Kion Fallah used to train robots to drive. Now he's teaching them how not to break your spreadsheets.
That might sound like a demotion—from robotaxis navigating rush-hour traffic to software agents clicking through Excel—but the underlying problem is remarkably similar. Both require something the AI industry spent years learning to build: detailed simulations of consequence. What happens if the autonomous vehicle changes lanes here? What happens if the AI agent deletes this file? The architecture that answers the first question, it turns out, also solves the second.
Fallah spent years at Waabi building mixed-reality simulators accurate enough to train robots that would eventually carry human passengers. The stakes were existential—get it wrong and people could die. Now, as CEO of Experiential Labs, he's aiming that same simulation rigor at a different autonomy problem: AI agents that need to understand what their actions will do before they do them. The company emerged from Y Combinator's summer cohort with a pitch that would have seemed impossibly niche not long ago. Today, it sits near the center of a market realignment that nobody quite predicted.
Because here's what changed: large language models can reason about text beautifully, but agents that interact with actual systems—software interfaces, factory equipment, codebases under version control—need to reason about physics, causality, and cascading consequences across time. That requires simulation fidelity the autonomous vehicle industry spent the better part of a decade perfining, often at brutal expense.
When a Billion Dollars Flows Toward Imaginary Worlds
The capital pouring into world model infrastructure isn't speculative froth. It reflects a pragmatic realization among the smartest investors in Silicon Valley: if you want agents that work reliably in production, you need them to rehearse in environments that behave like the real world.
World Labs raised $1 billion earlier this year to commercialize general-purpose world models, including $200 million from Autodesk. Waymo, meanwhile, has been vocal about its World Model—built atop Google DeepMind's Genie 3 architecture—which it uses to simulate rare driving events at industrial scale. Gartner estimates that up to $234 billion in enterprise application spend could face disruption by agents by 2030, much of it hinging on a deceptively simple question: can these agents reliably predict outcomes before executing tasks?
Perhaps the most telling indicator comes from NVIDIA, which announced its Cosmos platform with considerable fanfare. Cosmos offers world foundation models explicitly designed for physics-aware video generation across robotics and autonomy applications. The Cosmos-Drive variant generates spatiotemporally consistent multi-view driving videos for synthetic data pipelines—technology that Waymo, Waabi, and their competitors now use to manufacture edge cases too statistically rare to encounter during real-world testing, no matter how many millions of miles you log.
California DMV data shows autonomous vehicle operators logged over 9 million public road miles between December 2024 and November 2025. Impressive, certainly. Yet simulation remains the primary tool for validating the long-tail scenarios that matter most for safety—the pedestrian who darts between parked cars, the debris in the road, the sensor glare at sunset.
Meta's V-JEPA 2 research demonstrated that self-supervised video models could enable planning with minimal action-labeled data, a critical capability when you're trying to train an agent to navigate unfamiliar software interfaces without exhaustive human annotation. The subsequent V-JEPA 2.1 update tightened robotics performance enough to suggest that video world models trained in one domain might transfer to others with enough architectural headroom. Might being the operative word—researchers remain cautiously optimistic rather than certain.
Applied Intuition, which raised $600 million at a $15 billion valuation, is a notable simulation vendor serving AV and defense customers. While not a generative world model in the technical sense, its tooling increasingly sits alongside—and sometimes inside—world-model pipelines, handling scenario authoring and sensor emulation that generative models alone can't yet cover end-to-end. The distinction matters less than it used to.
Where the Technology Actually Travels
The use cases extending beyond autonomous vehicles reveal where knowledge transfer is happening fastest, and the results occasionally surprise even the engineers building these systems.
NVIDIA's Isaac GR00T platform now trains vision-language-action models for humanoid robots using Cosmos-generated synthetic trajectory data. Antioch, which raised an $8.5 million seed round positioning itself as "the Cursor for physical AI," is building developer tooling atop NVIDIA and World Labs models to close the sim-to-real gap for robotics teams lacking Waymo-scale engineering resources. The gap between research-grade simulation and production deployment, in other words, remains wide enough to drive a startup through.
Then there's the desktop. A cluster of recent papers introduced "computer-use world models"—systems that simulate desktop and web environments so agents can forecast what happens when they click a button or fill a form before actually committing the action. The CUWM, WebWorld, and WMA frameworks all demonstrated lookahead planning that reduced task failures in multi-step workflows. The approach borrows directly from how driving world models predict collision outcomes if you brake versus swerve.
Experiential Labs frames this crossover explicitly, and with some justification. Co-founder Silen Naihin helped build AutoGPT to over 160,000 GitHub stars and co-authored the first open-source agent benchmark accepted to NeurIPS—work that established early evaluation standards for agent safety. Fallah's Waabi background in "realistic AI-based simulation to validate autonomy stacks" maps directly onto what Experiential Labs is attempting to solve: how do you know an agent's world model is accurate enough to trust when money or uptime is on the line?
The company's Y Combinator listing claims 99.7% reconstruction fidelity based on self-reporting, though that figure awaits independent benchmarking against emerging standards. Their CLaaS (Continual Learning as a Service) paper proposes an API letting agents improve during deployment via replay buffers and asynchronous training—a concept borrowed from reinforcement learning in robotics but adapted for agents that need to learn from mistakes without halting user workflows. On paper, it's elegant. In production, we'll see.
The Measurement Problem Nobody Wants to Talk About
Fidelity claims like "99.7% reconstruction" highlight the industry's thorniest challenge: how do you actually measure whether a simulation is good enough?
The WorldLens benchmark introduced a five-axis framework for assessing driving world models: realism, geometry, physics, behavior, and user preferences. The core finding—that no single model dominated all axes—suggests different applications may demand very different fidelity tradeoffs. A robotics planner might tolerate photorealistic imperfections if physics remains accurate. A creative tool might invert that priority entirely.
DrivingGen provides a dedicated benchmark for generative driving world models, while ongoing research explores latent video prediction methods, comparing V-JEPA 2.1, VideoPrism, and VideoMAEv2 on robustness criteria. But benchmarks proliferate faster than consensus, and vendors understandably gravitate toward whichever metric makes their system look best.
The regulatory landscape adds pressure, though enforcement remains uneven. The EU AI Act's general-purpose AI obligations took effect recently, with broader application requirements phasing in gradually. NIST's AI Risk Management Framework, updated with a generative AI profile, offers voluntary guidelines that world-model vendors can align to—particularly as these systems underpin safety-critical autonomy applications.
There's an interesting footnote in California's approach: the DMV explicitly excludes simulation from its public autonomous vehicle disengagement reports. That detail highlights the gap between real-world validation and sim-based development. Internal fidelity metrics—ideally standardized benchmarks like WorldLens—become essential for building safety cases when regulators eventually scrutinize how companies validate agentic systems in simulation before deployment. That reckoning is coming, whether the industry is ready or not.
What Happens Next (Besides Hype)
Forrester's recent report on agentic AI describes the technology as "technically viable" but notes that most enterprises remain stuck in what one analyst called "pilot-to-production limbo"—a purgatory familiar to anyone who remembers the blockchain era. IDC's forward-looking vision document outlines scenario planning for agentic transformation across industries through the end of the decade. The infrastructure gap those reports identify—reliable, verifiable simulation environments where agents can safely learn and be tested—is precisely where the AV-to-agents technology transfer matters most.
Market-size estimates vary wildly, as they always do at this stage. One analysis projects embodied AI growing from $4.44 billion to $23.06 billion by 2030 at a 39% compound annual growth rate; another forecasts $67.6 billion by 2033. Autonomous vehicle simulation solutions alone could reach $2.4 billion by 2030, depending on whose methodology you trust. The directional signal, however, remains consistent: organizations building agents need world models, and the companies that learned how to build high-fidelity, physics-aware simulators for robotaxis now shape how those agents get validated.
Runway's messaging around its Gen-3 architecture and GWM-1 family positions "general world models" as something larger than creative media applications. DeepMind has publicly framed Genie 3 as a "stepping stone toward AGI" in media briefings. Perhaps that's hyperbolic marketing. Or maybe it reflects a genuine intuition that intelligence doesn't emerge from text prediction alone—that it requires models capturing how actions propagate through time and space, with all the messy physics that entails.
For engineers evaluating simulation infrastructure today, the questions aren't purely technical. They're strategic, and occasionally existential: Which fidelity axes actually matter for your application? Can you trust vendor claims without independent benchmarks? Do you build in-house or integrate with platforms like Cosmos, World Labs, or Applied Intuition? And if you're Experiential Labs—a two-person Y Combinator team competing in a space where NVIDIA and DeepMind set the pace—how exactly do you carve out differentiation beyond founder pedigree and a single fidelity metric?
The answers will likely arrive unevenly, as they usually do. World models for driving are shipping in production autonomous vehicles today. World models for desktop agents remain largely experimental, confined to research labs and pilot programs. But the pattern has clarified: the next wave of AI agents won't learn by trial and error in live environments, breaking production systems while users watch in horror.
They'll rehearse in simulations built by people who spent the last decade teaching cars to see around corners. Whether that's enough remains the open question.
