Radical Ventures partner Rob Toews on pre-training plateaus, robotics data gaps, and Waymo's civilizational implications

Summary

Rob Toews, partner at Radical Ventures and a longtime AI columnist, covers three distinct threads: where the pre-training scaling narrative stands, the data problem holding back robotics, and what Waymo's traction means for cities and investors.

Pre-training plateau

Toews is direct that pre-training scaling laws have plateaued. The GPT-2 to GPT-4 era of reliable capability gains from raw compute and data is over, and reasoning models — OpenAI's O-series, DeepSeek R1 — represent a real but still-unproven next runway. Whether inference-time compute unlocks years of further scaling or merely extracts more value from a ceiling that's already been hit is, in his view, genuinely unresolved.

What he is confident about: even if model capabilities were frozen today, there are trillions of dollars of economic value left to extract through productization. OpenAI and Anthropic are already moving further up the stack, and that's where he expects the action to concentrate.

On the data center buildout question — whether the shift toward inference reduces the case for massive superclusters — Toews broadly agrees that inference can be served from more distributed, less tightly interconnected infrastructure. He acknowledges the Jevons paradox counterargument: cheaper compute drives more usage, which could still justify frontier-scale pre-training runs. His read is that both things can be true simultaneously, and that future large training runs may increasingly happen in modalities other than text — robotics and biology being the clearest candidates.

Robotics data gap

Toews made a 2025 prediction that scaling laws would begin emerging in non-text modalities, and he sees early signs of that in robotics. But the fundamental constraint is data volume. There is no internet-scale corpus for robot training, which limits how large a robotics foundation model can practically get relative to a general-purpose language model.

On simulation as a substitute, he lands cautiously on the skeptical side — for now. The sim-to-real gap remains meaningful, and he believes real-world data is still essential rather than optional. His expectation is that simulation improves over time and the field ends up resembling autonomous vehicles, where real-world collection remains the foundation but simulation becomes a significant complement rather than a replacement.

Waymo versus Tesla

Toews worked in autonomous vehicles before moving into venture, and he frames the Waymo-Tesla debate around data quality rather than data volume. Tesla has far more raw miles from its consumer fleet, but without LiDAR the data is less multimodal and therefore less robust for training toward full autonomy. Waymo can absorb LiDAR's cost because its robotaxi model doesn't require selling hardware to individual consumers; Tesla, selling cars at consumer price points, structurally cannot.

LiDAR costs have fallen sharply — from roughly $64,000 per unit on early Google self-driving vehicles to a few thousand dollars today — and solid-state prototypes are being discussed at around $250 per unit, though those are not production-ready. Toews allows that the debate may eventually resolve itself simply because LiDAR gets cheap enough that Tesla could adopt it, noting there are already rumors of Tesla experimenting with the sensor.

Waymo's civilizational footprint

Waymo monthly rides have moved from gradual growth to a sharp inflection, and Toews says most people he knows in San Francisco now use Waymo more than Uber or Lyft. The service has expanded beyond the Bay Area to LA, Phoenix, and Austin.

The broader thesis he endorses: roughly a third of real estate in the average American city is currently devoted to parking, and autonomous vehicles could free most of that for housing or public space. Cities could be redesigned around people rather than cars. Longer commutes become productive time, enabling exurban living patterns that aren't viable today. The caveat is timing — the built environment adjusts slowly, so the full impact plays out over decades rather than years.

On downstream investment opportunities, Toews points most naturally to real estate as the clearest near-term surface area, though he notes that parking lots, while increasingly redundant, aren't cheap to redevelop.