Sam Skjonsberg

Mostly code. Always coffee. Sometimes bikes. Views are my own.

[Sam Skjonsberg @codeviking]

@jacobcares @finbarrtimbers Could call it Sleaker <3

[Cristóbal Valenzuela @c_valenzuelab]

https://t.co/zhh4IOU1mv

[Anastasis Germanidis @agermanidis]

Real-World Superintelligence

I wrote short essay on how to build AI systems that can take the right actions in an unpredictable world.

https://t.co/XXRL56LG8V

[Sam Skjonsberg @codeviking]

Runway is hiring a Founding Staff Analytics Engineer. You’ll take the lead in building AI first, self-serve tools that power decisions across the organization. This is a zero-to-one role; lots of potential for exploratory, greenfield development with a world class team.

If you’re interested or know someone who would be a good fit, send them my way.

Join us: https://t.co/bTG3mKuvFX

[Katie Lewis Grumman @KatieLewisMIT]

Excited to demo Runway Characters at SIGGRAPH with @yining_shi !

[Kamil Sindi @kamilsindi]

We open-sourced AVTensor, the Rust media decoder behind Runway's training pipelines. It decodes video and audio on one timeline straight into PyTorch tensors, with decode-time resize up to 6× faster than PyTorch's torchcodec. It improved our training MFU by 1.8pp. Check it out!
https://t.co/ANrfmvjwjU

[Kamil Sindi @kamilsindi]

As a follow-up to our Kueue post on sharing idle research compute, we wrote up how we reallocate GPUs from production to research overnight, using queueing theory to know just how many we can spare. More research FLOPs and shorter queue waits: https://t.co/lI1IZBaauQ

[Cristóbal Valenzuela @c_valenzuelab]

A very rare deep dive into Runway's culture, technology and mission, or how to build new worlds.

[Anastasis Germanidis @agermanidis]

Less than a year ago, we introduced Aleph 1.0, with the thesis that video models can become as general as language models. It was the first model of its kind, moving beyond rigid tasks like text-to-video or image-to-video, to accept combinations of image, video, and text inputs and generalize to tasks unseen during training.

Today, we're releasing Aleph 2.0, with the goal of making video editing models as powerful as possible for real-world use cases. It generates up to 30 seconds of video at 1080p and propagates edits consistently across shots. It preserves the details of the original video extremely well.

Based on all the learnings from 1.0, we've gotten more opinionated about what the right interface for a video editing model looks like. You can now preview edits on a single frame before generating the entire video, which makes for a much more interactive and controllable editing experience.

Hope you have as much fun with it as we've had these past weeks.

[Sam Skjonsberg @codeviking]

This is an amazing time for hackers, builders and dreamers. The magic of software is still very much alive and well.

[Sam Skjonsberg @codeviking]

When I first started writing code it was magical. A UNIX server in my friend's basement (his internet was faster), PHP and some HTML and JavaScript unlocked a whole new world of creation and agency.

[Sam Skjonsberg @codeviking]

Today instead of writing a custom blog, a high school kid might write their own hypervisor. Host a cloud on their friend's UNIX server. Write their own programming language. They still won't be perfect. But the floor has risen from a custom website to a distributed system.

[Runway @runwayml]

Meet Runway Agent. Your new AI creative partner that helps you ideate and execute fully finished, sound designed and edited videos. All with just a simple conversation. From ads to shorts to content for social, Runway Agent makes it easy to make more of what you need.

Get started on web at the link below.

[Anastasis Germanidis @agermanidis]

ICYMI: Over the past two weeks, we've published a series of in-depth looks into how we deploy frontier video and world models into production, across engineering, research, and trust and safety.

From how we optimized GWM-1 to serve Characters (which I believe is one of the largest deployments of real-time autoregressive diffusion today), to reducing cold start times by 60x for our inference workers via GPU-to-GPU transfer of model weights, to open-sourcing our configuration system for training and evaluation configs, and finally, the continuous work we do to ensure child safety across our platforms.

There are a lot of novel challenges we face as Runway increasingly becomes the world's infrastructure for generating pixels, media, and worlds, and we'll start sharing more of the behind-the-scenes in the weeks and months ahead.

[Kamil Sindi @kamilsindi]

We cut model cold-start times from minutes to seconds. 60x faster, saving thousands of GPU-minutes and hundreds of terabytes of transfers every day.

It turns out GPUs already holding weights are faster weight servers than cloud storage. The writeup covers how we built around that, including the failure modes we hit along the way.

https://t.co/BLUu5vf3n8

[Kamil Sindi @kamilsindi]

I get asked a lot about how we manage research GPU clusters at Runway, so we wrote it up.

We go deep on the design, tradeoffs, and what it actually takes to keep GPUs fully utilized.
https://t.co/JMWsEwu9zI

[Sam Skjonsberg @codeviking]

I don't agree w/ the take that LLMs aren't useful for large, "messy" codebases. They're amazing for exploration, quickly adding test coverage, or wiring up simplified repros. All of which are major multipliers for changing muddy/complex systems.