YC this week put a judgment on the table that the robotics industry finds uncomfortable: MIT professor Philip Isola argued last month that robot skills, which originally required tens of thousands of hours of video training, can now be handled by general-purpose large models with just a few lines of code. This week's Decoded podcast features two startups doing exactly this—Waddle Labs and RoboCurve. Our editorial team wants to use this opportunity to break the issue down clearly.
What this is
Past robot training worked like this: let AI watch ten thousand hours of industrial robotic arm video and train it to grasp. This is the "specialized model" approach.
Isola's judgment: stop doing that. Just take general-purpose large models like GPT and Claude, write a few lines of code as the "policy," and they can control different robots. This type of model is called VLA (Vision-Language-Action—"see images, understand human language, take action" in one) in academia.
The theory supporting this judgment is called "The Bitter Lesson" (proposed by reinforcement learning pioneer Rich Sutton): general methods always eventually defeat specialized methods. This pattern has been verified in NLP (natural language processing) and code generation—now it's the robots' turn. Waddle Labs and RoboCurve proved with demos that this approach is at least viable at the prototype stage.
Industry view
The optimists (the Waddle Labs, RoboCurve types) believe this is the right path: general-purpose large models iterate too fast for specialized models to keep up. The cost of training a specialized grasping model today may be wiped out by next month's foundation model upgrade. Historically, the story of specialized AI companies being replaced by general AI has already played out in image recognition, speech, and code.
But the opposing voices are hard. People who have actually deployed robots will tell you: the factory floor's demands for latency, stability, and repeatability are entirely different from writing code; a demo that runs doesn't mean it can run for a year without errors; the "black box" decisions of general models may hit safety and regulatory red lines in industrial scenarios. A repeatedly raised question is "how does a robot learn from one success to succeed next time"—which is precisely where general models are currently weakest. We note that YC itself hasn't given a definitive answer—the last section of the podcast is titled "How far are we from general-purpose robots."
Impact on regular people
For enterprise IT: factories and logistics companies evaluating "robot deployment" projects may need to bid separately for "specialized solutions" and "general-purpose solutions," rather than betting everything on one approach.
For individual careers: the medium-term value of positions like robot debugging and robot vision algorithm work needs reassessment; conversely, engineers who can translate business requirements into prompts and rapidly prototype robots using general models will be in higher demand.
For the consumer market: the update cadence for products like home robots and robot vacuums may accelerate, but they're still far from an "all-purpose butler"—don't get carried away by demo videos.