.

Most home robotics miss real user pain. This talk explores what changes when you build around a concrete problem—cooking sucks for many busy professionals—and turn a prototype into something people can rely on. You’ll see what we learned testing a cooking robot across different home environments, the feedback loops we built with early users, and the systems we shipped, from ingredient recognition to precision cooking mechanisms. We’ll connect technical choices to reliability and daily use. You’ll leave with lessons for making home robots indispensable by solving real problems, not chasing novelty.

AI agents are deep in the hype cycle, but real deployments lag behind the promise. This talk cuts through the noise with a hands-on walkthrough of how to agentify any business in a concrete way you can do tomorrow. You’ll leave ready to deploy agents that move from demo to production, fast.

The web was built for humans to click, not for agents to act. This talk explores what changes when AI agents need programmatic access instead of visual interfaces. You’ll see examples of how today’s smart models stall on HTML-driven workflows and what new infrastructure is needed to let agents operate the web directly. We’ll outline emerging standards, APIs, and service layers that redefine how the internet exposes functionality. You’ll leave with a clear view of why the path to AGI runs through building the infrastructure layer, not just smarter models.

In a world of parity features and decent UI/UX, products win by sparking emotion. This talk shows how to layer whimsy—intentional playful touches—without turning your product into a toy. You’ll see demos of micro interactions, playful empty states, tasteful animation, and “mascot” moments, plus how generative media (with fal) makes this fast to prototype and cheap to ship. We’ll map where whimsy helps (retention, completion, shareability) and where it hurts. You’ll leave with a checklist and principles to add delight that drives measurable outcomes.

Most AI still ships on vibes—it works until it doesn’t. This talk shows how to make evaluation part of the creative loop, not an afterthought. You’ll see examples from GitHub Copilot and modern enterprise stacks where better models alone couldn’t fix reliability. We’ll demo how context-aware metrics, developer-friendly ergonomics, and lightweight eval loops help teams decide when to ship and when to measure. You’ll leave with a practical framework for balancing speed and rigor—turning evaluation into a driver of faster learning, safer releases, and better user experiences.

Most AI ships as a side chatbox. This talk shows what happens when an agent lives inside Linear. You’ll see it summarize issues, apply labels, propose subtasks, surface duplicates, connect related work, post daily digests into Slack, and nudge teams when high-priority bugs sit unassigned. It operates through narrow, auditable functions like “set labels,” “create subtask,” and “post comment,” with clear guardrails. You’ll leave with patterns for embedding agents directly in workflows so teams move faster and make better decisions.

Join Steve Ruiz, founder of tldraw, for a tour of the latest AI experiments built with tldraw's infinite canvas SDK. You’ll see early prototypes and live demos that combine freeform drawing, agents, and collaboration. You’ll leave with concrete ideas for using the canvas as a platform for mixed collaboration and multi-agent coordination.

Slides off. Editor on. This talk is a live tour of prototypes and upcoming Cursor features that reimagine the experience of writing code with the assistance of AI and LLMs. You’ll see working demos that reshape how we write, review, and iterate, along with the reasoning behind these explorations. We’ll discuss where current IDE UI falls short and what changes when AI participates in the core loop. You’ll leave with concrete ideas for building AI-first flows in the tools developers use every day.

After an initial burst of AI-tool adoption, teams now juggle overlapping solutions and growing governance challenges. This talk reviews what surveys reveal about adoption trends, how teams are responding, and why governance must move faster than adoption. You’ll see common pain points, emerging best practices, and a pragmatic view of what sustainable AI adoption in engineering looks like.