We are excited to announce the first night of talks in the NYC Systems series in 2026! Talks are agnostic of language, framework, operating system, etc. And they are focused on engineering challenges, not product pitches.
We are pleased to have Sean Allen and Yining Shi speak, and glad to have Trail of Bits as a partner for the venue.
Sean T. Allen is a founding member of the Pony core team and currently works at Antithesis. His turn-ons include programming languages, distributed computing, Hiwatt amplifiers, and Fender Telecasters. His turn-offs include mayonnaise and stirring yogurt.
Most engineers think garbage collection is slow. They're not wrong about their experience — stop-the-world collectors really do destroy your tail latencies. But "GC is slow" confuses the symptom with the cause. The real villain is coordination. Every stop-the-world pause, every atomic reference count, every lock-free reclamation scheme is coordination over memory, and coordination is what kills performance.
Pony takes a different path. Actors own their own heaps. Each actor collects independently, using only local knowledge. No stop-the-world. Ever. When actors need to share information about cross-actor references, they gossip — piggybacking on the application messages they're already sending. They share information but never wait for agreement to act.
This isn't magic and it isn't free. But the costs are local, not coordination costs. And that changes everything about what "fast GC" can mean.
Yining Shi is Senior Director of Applications and a founding engineer at Runway, where she has spent nearly seven years building the products and infrastructure behind AI-powered generative video. She is also an adjunct professor at NYU Tisch, where she teaches Machine Learning for the Web, and the author of Make: Jumpstarting the Arduino 101 (O'Reilly).
This talk covers how Runway engineered the infrastructure required to run generative video models in real time — from model optimization techniques to the serving architecture that makes mutli-step low-latency inference possible. Yining will walk through the specific constraints of real-time avatar generation and the tradeoffs the team navigated, as well as show a live demo of Runway Characters.