The problem with physical robots replacing human workers has been the programming necessary to duplicate human level knowledge and skills.
That problem is now being addressed by installing AI into robots and allowing robots to learn tasks by simply watching, just as humans do. This is the key technology that will make replacement of blue-collar workers a reality sooner rather than later.
Looking around Ai is still reserved for those who have the time capital, who have the basic knowledge. For a chef or a builder on site everyday they don’t have the time to sit down and ask ai chatbot for thoughts before proceed. I would say that ai is yet at the “IPhone moment” yet, where it is so penetrating, that everyone use or at least a version of phone.
At the same time, most industries isn’t ready for Ai, not in terms of not receptive to it, more often than not, they love the idea of Ai, as if it’s the answer to the manual work they been doing. They don’t need Ai, they are still step below it, they need automation.
The six-month validation cycle stuck with me. By then the model may have changed again. That lag seems like the real problem for physical AI. You can’t ship a truck the same way you ship an app update.
I agree completely and I have thought about that all week! As AI able to do more work , we have to find ways to define what “good” work looks like! That’s the validation!
Twenty years from now, we won’t remember which company had the best world model in 2027. We’ll remember which company had the best LAM - Large Action Model, as I wrote in 1982
Aviation solved the easy half of this decades ago: writing the software is the cheap part, certifying it sets the schedule. A model that improves every few weeks against a validation cycle that takes months just stacks up capability nobody's allowed to ship. My guess is the first physical-AI shop that makes validation continuous instead of episodic wins on tempo, not model quality.
Our biggest mistake is looking at today’s interconnected global crises as separate issues when they are in fact symptoms of one deeply interlocked system. Economic volatility, the climate crisis, cyber attacks, misinformation, broken governance, declining public trust, geopolitic tensions, public health crises, and the exponential pace of technological innovation now feed into each other. A cyberattack could cripple global markets; conflict can drive scarcity in food and energy; environmental ruin can increase public health crises and migration; and information pollution weakens our faith in the systems we need to build to confront challenges. By simply responding to the symptoms-the job losses, the gas prices, the virus, the attacks, the protests, the rising temperatures-and not the underlying system failures that produce them, leaders and decision makers simply export the problem to another sector or to another generation.
This requires a new model of strategic leadership with a bias toward long-term, integrated responses rather than crisis-driven reactions. This should include governments investing in clear and trusted institutions, in resilient infrastructure, education that reflects a world in flux, modern and inclusive public health systems, and governance systems that ensure innovative technologies and advancements are deployed in responsible ways. This requires companies going beyond quarterly expectations and investing in a form of innovation that addresses social, economic, and environmental concerns over the long term. And it involves researchers, engineers, educators, civil society organizations, entrepreneurs, and communities coming together with the shared understanding of the necessary and accelerated progress in AI ethics, Cybersecurity, renewable energy, advanced materials science, and scientific understanding – that will deliver broadly shared benefits, not just individual or sectoral advantages.
Twenty-first-century success will be measured less by military power or gross output and more by a capacity to build and maintain resilient systems capable of weathering disruption, managing technological transitions, conserving our natural world and building public trust. The leaders who will shape the future will be those who embrace every crisis as part of a larger interconnected web of systems that they address collectively and strategically through science, empathy, transparency and well-informed public-policy processes.
The future of the world isn’t under threat by a single crisis; it is being threatened by a failure of imagination and integration in policy and leadership. The solution will not be single, isolated technologies or policies but rather an integrated strategy that connects the pillars of security, economics, environment, technology, education and social cohesion to drive collectively agreed upon outcomes. Perhaps our most important duty as individuals and as societies isn’t to simply solve today’s problems but to build the capacity to prevent tomorrow’s. The future they inherit depends upon our foresight today.
If models really do commoditize toward the frontier, the interesting consequence is that the moat inverts: it stops being the intelligence and becomes the proprietary field data plus the loop that turns it into validated software. And that's a much better moat because data and validation cycles compound, model weights leak.
The iterative development loop in physical AI, from requirement to tested deployment, is at least one order of magnitude more challenging than it is in traditional web software application development.
As a result, there should be special attention and a strong focus on engineering and improving that iterative loop itself, especially in physical AI.
Every engineer should understand this and have it on the top of mind when approaching intelligence in this domain. Stagnation is inevitable otherwise.
The point: advanced models have helped them build the "thousands" of apps they actually need to do the work with physical world sensor layers (sense, categorize, synthesize) that actually moves their company's value proposition forward (decisions that improve and speed up responsible real world building). Thus, the important constraint to solve for the future is found in the speed to develop (and quality of) those apps / that app building work - which equates to engineering, or deployed intelligence.
Good piece. My last startup built RFID sensor systems at the edge, before anything agentic existed, and the hard part was never the intelligence. It was getting the thing validated, integrated, and running reliably in the field. Same story in enterprise AI today. Everyone has the same models, most teams still die in the pilot. Models don't scale, systems do. Especially well-engineered ones, built to hold up before they hit production and learn once they're there. That's been true the whole time.
There's not moat in so called “ai” . It's all commodities already
The problem with physical robots replacing human workers has been the programming necessary to duplicate human level knowledge and skills.
That problem is now being addressed by installing AI into robots and allowing robots to learn tasks by simply watching, just as humans do. This is the key technology that will make replacement of blue-collar workers a reality sooner rather than later.
Looking around Ai is still reserved for those who have the time capital, who have the basic knowledge. For a chef or a builder on site everyday they don’t have the time to sit down and ask ai chatbot for thoughts before proceed. I would say that ai is yet at the “IPhone moment” yet, where it is so penetrating, that everyone use or at least a version of phone.
At the same time, most industries isn’t ready for Ai, not in terms of not receptive to it, more often than not, they love the idea of Ai, as if it’s the answer to the manual work they been doing. They don’t need Ai, they are still step below it, they need automation.
If you need a chatgpt moment for physical AI
The six-month validation cycle stuck with me. By then the model may have changed again. That lag seems like the real problem for physical AI. You can’t ship a truck the same way you ship an app update.
I agree completely and I have thought about that all week! As AI able to do more work , we have to find ways to define what “good” work looks like! That’s the validation!
Twenty years from now, we won’t remember which company had the best world model in 2027. We’ll remember which company had the best LAM - Large Action Model, as I wrote in 1982
Aviation solved the easy half of this decades ago: writing the software is the cheap part, certifying it sets the schedule. A model that improves every few weeks against a validation cycle that takes months just stacks up capability nobody's allowed to ship. My guess is the first physical-AI shop that makes validation continuous instead of episodic wins on tempo, not model quality.
Seems like intelligence gets you into the race. The speed of learning may decide who wins it. 😊
Our biggest mistake is looking at today’s interconnected global crises as separate issues when they are in fact symptoms of one deeply interlocked system. Economic volatility, the climate crisis, cyber attacks, misinformation, broken governance, declining public trust, geopolitic tensions, public health crises, and the exponential pace of technological innovation now feed into each other. A cyberattack could cripple global markets; conflict can drive scarcity in food and energy; environmental ruin can increase public health crises and migration; and information pollution weakens our faith in the systems we need to build to confront challenges. By simply responding to the symptoms-the job losses, the gas prices, the virus, the attacks, the protests, the rising temperatures-and not the underlying system failures that produce them, leaders and decision makers simply export the problem to another sector or to another generation.
This requires a new model of strategic leadership with a bias toward long-term, integrated responses rather than crisis-driven reactions. This should include governments investing in clear and trusted institutions, in resilient infrastructure, education that reflects a world in flux, modern and inclusive public health systems, and governance systems that ensure innovative technologies and advancements are deployed in responsible ways. This requires companies going beyond quarterly expectations and investing in a form of innovation that addresses social, economic, and environmental concerns over the long term. And it involves researchers, engineers, educators, civil society organizations, entrepreneurs, and communities coming together with the shared understanding of the necessary and accelerated progress in AI ethics, Cybersecurity, renewable energy, advanced materials science, and scientific understanding – that will deliver broadly shared benefits, not just individual or sectoral advantages.
Twenty-first-century success will be measured less by military power or gross output and more by a capacity to build and maintain resilient systems capable of weathering disruption, managing technological transitions, conserving our natural world and building public trust. The leaders who will shape the future will be those who embrace every crisis as part of a larger interconnected web of systems that they address collectively and strategically through science, empathy, transparency and well-informed public-policy processes.
The future of the world isn’t under threat by a single crisis; it is being threatened by a failure of imagination and integration in policy and leadership. The solution will not be single, isolated technologies or policies but rather an integrated strategy that connects the pillars of security, economics, environment, technology, education and social cohesion to drive collectively agreed upon outcomes. Perhaps our most important duty as individuals and as societies isn’t to simply solve today’s problems but to build the capacity to prevent tomorrow’s. The future they inherit depends upon our foresight today.
If models really do commoditize toward the frontier, the interesting consequence is that the moat inverts: it stops being the intelligence and becomes the proprietary field data plus the loop that turns it into validated software. And that's a much better moat because data and validation cycles compound, model weights leak.
The iterative development loop in physical AI, from requirement to tested deployment, is at least one order of magnitude more challenging than it is in traditional web software application development.
As a result, there should be special attention and a strong focus on engineering and improving that iterative loop itself, especially in physical AI.
Every engineer should understand this and have it on the top of mind when approaching intelligence in this domain. Stagnation is inevitable otherwise.
The point: advanced models have helped them build the "thousands" of apps they actually need to do the work with physical world sensor layers (sense, categorize, synthesize) that actually moves their company's value proposition forward (decisions that improve and speed up responsible real world building). Thus, the important constraint to solve for the future is found in the speed to develop (and quality of) those apps / that app building work - which equates to engineering, or deployed intelligence.
Good piece. My last startup built RFID sensor systems at the edge, before anything agentic existed, and the hard part was never the intelligence. It was getting the thing validated, integrated, and running reliably in the field. Same story in enterprise AI today. Everyone has the same models, most teams still die in the pilot. Models don't scale, systems do. Especially well-engineered ones, built to hold up before they hit production and learn once they're there. That's been true the whole time.