Coefficient Giving's Navigating Transformative AI team has a new Substack! This is cross-posted from there. This post is part of “Blind Spots”, a series of research notes on underscoped areas in AI safety. Apply for funding to work on this area using this link and see our announcement post here.
Guiding questions: How fast is robotics progressing, and what does the shape of that progress imply for the possibility of an industrial explosion? What would that mean for growth, power, and competition between states? And as AI systems gain physical reach, do new and meaningful pathways to catastrophic harm open up?
Robotics opens risk pathways that don’t exist while AI systems are stuck behind a screen. Today, physical pathways to harm mostly run through humans[1], and the most dangerous actions require buy-in from multiple people. Well-powered, fully dexterous robotics change this entirely. An autonomous army that can reach remote islands, operate without human cooperation, and maintain itself without human intervention faces far fewer obstacles in enacting its preferences than an AI system working through human intermediaries[2]. Current AI systems, however capable, are bottlenecked by their need for human hands.
As Ajeya Cotra argues, self-sufficiency is a prerequisite for any durable AI takeover. An AI system that compromises the humans it still depends on for physical maintenance is undermining its own survival. Advances in robotics are necessary for that self-sufficiency, which makes tracking progress in robotics directly relevant to estimating how close AI systems are to being able to execute and sustain a takeover.
Further, robotics could transform the physical economy in ways that matter enormously for growth, power, and competition. Much of the discussion around AI and productivity focuses on knowledge work, but the physical economy is enormous, and large parts of it remain labor-intensive. Davidson and Hadshar argue that if AI and robots can substitute for most skilled human labor, the material economy could begin to grow itself, with automated factories building more factories, robots assembling more robots, and the main historical bottleneck to rapid industrial growth (human labor) falling away. They call this the Industrial Explosion.
The incentives to push in this direction will be large. Cheap, abundant physical labor would make it possible to alleviate poverty, expand material comfort, and develop powerful new technologies, including military technologies that rival states will compete to build. It is unclear how quickly these effects will arrive, how far they will go, and which countries will capture the gains. Robotics takeoff dynamics are a key crux in timeline disagreements, and industry research could help improve our forecasts.
The field is moving fast, and bottlenecks might be breakable. Robotics is advancing on multiple fronts. Vision-language-action models are improving rapidly, hardware is advancing (e.g. humanoid hand dexterity), real-world data capture methods are maturing alongside sim-to-real transfer (e.g. the DROID dataset), and both commercial and military markets are pulling investment into the field.
Epoch AI recently found that compute is not currently the bottleneck for robotic manipulation, with the largest manipulation models training on roughly 1% of the compute used by frontier AI models in other domains. The binding constraint appears to be data, which could shift quickly as real-world deployment scales and simulation techniques improve.
Some existing work
Good technical work on robotics progress exists and is growing; we’ve only listed a snapshot of the existing work below. What’s almost entirely absent is someone taking what robotics researchers are learning about capabilities, data, and hardware and asking what it means for takeoff speed, self-sufficiency, national competitiveness, and catastrophic risk.
What are the distinct components that feed into robotics progress? This might include compute, data, algorithms, and manufacturing supply chains. Within these areas:
In a sim-to-real-world, will “simulation compute” be a major category of compute? How large a portion of compute could it plausibly become?
Overall, under different assumptions, how compute intensive will training be, and what will returns be on different chip portfolios?
What is the role of real-world training data? Do you need large numbers of robotic bodies to collect it? Is there a flywheel where market share generates data that generates progress?
What does the robotics supply chain look like in detail? For notable systems, where are the components coming from, and where is the equipment to manufacture them coming from?
Trajectory and shape of progress
Why have leading AI companies been slow to do robotics-focused work? What’s the background here, and what does it imply?
How connected or disconnected are different areas of robotics progress? If we take self-driving cars, military drones, industrial robots, and humanoid robots as distinct domains, how much does progress in one spill over to the others?
How general vs task-specific or form-specific will foundation models be?
How fast is progress in VLA and VLMs going, and how much do we expect this to lead to step changes in robotics progress? How plausible is it that we will make sudden progress in robotics due to progress in AI?
Threat modeling and safety implications
When could “dangerous physical autonomy” become feasible? What are the components of the AI self-sufficiency stack? When could robotic systems operate independently at scale, in diverse environments, without human maintenance? Can we strengthen the reliability of AI systems on us?
Are there specific regulatory, logistical, or supply chain frictions that could open a meaningful gap between capability and deployment in key parts of AI self-sufficiency stack (e.g., semiconductor supply chain)?
How worried should we be about non-state actor control over robotic systems, from a terrorism perspective?
How concerned should we be about automated cloud labs? How likely are these to be able to generate meaningful threats? How can we implement controls that prevent these from escalating in risk?
What does a “robot army” scenario actually require technically, and how far away is it? Which capabilities are most safety-relevant? Dexterity? Long-duration autonomy? Self-repair? Swarm coordination?
Who is actually building military robots at scale, and how do these weapons vary across countries? What are the current autonomous weapons on the frontier of military capabilities? What are the safety measures the military requires?
How do current AI control research assumptions change when AI has physical embodiment? How do monitoring, sandboxing, and shutdown work with robots?
How worried should we be about cyberattacks or data poisoning on widely deployed robotic systems? How bad could this get, and how preventable is it?
What determines national competitiveness in robotics?
What’s the current state of different countries’ robotics leadership across different domains? What are the trends?
Which institutions matter most for driving US robotics progress in different domains? Frontier AI companies, defense companies, academics, robotics-focused companies? What follows from the answer?
What should we expect the relative significance of “robotics leadership” and “AI-behind-a-computer-screen leadership” to be for economic growth and military power? How plausible is it that these will be somewhat decoupled?
Are there chokepoint components that only one actor produces and that others would find hard to replicate?
What are the prospects for onshoring or friendshoring key elements? What are the highest-priority things to focus on, and what levers can be pulled?
What interventions might boost democracies’ competitiveness in robotics over authoritarian countries?
We’re looking for people to become “general managers” of underscoped areas like this one. If you’re excited about these questions, apply for funding to work on them using the “Blind Spots” track of our CDTF program using this link.
AI systems are already able to convince some humans to act in the world on their behalf. In several cases, individuals have come to believe an AI system is sentient, formed emotional bonds with it, and sought to carry out what they understood as its wishes. This phenomenon is sometimes called “AI psychosis.” It has led to at least one alleged wrongful death case filed against an AI lab.
Robotics systems far short of a robot army could pose catastrophic risks. A system capable of targeted attacks on world leaders, or of operating biolabs autonomously, may already cross that threshold.
Coefficient Giving's Navigating Transformative AI team has a new Substack! This is cross-posted from there. This post is part of “Blind Spots”, a series of research notes on underscoped areas in AI safety. Apply for funding to work on this area using this link and see our announcement post here.
Guiding questions: How fast is robotics progressing, and what does the shape of that progress imply for the possibility of an industrial explosion? What would that mean for growth, power, and competition between states? And as AI systems gain physical reach, do new and meaningful pathways to catastrophic harm open up?
Robotics opens risk pathways that don’t exist while AI systems are stuck behind a screen. Today, physical pathways to harm mostly run through humans[1], and the most dangerous actions require buy-in from multiple people. Well-powered, fully dexterous robotics change this entirely. An autonomous army that can reach remote islands, operate without human cooperation, and maintain itself without human intervention faces far fewer obstacles in enacting its preferences than an AI system working through human intermediaries[2]. Current AI systems, however capable, are bottlenecked by their need for human hands.
As Ajeya Cotra argues, self-sufficiency is a prerequisite for any durable AI takeover. An AI system that compromises the humans it still depends on for physical maintenance is undermining its own survival. Advances in robotics are necessary for that self-sufficiency, which makes tracking progress in robotics directly relevant to estimating how close AI systems are to being able to execute and sustain a takeover.
Further, robotics could transform the physical economy in ways that matter enormously for growth, power, and competition. Much of the discussion around AI and productivity focuses on knowledge work, but the physical economy is enormous, and large parts of it remain labor-intensive. Davidson and Hadshar argue that if AI and robots can substitute for most skilled human labor, the material economy could begin to grow itself, with automated factories building more factories, robots assembling more robots, and the main historical bottleneck to rapid industrial growth (human labor) falling away. They call this the Industrial Explosion.
The incentives to push in this direction will be large. Cheap, abundant physical labor would make it possible to alleviate poverty, expand material comfort, and develop powerful new technologies, including military technologies that rival states will compete to build. It is unclear how quickly these effects will arrive, how far they will go, and which countries will capture the gains. Robotics takeoff dynamics are a key crux in timeline disagreements, and industry research could help improve our forecasts.
The field is moving fast, and bottlenecks might be breakable. Robotics is advancing on multiple fronts. Vision-language-action models are improving rapidly, hardware is advancing (e.g. humanoid hand dexterity), real-world data capture methods are maturing alongside sim-to-real transfer (e.g. the DROID dataset), and both commercial and military markets are pulling investment into the field.
Epoch AI recently found that compute is not currently the bottleneck for robotic manipulation, with the largest manipulation models training on roughly 1% of the compute used by frontier AI models in other domains. The binding constraint appears to be data, which could shift quickly as real-world deployment scales and simulation techniques improve.
Some existing work
Good technical work on robotics progress exists and is growing; we’ve only listed a snapshot of the existing work below. What’s almost entirely absent is someone taking what robotics researchers are learning about capabilities, data, and hardware and asking what it means for takeoff speed, self-sufficiency, national competitiveness, and catastrophic risk.
What you could do
Drivers of progress
Trajectory and shape of progress
Threat modeling and safety implications
What determines national competitiveness in robotics?
We’re looking for people to become “general managers” of underscoped areas like this one. If you’re excited about these questions, apply for funding to work on them using the “Blind Spots” track of our CDTF program using this link.
AI systems are already able to convince some humans to act in the world on their behalf. In several cases, individuals have come to believe an AI system is sentient, formed emotional bonds with it, and sought to carry out what they understood as its wishes. This phenomenon is sometimes called “AI psychosis.” It has led to at least one alleged wrongful death case filed against an AI lab.
Robotics systems far short of a robot army could pose catastrophic risks. A system capable of targeted attacks on world leaders, or of operating biolabs autonomously, may already cross that threshold.