Robotics' End Game: Nvidia's Jim Fan Declares VLA Dead, Ushers in WAM机器人的终局：英伟达 Jim Fan 宣告 VLA 时代结束，WAM 登场

Jim Fan declares VLA (Vision-Language-Action) — the architecture he championed until recently — obsolete, because its parameters are dominated by language while action lags behind; the new paradigm, World Action Models (WAM), swaps in video world models as the pretraining substrate.

He calls this shift 'the Great Parallel' — replaying the LLM playbook of pretrain→align→RL, replacing teleoperation data with egocentric human video. The flagship model DreamZero is at a 'GPT-2 stage' — directionally right but not yet reliable.

The talk's narrative structure is itself part of the argument: Jim opens with signing Nvidia's first DGX-1 at OpenAI in 2016, casting himself as a firsthand witness to deep learning's paradigm shifts, then quotes Ilya Sutskever to imply this pivot belongs to the same category of event. This 'I've seen one paradigm shift, now I'm seeing another' framing is doing real persuasive work for an architecture (WAM) not yet validated in production — worth separating from the technical argument that follows.

The critique of VLA is structurally diagnostic: vision, language, and action share one transformer backbone, but language corpora vastly outweigh action data in scale and maturity, so parameters and gradients naturally skew toward language. This isn't a tuning flaw in one model but an inherited weakness across the whole family (GR00T, π0). The RT-2 soda-can-near-Taylor-Swift example is persuasive precisely because it separates cognitive generalization (recognizing an unseen noun) from motor generalization (executing an unseen action policy) — the latter being robotics' real bottleneck.

WAM's technical core merges video generation and action prediction into one decoding objective: DreamZero imagines several seconds of future frames before acting, effectively wiring the world model's imagination directly into the policy output. But the deployment cost disclosed in the paper is steep — the 14B model needs 38x system-level optimization plus GB200 hardware just to hit 7Hz closed-loop control, meaning the current version is far from practical deployment, something Jim himself likens to a 'GPT-2 stage.'

The data-side pivot deserves scrutiny too: teleoperation is capped by robot hardware uptime and failure rates and can't scale. EgoScale's exoskeleton-and-egocentric-video approach removes the robot from data collection entirely, theoretically riding the scale effects of consumer video, à la FSD's data flywheel. The R²=0.998 scaling law is the talk's strongest empirical evidence, but it only establishes a correlation between more data and lower validation loss — not that downstream task success rate improves in lockstep, a distinction worth keeping separate.

Dream Dojo converts RL's 'environment cost' from physics-engine engineering into data-and-compute cost: instead of hand-coding collision detection and friction coefficients, the video world model learns 'what the next frame looks like given this action.' The appeal is sidestepping the sim-to-real gap, but the simulator's reliability now depends entirely on training-data coverage — out-of-distribution action combinations may produce physically implausible feedback that's harder to diagnose than a traditional physics engine's failure modes.

The talk's close embeds the paradigm shift in a larger commercial logic: whether it's WAM training or Dream Dojo simulation, compute demand far exceeds traditional physics-engine approaches, and Jim's line 'buy more, save more' openly reveals Nvidia's incentive as a chip vendor. That doesn't necessarily invalidate the technical argument, but readers should separate 'the paradigm has shifted' from 'who benefits from the shift' — especially since Nvidia's own GR00T N1.7 paper from April still uses 'VLA model' language, suggesting the internal transition isn't complete.