Summary

West sees signs that robotic manipulation is beginning to scale, with laundry folding moving “from being impossible to being boring over the last year.” Teleoperated demonstrations made imitation learning work in more serious ways about two or three years ago; more recently, combining it with reinforcement learning has produced substantially more robust manipulation. Transformers mattered, but so did the post-ChatGPT conviction that scalable learning could attract more capital, data collection, compute, and cheaper hardware.
The investable reality is narrower than the demos but stronger than the consumer market suggests. Biewald pressed West on the gap between fluid videos and underwhelming customer systems; West’s answer was that practical vendors already have “tens to 100 robots” performing learning-based pick-and-place in manufacturing, although deployment has not reached massive scale. The near-term product path may favor scrappy, application-specific operators rather than companies “swinging for the fences.”
Consumer robotics remains a reliability and productization problem, not merely a task-performance problem. A robot may fold laundry under controlled conditions yet fail commercially because homes contain a “huge long tail” of variations, while installation, onboarding, servicing, and exception handling all matter. West expects incremental products first, such as vacuums that eventually also pick up toys, because enough small failures make owners stop using the machine.
Physical AI requires a distinct data architecture because its data is “multimodal, multirate and often episodic.” Images, motion, joint state, LiDAR, text, and model outputs arrive at different rates, so West argues they cannot be forced cleanly into a conventional table. Today, answering a question such as “nighttime runs where the left gripper failed to open and failure type B occurred” often requires custom parallel code; it should resemble a SQL query.
Rerun’s commercial thesis is that visualization must be ubiquitous and therefore cannot be the primary tollbooth. Engineers need the same viewer in notebooks, beside a robot, inside training-data pipelines, and during operations; per-seat pricing breaks when occasional support users also need access. Rerun consequently open-sourced its client-side viewer and logging stack, while monetizing the scalable cloud backend and the record-curate-train data path.
That strategy rests on unusually deep infrastructure rather than a thin visualization layer. Rerun built an in-memory database, a Rust-and-Arrow stack, and a sparse-Parquet-like physical-data format; it also redesigned its data model roughly four times. The governing API principle is “dump and forget,” paired with an entity-component model flexible enough to attach arbitrary confidence values, point clouds, or neural-network outputs without predeclaring every schema.
Robotics still lacks shared evaluation infrastructure, and simulation remains an imperfect substitute. West said there are “no great benchmarks”: useful systems co-train and co-design models with hardware, while simulation-based comparisons inherit the simulator’s weaknesses in contact and manipulation. ROS persists despite widespread frustration because it is “proof that network effects work,” with standard messages and reusable modules outweighing architectural dissatisfaction.
A bubble verdict depends on whether general-purpose platforms can turn capital into hardware, data, and compute scale. West would not dismiss pre-revenue valuations because serving many tasks on the same hardware could lower hardware cost, but doing so demands enormous data and compute. His nearer-term forecast is better robustness—longer tasks, self-correction, and some learning on the fly—while strong high-level reasoning may take the following year.

Deep dive

Not yet available upstream; scheduled sync will retry.

The Fairfax Way with David Thomas $FFH.TO Finding The 1% of Stocks That Matter | Henry Ellenbogen Interview