I recently received a panicked message from my girlfriend. She’d been following the news about Jacob Coxon’s resignation from Anthropic, and was rightly concerned about the flood of news articles repeating the “end of the decade” timeline for the AI-powered apocalypse.
The purpose of this essay isn’t to dive into the AI safety debate; I do want to highlight one implication of this prediction, one that has largely slipped under the prevailing narrative. Whilst it’s unsurprising that the general public are unnerved by the exponential increases in LLM performance and capacity over the last few years, the assumption that digital text-based chatbot and agentic knowledge work performance increases will mirror across into rapidly scaling capabilities for embodied AI in the physical world ignores a problem that researchers in robotics have been grappling with for decades: in purely digital tasks, advances in reinforcement learning and feedback loops produce new, improved iterations within days; embodied, physical AI’s timescale is glacial by comparison.
I’m not dismissing digital AI risk; code can easily engender physical harm through attacks on infrastructure, for example. When faced with the deluge of bio, cyber, and autonomous-agent risks that current AI progress enables, many outside of the tech sector, my girlfriend included, are rightly bewildered. However, one small reassurance that has largely been ignored by the public debate (and also in a frantic text conversation with my girlfriend) is that there are structural constraints that mean the disorienting pace that LLMs have progressed over the last few years will struggle to be replicated in physical AI by 2030.
The first of these is data. Scaling laws are centred on three variables: data size, model size, and compute size. For LLMs, the internet has provided a vast, publicly available text-based data source that just needed to be scraped, labelled and fed into a model. The bottleneck came with compute, and once rent for chips such as NVIDIA’s H100 dropped by as much as 40% since 2022 (Sourced), the possibility of a GPT-level model became a matter of time.
General-purpose applied AI is still blocked at the first hurdle. The training tokens required to bootstrap a GPT for robotics simply do not exist at the scale needed. This means the only way to generate the real-world data needed to train embodied AI is through actual deployments. However, the question then becomes: how do you sufficiently pretrain a model to the point where it can even act meaningfully in the world to collect this data? There is no ImageNet database, or years of internet data and chatbot interactions as there were for computer vision models or LLMs respectively, that you can use to get a minimum viable product to deployment-grade performance in order to start generating real-world data to improve performance. This is the ‘cold-start’ problem in robotics: you need real-world training data to deploy, and you need deployments to generate real-world training data.
This leaves physical AI startups with the lab, the theory being that if you don’t have the real-world training data, you can generate it artificially by running thousands of hours of simulations. But after the scarcity of training data, the second constraint on rapid, runaway progress in embodied AI is the difficulty of transferring lab-based training into real-world deployments. This is the infamous Sim2Real gap that robotics researchers have been grappling with for decades. A robotics policy that achieves 95% success in Isaac Gym, NVIDIA’s GPU-accelerated physics simulation platform, can fail when deployed on real hardware with even the most seemingly basic tasks (Claru).
It might seem trite that our physical world is infinitely complex, but it’s only when you try to recreate physical interactions in a predictable manner that you realise just how true this is. Something as simple as a paper cup, our interactions with which we would not pause to register in day-to-day life, becomes a significant challenge for lab-based simulations to accurately model in a way that reflects the physical world.
Rather than a single chasm, the Sim2Real gap is actually many smaller cracks, that when aggregated, contribute to serious performance drop-offs between the lab and reality. Real-world contact interactions are inherently complex and nonlinear, and simulations approximate friction, contact states, and deformable bodies in order to cope with this. MuJoCo (Multi Joint dynamics with Contact, an open-source physics engine) struggles to match contact dynamics for soft, pliable materials. These engines can partially compensate with domain randomisation across mass, friction coefficients, and joint damping, but they are systematically limited; if the policy hits real contact physics that fall outside the randomisation range, the policy will fail. To return to the example of the paper cup, it is simply unbelievably hard to model the exact way a pliable, deformable material will interact with a physical force.
Similarly, policies created on simulations run into difficulties on production systems with sensors. All sensors output noise and simulations’ noise profiles are usually just assumed to be Gaussian to keep compute costs manageable. The problem is that real-world sensor noise is not simply Gaussian: RGB cameras have lens flares, chromatic aberrations and flying pixels; joint encoders accumulate drift; and tactile arrays produce complex noise (arxiv.org). None of this can be reproduced in simulators like Isaac Gym. Robotics policies trained on models that cannot account for non-Gaussian noise will hit patterns in production that they are not trained to handle and fail.
If training data is scarce, and the Sim2Real gap holds back lab-based development, the only break in the cycle can come through forcing an MVP into deployment. But this requires a serious amount of capex. Getting physical AI systems into live production environments is enormously expensive and incurs a series of costs that pure-play software deployments do not have to contend with around maintenance, transport, part deterioration, replacements, and supervision. Not only that, but robot generalisation requires diverse datasets across many contexts, tasks and environments. So capex costs do not necessarily fall as you deploy under the economies of scale, as domain-shifting generates new manufacturing and servicing outlays.
People are right to be concerned about the exponential increase in language models’ capabilities over the last few years. AI is not a monolithic technology, and although LLMs are the dominant structure that we interact with day-to-day, this is only within a digital context. Out there in the real world, we’re still grappling with some very physical, very hard problems, that mean the dizzying, disorienting pace of change that people associate with AI will not be replicated so quickly.