No LLM generated, heavily assisted/co-written, or otherwise reliant work.
Read full explanation
I work at a semiconductor fabrication facility in the USA for Samsung as an engineer developing and deploying a variety of AI models (traditional CV and LLMs) for manufacturing operations. This post analyzes whether AGI—defined as AI capable of all digital white collar work at 10x human speed—could meaningfully accelerate current semiconductor manufacturing growth trends with or without access to humanoid robots. My conclusion is that even in the highly unlikely case that humanoid robots are brought online tomorrow, chip manufacturing would have a 8.4-9.1% CAGR instead of a baseline 7% CAGR, an increase of 20-30% over the next 2-3 years. Same goes for advanced (<7nm) chips, increasing from 14% CAGR to 16.8-18.2% It seems like manufacturing is a strong blind spot for the AI safety community, and so I am hoping to add to the conversation with domain expertise.
Epistemic status: Moderately confident (70-80%) about the analysis of existing bottlenecks which could be addressed by AGI based on 3+ years building and deploying AI in semiconductor fabs. Much less confident about AGI capability assumptions and what they would do to chip production. The core claim that manufacturing growth rates would only see 20-30% uplift seems plausible but is ultimately a guess based on my personal experience and reading about current fab scaling times. I am likely wrong about specific timelines and lots of information needed to make a better prediction is proprietary to private corporations with a strong commercial interest in not making this information available.
Why Chip Production Growth Matters for AI Safety
What would it be worth epistemically if we could confidently rule out a contributing factor or sole path to runaway AI progress? It is important to separate chip production out from pathways like algorithmic progress, hardware design progress, and data scaling because it seems like it might be safely ruled out as a path (or even a factor contributing more than 1% to capabilities growth) for recursive self improvement. People in AI safety could reallocate their attention to potentially more fruitful areas of research such as identifying timelines related to algorithmic progress, work which Epoch AI has recently put out a call for and which I would like to highlight here.
How Would Semiconductor Manufacturing Growth Rates Change if AGI was Built Tomorrow?
If we achieved AGI tomorrow (AI doing all digital white collar work at 10x human speed), would it significantly accelerate cutting-edge AI chip fab production growth trends over the next decade? This is narrower than "will AGI cause fast takeoff?" but addresses one specific pathway: whether AGI could bootstrap itself by accelerating hardware manufacturing.
Why "All Digital White Collar Work at 10x Human Speed"?
I chose this definition because it represents a plausible order of magnitude estimate on what near-term AI systems might achieve and this order of magnitude of speed and intelligence often shows up in projects like AI 2027 as a first step towards dangerously accelerated timelines.
Semiconductor Manufacturing Primer
For readers unfamiliar with chip manufacturing, here are the key concepts:
What is a fab? A semiconductor fabrication facility ("fab") is where computer chips are manufactured. Building a cutting-edge fab costs $10-25 billion USD and takes 2-4 years. They're essentially the most complex factories humans build. They must be very clean, with a maximum of 10 or 100 particles 0.5 microns or larger in each cubic foot of air. They contain hundreds or thousands of specialized machines performing thousands of process steps on silicon wafers which make up the majority of the cost of the fab.
Why is ASML critical? ASML manufactures extreme ultraviolet (EUV) lithography machines, which are the only equipment capable of printing features small enough for cutting-edge chips (3nm, 5nm nodes). Each machine costs ~$300 million, weighs 180 tons, and ASML ships only ~65 of these units per quarter globally. Without EUV machines, you cannot manufacture cutting-edge AI chips. No other company makes them.
What is yield and why does it require physical testing? Yield is the percentage of chips that work correctly after manufacturing. A 70% yield means 30% of chips are defective and must be discarded. The challenge is that thousands of process steps interact in ways we don't fully understand. Defects can have multiple root causes—contamination, equipment drift, material variations, quantum effects at nanometer scale. Even with perfect AI analysis, you must run actual production wafers to collect sensor, scan, and test data to observe how defects manifest in practice.
Hypothesis: AGI Would not Significantly Accelerate Semiconductor Manufacturing
Even with AGI and a capacity to be embodied in humanoid robots, chip manufacturing growth rates may only increase by 20-30% in the next 2-3 years.
This is because:
Equipment manufacturing is hard to expand rapidly: ASML can only produce ~65 EUV machines/quarter with 12-18 month lead times and like other manufacturing processes this seems unlikely to be dramatically accelerated by AGI.
Empirical yield optimization requires physical iteration: Fabs must run production wafers through thousands of process steps, analyze failures, adjust parameters, and repeat in order to collect data, even if AGI was a perfect analyzer.
Physical/chemical processes have inherent time constants: Concrete curing, equipment burn-in testing, time-dependent failure modes, reliability qualification.
Lead times scale with infrastructure requirements:Epoch AI finds "lead time grows roughly one year for every ten-fold increase in compute"
How Much Could AGI Accelerate Growth Rates?
Baseline Growth Projections
SEMI projects global semiconductor capacity growing for the leading-edge (≤7nm) technology used in making AI chips by +14% CAGR.
What AGI Could Accelerate
High-impact areas:
Yield optimization: AGI could accelerate defect analysis, parameter search, and root cause correlation preventing large wafer loss events and low yield which can render making semiconductor wafers unprofitable (this is what I work on).
Planning and design: AGI can optimize facility design, logistics, and planning at 10x speed, probably by replacing and/or uplifting line planning departments.
Construction planning: LLMs are already quite good at planning, and despite some early struggles with physical systems, it seems plausible that AGI could greatly accelerate these timelines.
Engineering work: Process engineers, equipment engineers, yield engineers all replaced/uplifted by AGI so that their work might easily be completed 100-1000x faster, essentially removing any constraints from white collar work.
Training Human Technicians: Semiconductor fabs are currently constrained for blue collar workers able to perform repairs and maintenance on the fab and the equipment inside of it. AGI could rapidly scale such training programs and remove the supply constraint.
Strong Physical and Operational Growth Constraints
These are the constraints that remain even with white collar AGI doing all engineering and planning work, but without humanoid robots:
1. ASML Equipment Manufacturing Bottleneck
Current capacity: ~65 EUV machines/year, 12-18 month lead times
Even with AGI: ASML is limited by:
Specialized component suppliers with their own manufacturing constraints.
Precision manufacturing processes requiring physical time (grinding optics, assembling systems).
Assembly: I couldn't nail down a consistent assembly time to a single source, but generally all of them say 4-5 months, and more like 7-9 months if you include setting it up inside the fab, which is consistent with my experience.
Testing and calibration of extreme precision equipment.
AGI can't fix this quickly: Even if ASML uses AGI for all design/planning, their suppliers need time to scale production of specialized components, and physical assembly/testing cannot be arbitrarily accelerated.
Must run production wafers through thousands of process steps, analyze failure modes, adjust process parameters, and run again to validate improvements. Historically, this has taken fabs pushing the leading edge (which is necessary for AI chips) 1-2 years.
What AGI accelerates: Defect analysis, root cause identification, and parameter optimization.
What AGI cannot accelerate: A single physical wafer run can take months from entering as bare silicon to exiting as a production ready wafer before packaging.
Why can't we just simulate yield instead of running physical wafers?
The fundamental challenge is the high volume of process steps with interactions we don't fully understand:
Unpredictable interactions: Each lithography, etching, deposition, and cleaning step affects subsequent steps in ways that aren't fully captured by models. A parameter change in step 50 might cause defects in step 350.
Multiple root causes: Defects can come from contamination, equipment drift, material variations, pattern-dependent effects, or combinations thereof. AI can help identify patterns faster, but you need real defect data first. Also, very often your sensors may simply not be able to capture a signal which could be detected by an AI in the first place, and so you need empirical trial runs.
Scale-dependent failures: Many failure modes only appear in volume production. Patterns that work perfectly on test wafers fail when you're running 1,000 wafers/day.
Quantum effects: At 3nm nodes, quantum tunneling and other nano-scale effects create unpredictable behavior that cannot be fully simulated with current physics models (to my knowledge).
Unknown unknowns: You discover new types of defects as you scale production. You can't simulate what you don't know exists yet.
Even with perfect AI analysis, it is necessary to run actual production wafers to observe how defects manifest in practice. AGI can dramatically speed up the analysis and learning from each iteration, but it seems extremely unlikely to eliminate the need for physical iterations.
3. Physical and Chemical Time Constants
Concrete curing: Often around one month to reach full strength with existing chemical processes.
Equipment burn-in testing: Must run equipment for weeks/months to reveal time-dependent failures.
Cleanroom particle settling: Time-dependent process for achieving ISO Class 1 standards.
What AGI helps: Better planning around these constraints, optimal scheduling.
What AGI cannot help: Making chemical reactions happen faster, revealing time-dependent failures without waiting.
Physical construction work (concrete, steel, installation) takes time
Weather delays
Sequential dependencies (can't install equipment until structure is complete)
What About Humanoid Robots?
Humanoid robots seems extremely unlikely to come about in the next ten years through normal human research (in my opinion, which is based largely on those of my robotics professor), but what if AGI somehow accelerated this progress and also came with capable humanoid robots? It seems that even with fully functional humanoid robots, AGI would not significantly help accelerate chip manufacturing for fundamental reasons:
Reliability gap: Fabs need as much uptime as possible when EUV machines cost $300M. One pathway AGI could affect this is by replacing equipment and process technicians with humanoid robots. Less specialized, more complex robotics introduce additional room for error, and the AGI would have to be more reliable than humans in performing physical tasks like maintenance on equipment or construction of fabs. What happens when the computer vision model mistakes two very similar seals and puts the wrong one in a piece of equipment? The answer is enormous losses, and in order for these humanoids to be adopted, they would have to undergo at least months of testing to demonstrate they are substantially more reliable than humans, which is a strong weakness of current transformer architectures. If humanoid robots proved to be perfectly reliable, they would still be closing a tiny gap in production. Fabs have extremely high reliability standards, so you might see something like a less than one percent increase in production due to reduced scrap and improved yield.
Existing automation is better: Fabs creating cutting edge semiconductor wafers for AI are already fully automated with the exception of human maintenance and repair on the equipment itself in the fab. Purpose-built systems are more reliable, don't contaminate cleanrooms, and don't need the flexibility humanoids provide. To make wafers all you need to do is repeat the same motions millions of times. This one is hard to measure and prove because of how close to its chest the industry plays its cards, but think about it this way: Would it be easier for 100 humans to make waffles with their hands, or a giant automated oven full of waffle irons with a single human sitting next to it who could fix it if something broke once every few months?
Technicians and other physical human labor are not the constraint: Current training timelines are fast and you could get people up to speed rapidly. Quick-start programs can take 8-10 weeks for entry-level roles, while certificate programs can last a year and it takes two years for an associate's degree for a full technician qualification. Most of these options are much faster than trying to manufacture millions of robots to scale semiconductor fab construction and operations. There is a workforce shortage:SIA projects that 39% of chip factory technician jobs may remain vacant by 2030, withMcKinsey estimating a deficit of 90,000 skilled technicians. However, unlike other constraints on growth, this one is highly responsive to investment:
CHIPS Act allocated $200 million for semiconductor workforce development, with broader STEM workforce programs totaling $33 billion.
White collar / digital AGI could dramatically accelerate this by: optimizing curricula, creating better training materials, scaling online/hybrid programs, improving retention through better teaching.
So while technician supply is currently constrained, it's not serious physical constraint or bottleneck. I think it's an organizational/educational scaling challenge that AGI could significantly help solve.
As a result, it seems that while having more technicians in the form of humanoid robots could be useful and potentially accelerate growth, because modern fabs are already fully autonomous outside of repairs, humanoid robots would not significantly uplift AGI's ability to accelerate growth. Given that the technology for humanoid robots following shortly after AGI is a already in question due to Moravec's paradox, it seems that humanoid robots would not be a significant part of the growth equation.
What Would Change My Mind
What evidence would make me update toward AGI significantly accelerating manufacturing?
1. Breakthrough in simulation accuracy: If AGI could accurately predict yield outcomes without physical wafer runs. This would require solving the "thousands of interacting process steps with quantum effects" problem. I think this would require essentially perfect physics simulation at the nano-scale. This doesn't seem plausible to me given we can't even fully simulate much simpler systems like waves on the beach at this level.
2. ASML production breakthrough: If AGI helped ASML scale from ~65 to 200+ EUV machines/year within 2 years, despite specialized component supplier constraints. This would demonstrate both that manufacturing bottlenecks in general are primarily organizational rather than physical and that EUVs would not prevent semiconductor manufacturing acceleration.
3. Chemical construction and manufacturing process acceleration: If AGI discovered techniques to achieve full-strength concrete in hours instead of weeks while maintaining structural integrity. This would show that "physical time constants" aren't as fundamental as I claim.
4. Alternative lithography breakthrough: If AGI discovered a completely different approach to nano-scale patterning that bypasses EUV entirely and can be manufactured at scale quickly. This would make ASML's production constraints irrelevant.
5. Historical analogues: Examples of previous automation breakthroughs that eliminated empirical iteration requirements in complex physical systems. If similar transitions happened in other industries (aerospace, pharmaceuticals, etc.), that would suggest I'm underestimating AGI's potential.
Current assessment: I don't see plausible pathways to these breakthroughs based on my understanding of the physical constraints, but I'd update significantly if presented with concrete evidence or compelling mechanistic arguments for any of them.
Implications
Manufacturing may provide a significant bottleneck on hardware-limited capability growth. This means observable leading indicators years before impact such as fab construction and equipment orders. It also means multiple opportunities for iterative technical safety work and political institutional adaptation. It also seems like it makesPaul Christiano's slow takeoff scenarios more plausible.
However, this does not rule out a fast takeoff scenario in under a year via a recursion loop as AGI might self-improve through other pathways that don't require more hardware. For example, discovering more compute-efficient algorithms, better training techniques that achieve more with same FLOP, and novel architectures that are qualitatively more capable with the same amount of compute.
AGI (defined as 10x speed on all digital white collar work), even with capable humanoid robots, would probably not accelerate cutting-edge chip fab production beyond 20-30% of current growth trends. The binding constraints are physical manufacturing time floors, fab construction, and equipment construction, not lack of engineering capability or planning intelligence.
The manufacturing floor teaches that intelligence cannot optimize away physical time constants. Chemical reactions take time. Equipment manufacturing has production rate limits. Empirical optimization requires physical iteration. But it remains an open question whether algorithmic improvements can make manufacturing constraints largely irrelevant.
Views and opinions expressed in this article are my own and do not reflect the official policy, opinions, or position of Samsung or its subsidiaries. I used Claude Sonnet 4.5 heavily in the research and writing of this piece, and have manually reviewed every line, as well as hand written most of the sections on humanoid robotics and growth constraints. Every source cited from Epoch AI I have read in full, but other sources were provided to me by Claude, which I double checked. Feedback is especially welcome from those with expertise in: (1) semiconductor manufacturing, (2) chip design, (3) AI algorithmic progress modeling, (4) AI R&D acceleration dynamics.
I work at a semiconductor fabrication facility in the USA for Samsung as an engineer developing and deploying a variety of AI models (traditional CV and LLMs) for manufacturing operations. This post analyzes whether AGI—defined as AI capable of all digital white collar work at 10x human speed—could meaningfully accelerate current semiconductor manufacturing growth trends with or without access to humanoid robots. My conclusion is that even in the highly unlikely case that humanoid robots are brought online tomorrow, chip manufacturing would have a 8.4-9.1% CAGR instead of a baseline 7% CAGR, an increase of 20-30% over the next 2-3 years. Same goes for advanced (<7nm) chips, increasing from 14% CAGR to 16.8-18.2% It seems like manufacturing is a strong blind spot for the AI safety community, and so I am hoping to add to the conversation with domain expertise.
Epistemic status: Moderately confident (70-80%) about the analysis of existing bottlenecks which could be addressed by AGI based on 3+ years building and deploying AI in semiconductor fabs. Much less confident about AGI capability assumptions and what they would do to chip production. The core claim that manufacturing growth rates would only see 20-30% uplift seems plausible but is ultimately a guess based on my personal experience and reading about current fab scaling times. I am likely wrong about specific timelines and lots of information needed to make a better prediction is proprietary to private corporations with a strong commercial interest in not making this information available.
Why Chip Production Growth Matters for AI Safety
What would it be worth epistemically if we could confidently rule out a contributing factor or sole path to runaway AI progress? It is important to separate chip production out from pathways like algorithmic progress, hardware design progress, and data scaling because it seems like it might be safely ruled out as a path (or even a factor contributing more than 1% to capabilities growth) for recursive self improvement. People in AI safety could reallocate their attention to potentially more fruitful areas of research such as identifying timelines related to algorithmic progress, work which Epoch AI has recently put out a call for and which I would like to highlight here.
How Would Semiconductor Manufacturing Growth Rates Change if AGI was Built Tomorrow?
If we achieved AGI tomorrow (AI doing all digital white collar work at 10x human speed), would it significantly accelerate cutting-edge AI chip fab production growth trends over the next decade? This is narrower than "will AGI cause fast takeoff?" but addresses one specific pathway: whether AGI could bootstrap itself by accelerating hardware manufacturing.
Why "All Digital White Collar Work at 10x Human Speed"?
I chose this definition because it represents a plausible order of magnitude estimate on what near-term AI systems might achieve and this order of magnitude of speed and intelligence often shows up in projects like AI 2027 as a first step towards dangerously accelerated timelines.
Semiconductor Manufacturing Primer
For readers unfamiliar with chip manufacturing, here are the key concepts:
What is a fab? A semiconductor fabrication facility ("fab") is where computer chips are manufactured. Building a cutting-edge fab costs $10-25 billion USD and takes 2-4 years. They're essentially the most complex factories humans build. They must be very clean, with a maximum of 10 or 100 particles 0.5 microns or larger in each cubic foot of air. They contain hundreds or thousands of specialized machines performing thousands of process steps on silicon wafers which make up the majority of the cost of the fab.
Why is ASML critical? ASML manufactures extreme ultraviolet (EUV) lithography machines, which are the only equipment capable of printing features small enough for cutting-edge chips (3nm, 5nm nodes). Each machine costs ~$300 million, weighs 180 tons, and ASML ships only ~65 of these units per quarter globally. Without EUV machines, you cannot manufacture cutting-edge AI chips. No other company makes them.
What is yield and why does it require physical testing? Yield is the percentage of chips that work correctly after manufacturing. A 70% yield means 30% of chips are defective and must be discarded. The challenge is that thousands of process steps interact in ways we don't fully understand. Defects can have multiple root causes—contamination, equipment drift, material variations, quantum effects at nanometer scale. Even with perfect AI analysis, you must run actual production wafers to collect sensor, scan, and test data to observe how defects manifest in practice.
Hypothesis: AGI Would not Significantly Accelerate Semiconductor Manufacturing
Even with AGI and a capacity to be embodied in humanoid robots, chip manufacturing growth rates may only increase by 20-30% in the next 2-3 years.
This is because:
How Much Could AGI Accelerate Growth Rates?
Baseline Growth Projections
SEMI projects global semiconductor capacity growing for the leading-edge (≤7nm) technology used in making AI chips by +14% CAGR.
What AGI Could Accelerate
High-impact areas:
Strong Physical and Operational Growth Constraints
These are the constraints that remain even with white collar AGI doing all engineering and planning work, but without humanoid robots:
1. ASML Equipment Manufacturing Bottleneck
2. Empirical Yield Optimization Requires Physical Wafer Runs
Why can't we just simulate yield instead of running physical wafers?
The fundamental challenge is the high volume of process steps with interactions we don't fully understand:
Even with perfect AI analysis, it is necessary to run actual production wafers to observe how defects manifest in practice. AGI can dramatically speed up the analysis and learning from each iteration, but it seems extremely unlikely to eliminate the need for physical iterations.
3. Physical and Chemical Time Constants
4. Construction Physical Work
What About Humanoid Robots?
Humanoid robots seems extremely unlikely to come about in the next ten years through normal human research (in my opinion, which is based largely on those of my robotics professor), but what if AGI somehow accelerated this progress and also came with capable humanoid robots? It seems that even with fully functional humanoid robots, AGI would not significantly help accelerate chip manufacturing for fundamental reasons:
Reliability gap: Fabs need as much uptime as possible when EUV machines cost $300M. One pathway AGI could affect this is by replacing equipment and process technicians with humanoid robots. Less specialized, more complex robotics introduce additional room for error, and the AGI would have to be more reliable than humans in performing physical tasks like maintenance on equipment or construction of fabs. What happens when the computer vision model mistakes two very similar seals and puts the wrong one in a piece of equipment? The answer is enormous losses, and in order for these humanoids to be adopted, they would have to undergo at least months of testing to demonstrate they are substantially more reliable than humans, which is a strong weakness of current transformer architectures. If humanoid robots proved to be perfectly reliable, they would still be closing a tiny gap in production. Fabs have extremely high reliability standards, so you might see something like a less than one percent increase in production due to reduced scrap and improved yield.
Existing automation is better: Fabs creating cutting edge semiconductor wafers for AI are already fully automated with the exception of human maintenance and repair on the equipment itself in the fab. Purpose-built systems are more reliable, don't contaminate cleanrooms, and don't need the flexibility humanoids provide. To make wafers all you need to do is repeat the same motions millions of times. This one is hard to measure and prove because of how close to its chest the industry plays its cards, but think about it this way: Would it be easier for 100 humans to make waffles with their hands, or a giant automated oven full of waffle irons with a single human sitting next to it who could fix it if something broke once every few months?
Technicians and other physical human labor are not the constraint: Current training timelines are fast and you could get people up to speed rapidly. Quick-start programs can take 8-10 weeks for entry-level roles, while certificate programs can last a year and it takes two years for an associate's degree for a full technician qualification. Most of these options are much faster than trying to manufacture millions of robots to scale semiconductor fab construction and operations. There is a workforce shortage: SIA projects that 39% of chip factory technician jobs may remain vacant by 2030, with McKinsey estimating a deficit of 90,000 skilled technicians. However, unlike other constraints on growth, this one is highly responsive to investment:
So while technician supply is currently constrained, it's not serious physical constraint or bottleneck. I think it's an organizational/educational scaling challenge that AGI could significantly help solve.
As a result, it seems that while having more technicians in the form of humanoid robots could be useful and potentially accelerate growth, because modern fabs are already fully autonomous outside of repairs, humanoid robots would not significantly uplift AGI's ability to accelerate growth. Given that the technology for humanoid robots following shortly after AGI is a already in question due to Moravec's paradox, it seems that humanoid robots would not be a significant part of the growth equation.
What Would Change My Mind
What evidence would make me update toward AGI significantly accelerating manufacturing?
1. Breakthrough in simulation accuracy: If AGI could accurately predict yield outcomes without physical wafer runs. This would require solving the "thousands of interacting process steps with quantum effects" problem. I think this would require essentially perfect physics simulation at the nano-scale. This doesn't seem plausible to me given we can't even fully simulate much simpler systems like waves on the beach at this level.
2. ASML production breakthrough: If AGI helped ASML scale from ~65 to 200+ EUV machines/year within 2 years, despite specialized component supplier constraints. This would demonstrate both that manufacturing bottlenecks in general are primarily organizational rather than physical and that EUVs would not prevent semiconductor manufacturing acceleration.
3. Chemical construction and manufacturing process acceleration: If AGI discovered techniques to achieve full-strength concrete in hours instead of weeks while maintaining structural integrity. This would show that "physical time constants" aren't as fundamental as I claim.
4. Alternative lithography breakthrough: If AGI discovered a completely different approach to nano-scale patterning that bypasses EUV entirely and can be manufactured at scale quickly. This would make ASML's production constraints irrelevant.
5. Historical analogues: Examples of previous automation breakthroughs that eliminated empirical iteration requirements in complex physical systems. If similar transitions happened in other industries (aerospace, pharmaceuticals, etc.), that would suggest I'm underestimating AGI's potential.
Current assessment: I don't see plausible pathways to these breakthroughs based on my understanding of the physical constraints, but I'd update significantly if presented with concrete evidence or compelling mechanistic arguments for any of them.
Implications
Manufacturing may provide a significant bottleneck on hardware-limited capability growth. This means observable leading indicators years before impact such as fab construction and equipment orders. It also means multiple opportunities for iterative technical safety work and political institutional adaptation. It also seems like it makes Paul Christiano's slow takeoff scenarios more plausible.
However, this does not rule out a fast takeoff scenario in under a year via a recursion loop as AGI might self-improve through other pathways that don't require more hardware. For example, discovering more compute-efficient algorithms, better training techniques that achieve more with same FLOP, and novel architectures that are qualitatively more capable with the same amount of compute.
Epoch AI's work on software intelligence explosion identifies this as a key crux.
Conclusion
AGI (defined as 10x speed on all digital white collar work), even with capable humanoid robots, would probably not accelerate cutting-edge chip fab production beyond 20-30% of current growth trends. The binding constraints are physical manufacturing time floors, fab construction, and equipment construction, not lack of engineering capability or planning intelligence.
The manufacturing floor teaches that intelligence cannot optimize away physical time constants. Chemical reactions take time. Equipment manufacturing has production rate limits. Empirical optimization requires physical iteration. But it remains an open question whether algorithmic improvements can make manufacturing constraints largely irrelevant.
Views and opinions expressed in this article are my own and do not reflect the official policy, opinions, or position of Samsung or its subsidiaries. I used Claude Sonnet 4.5 heavily in the research and writing of this piece, and have manually reviewed every line, as well as hand written most of the sections on humanoid robotics and growth constraints. Every source cited from Epoch AI I have read in full, but other sources were provided to me by Claude, which I double checked. Feedback is especially welcome from those with expertise in: (1) semiconductor manufacturing, (2) chip design, (3) AI algorithmic progress modeling, (4) AI R&D acceleration dynamics.