Why general-purpose LLMs may win at robotics — Waddle Labs and RoboCurve

Key points

Key takeaways from YC Decoded interview with Waddle Labs Co-founders Hanming Ye and Yiding Song (Vincent), and RoboCurve Co-founder Jay Chooi (September 2026):

Robot-use agents. Both startups work at making robots more capable with LLMs: Waddle Labs builds LLMs that control robots via a harness plus data collection and training, while RoboCurve is an evaluation company for physical AI, measuring any robot and any model, including vision-language-action (VLA) and world action models.

From VLA to code. Google DeepMind's RT2 (2023) used a fine-tuned language model that output end-effector poses instead of text; the speakers contrast that with today's coding agents, which write code directly and can allocate more compute for complex tasks via chain-of-thought style reasoning.

Code-as-policy lineage. Early code-as-policy work, including Google DeepMind's 2022 papers and Voyager for Minecraft, let a coding agent write Python-like functions to control a robot one-shot, without needing extra robot data, because the model already knew code.

Bitter lesson and data. The founding view is that the bit-per-data lesson favours general-purpose models trained on computer-use data, coding data and egocentric video — not robotics-specific models dependent only on robot data.

In-context learning limits. Hanming Ye described experiments showing in-context learning improves on a held-out task cheaply without SGD, but is non-monotonic, saturates after roughly 20-40 examples and stops improving once the context window is exceeded.

Consolidating skills. Waddle Labs compresses in-context learning into reusable programs and skills, a form of distillation (like Dream Coder's refactoring and sleep phases), letting smaller, faster models run routine tasks while a large model handles variation.

Live demo constraints. Direct model control is often bottlenecked by latency; frontier-lab LLM output-token speed is improving about 2x per month, and the speakers expect that trend could reach real-time robot control by the end of 2026.

Platonic representation hypothesis. Citing Philip Isola's essay and the Platonic Representation Hypothesis, the founders argue strong language models and strong robotics models converge on similar world representations, so a single strong general-purpose model may outcompete purpose-built robotics models.

General-purpose robots. The guests predict general-purpose robots — capable of any natural-language instruction a competent teenager could do with their hands — within roughly two years, a shift they say society is largely unprepared for.

Read more: YC Decoded — YouTube