CThe Chats of OthersSubmit a conversation
Back to the archive

Summarize Situational Awareness

https://situational-awareness.ai/(https://situational-awareness.ai/) Sumarrize this Situational Awareness: The Decade Ahead — Summary <row align="start" gap={3} <AsyncImage query="Situational Awareness The Decade Ahea…

Human participant

Language model

Situational Awareness: The Decade Ahead — Summary

<row align="start" gap={3}> <AsyncImage query="Situational Awareness The Decade Ahead Leopold Aschenbrenner June 2024 essay cover" aspectRatio="3:4" maxWidth="132px"/> <box flex="1" gap={1}> **Author:** <Entity category="people" value="Leopold Aschenbrenner"/>, former OpenAI researcher
**Published:** June 2024

**Central prediction:** Artificial general intelligence (AGI) around 2027, followed by superintelligence potentially within a year.

**Central argument:** Most of the world dramatically underestimates how quickly AI will advance and how profoundly it will reshape economics, geopolitics, and national security.
</box> </row>

The essay is not primarily about whether AI will become important. It argues that we are approaching an intelligence explosion comparable in historical significance to the development of nuclear weapons, and that almost nobody is preparing adequately.

Aschenbrenner builds his argument in several stages.

1. AGI by 2027 is plausible

The first argument is that AI progress is surprisingly predictable if you extrapolate three trends:

  • Compute scaling: Training runs become roughly 3× larger each year.
  • Algorithmic improvements: Researchers continually discover ways to get more intelligence from the same compute.
  • Unhobbling: AI systems become more capable when given tools, memory, reasoning time, and the ability to act autonomously.

He estimates that compute and algorithmic improvements each contribute approximately 0.5 orders of magnitude annually.

His claim is that the jump from GPT-2 to GPT-4 took us from relatively primitive language abilities to something resembling a capable high-school student.

Extrapolate another three or four years, and AI systems could surpass most college graduates across many cognitive tasks.

His conclusion: AGI around 2027 is not an extraordinary prediction. It's a relatively straightforward extrapolation of existing trends.

2. AGI will trigger an intelligence explosion

This is arguably the most important part of the essay.

Aschenbrenner argues that AGI won't represent the endpoint of AI development. Instead, it will accelerate the development of even more powerful AI.

The reasoning is:

  1. We develop AI capable of performing AI research at approximately human-expert level.
  2. We deploy millions of copies of these AI researchers.
  3. They collectively accomplish years or decades of research in a fraction of the time.
  4. Their discoveries produce smarter AI systems.
  5. Those systems become even better AI researchers.

This creates a powerful feedback loop.

He suggests that a decade of algorithmic progress could be compressed into a single year, potentially producing systems vastly more intelligent than any human.

This is the transition from AGI to superintelligence.

The implication is that humanity might have very little time to adapt between developing human-level AI and developing something qualitatively beyond human comprehension.

3. Trillions of dollars will flow into compute

Aschenbrenner argues that AI development is becoming an industrial mobilization rather than merely a software engineering project.

Training and operating frontier systems will require extraordinary investments in:

  • GPU clusters costing hundreds of billions or even trillions of dollars.
  • New power plants and electricity transmission infrastructure.
  • Massive data centers.
  • Semiconductor manufacturing and supply chains.

He anticipates AI becoming one of the largest drivers of capital investment in the global economy.

The critical observation is that intelligence is becoming something you can manufacture by investing capital, electricity, and computing infrastructure.

That changes the economics of intelligence itself.

4. AI security is dangerously inadequate

He argues that frontier AI model weights and research breakthroughs will become some of the most valuable strategic assets on Earth.

Yet AI labs historically operated with security practices more appropriate for ordinary technology startups than for organizations developing potentially decisive military capabilities.

He is particularly concerned about Chinese intelligence operations stealing model weights and algorithmic secrets.

His conclusion is that frontier AI development will eventually require security approaching that of nuclear weapons programs.

5. Controlling superintelligence is unsolved

Even if the United States successfully develops superintelligence first, there's another problem: ensuring it does what humans actually want.

Aschenbrenner emphasizes that current alignment techniques are not proven to work on systems vastly smarter than humans.

A system that is smarter than its operators, capable of autonomous action, and able to improve itself presents unprecedented control problems.

The faster the intelligence explosion, the less time researchers will have to solve them.

He believes alignment may be solvable, but the transition could be extraordinarily dangerous.

6. AI becomes a geopolitical arms race

Aschenbrenner sees superintelligence as potentially delivering a decisive military, economic, and technological advantage.

Consequently, he expects the United States and China to enter an increasingly intense competition for AI supremacy.

His argument is that whichever country develops superintelligence first might gain an advantage so overwhelming that conventional balances of power cease to function.

This creates an uncomfortable tension:

  • Slowing AI development might improve safety.
  • Slowing development might also allow an adversarial government to develop superintelligence first.

He strongly favors maintaining American technological leadership while improving security and alignment.

7. Eventually, the government takes over

His final major prediction is that frontier AI development becomes too strategically important to remain an ordinary private-sector activity.

He anticipates something resembling a modern Manhattan Project around 2027–2028.

The US government would become deeply involved in securing, financing, and potentially directing frontier AI development.

He calls this "The Project."

In his scenario, the decisive stages of superintelligence development might happen inside highly classified facilities rather than Silicon Valley startups.


Language model

Situational Awareness: The Decade Ahead — Summary

<row align="start" gap={3}> <AsyncImage query="Situational Awareness The Decade Ahead Leopold Aschenbrenner June 2024 essay cover" aspectRatio="3:4" maxWidth="132px"/> <box flex="1" gap={1}> **Author:** <Entity category="people" value="Leopold Aschenbrenner"/>, former OpenAI researcher
**Published:** June 2024

**Central prediction:** Artificial general intelligence (AGI) around 2027, followed by superintelligence potentially within a year.

**Central argument:** Most of the world dramatically underestimates how quickly AI will advance and how profoundly it will reshape economics, geopolitics, and national security.
</box> </row>

The essay is not primarily about whether AI will become important. It argues that we are approaching an intelligence explosion comparable in historical significance to the development of nuclear weapons, and that almost nobody is preparing adequately.

Aschenbrenner builds his argument in several stages.

1. AGI by 2027 is plausible

The first argument is that AI progress is surprisingly predictable if you extrapolate three trends:

  • Compute scaling: Training runs become roughly 3× larger each year.
  • Algorithmic improvements: Researchers continually discover ways to get more intelligence from the same compute.
  • Unhobbling: AI systems become more capable when given tools, memory, reasoning time, and the ability to act autonomously.

He estimates that compute and algorithmic improvements each contribute approximately 0.5 orders of magnitude annually.

His claim is that the jump from GPT-2 to GPT-4 took us from relatively primitive language abilities to something resembling a capable high-school student.

Extrapolate another three or four years, and AI systems could surpass most college graduates across many cognitive tasks.

His conclusion: AGI around 2027 is not an extraordinary prediction. It's a relatively straightforward extrapolation of existing trends.

2. AGI will trigger an intelligence explosion

This is arguably the most important part of the essay.

Aschenbrenner argues that AGI won't represent the endpoint of AI development. Instead, it will accelerate the development of even more powerful AI.

The reasoning is:

  1. We develop AI capable of performing AI research at approximately human-expert level.
  2. We deploy millions of copies of these AI researchers.
  3. They collectively accomplish years or decades of research in a fraction of the time.
  4. Their discoveries produce smarter AI systems.
  5. Those systems become even better AI researchers.

This creates a powerful feedback loop.

He suggests that a decade of algorithmic progress could be compressed into a single year, potentially producing systems vastly more intelligent than any human.

This is the transition from AGI to superintelligence.

The implication is that humanity might have very little time to adapt between developing human-level AI and developing something qualitatively beyond human comprehension.

3. Trillions of dollars will flow into compute

Aschenbrenner argues that AI development is becoming an industrial mobilization rather than merely a software engineering project.

Training and operating frontier systems will require extraordinary investments in:

  • GPU clusters costing hundreds of billions or even trillions of dollars.
  • New power plants and electricity transmission infrastructure.
  • Massive data centers.
  • Semiconductor manufacturing and supply chains.

He anticipates AI becoming one of the largest drivers of capital investment in the global economy.

The critical observation is that intelligence is becoming something you can manufacture by investing capital, electricity, and computing infrastructure.

That changes the economics of intelligence itself.

4. AI security is dangerously inadequate

He argues that frontier AI model weights and research breakthroughs will become some of the most valuable strategic assets on Earth.

Yet AI labs historically operated with security practices more appropriate for ordinary technology startups than for organizations developing potentially decisive military capabilities.

He is particularly concerned about Chinese intelligence operations stealing model weights and algorithmic secrets.

His conclusion is that frontier AI development will eventually require security approaching that of nuclear weapons programs.

5. Controlling superintelligence is unsolved

Even if the United States successfully develops superintelligence first, there's another problem: ensuring it does what humans actually want.

Aschenbrenner emphasizes that current alignment techniques are not proven to work on systems vastly smarter than humans.

A system that is smarter than its operators, capable of autonomous action, and able to improve itself presents unprecedented control problems.

The faster the intelligence explosion, the less time researchers will have to solve them.

He believes alignment may be solvable, but the transition could be extraordinarily dangerous.

6. AI becomes a geopolitical arms race

Aschenbrenner sees superintelligence as potentially delivering a decisive military, economic, and technological advantage.

Consequently, he expects the United States and China to enter an increasingly intense competition for AI supremacy.

His argument is that whichever country develops superintelligence first might gain an advantage so overwhelming that conventional balances of power cease to function.

This creates an uncomfortable tension:

  • Slowing AI development might improve safety.
  • Slowing development might also allow an adversarial government to develop superintelligence first.

He strongly favors maintaining American technological leadership while improving security and alignment.

7. Eventually, the government takes over

His final major prediction is that frontier AI development becomes too strategically important to remain an ordinary private-sector activity.

He anticipates something resembling a modern Manhattan Project around 2027–2028.

The US government would become deeply involved in securing, financing, and potentially directing frontier AI development.

He calls this "The Project."

In his scenario, the decisive stages of superintelligence development might happen inside highly classified facilities rather than Silicon Valley startups.


How convincing is the argument?

I think it's useful to distinguish the essay's underlying observations from its predictions.

ClaimMy assessment
AI capabilities will continue improving rapidlyStrong argument
AI agents will automate substantial cognitive workStrong argument
Compute investment will become enormousStrong argument
AI will accelerate AI researchPlausible, potentially transformative
AGI arrives by 2027Plausible but highly uncertain
AGI produces superintelligence within a yearSpeculative
AI creates a decisive geopolitical advantagePlausible, but the magnitude is uncertain
The US establishes a classified Manhattan Project for AISpeculative

The biggest weakness is that the essay treats several uncertain developments as a connected sequence.

For example, automating AI research doesn't necessarily imply that progress accelerates without limit. Research could encounter bottlenecks in experiments, compute availability, validation, or fundamental scientific understanding.

Similarly, exponential improvements on benchmarks don't necessarily translate into equally rapid improvements in reliable, autonomous real-world work.

These are substantial uncertainties, not minor objections.

What I find most interesting

The deepest idea in the essay isn't actually the 2027 prediction.

It's that intelligence might cease to be a fundamentally scarce human resource.

Throughout economic history, technological progress has made physical labor, manufacturing, transportation, communication, and computation dramatically cheaper.

But human cognitive labor has remained relatively expensive.

Aschenbrenner is arguing that we're approaching the point where cognitive labor becomes something that can be scaled almost arbitrarily through computation.

If he's right, this changes the nature of economic growth itself.

It also has a particularly interesting implication for software businesses like Pickaxe and Fork: the scarce resource may increasingly shift away from building software and toward identifying valuable problems, understanding customers, distributing solutions, and managing trust.

That's my extrapolation, rather than an explicit thesis of the essay.

The entire argument in one sentence: We're not merely building better software; we're potentially building an industrial process for manufacturing intelligence, and the resulting economic and geopolitical transformation could happen much faster than our institutions can adapt.

The original series is worth reading, especially Part I: Counting the OOMs and Part II: The Intelligence Explosion. Those contain the core technical argument; the later sections explore its strategic consequences. <Cite refs={["turn300217view0","turn300217view1","turn300217view2"]}/>

Language model

How convincing is the argument?

I think it's useful to distinguish the essay's underlying observations from its predictions.

ClaimMy assessment
AI capabilities will continue improving rapidlyStrong argument
AI agents will automate substantial cognitive workStrong argument
Compute investment will become enormousStrong argument
AI will accelerate AI researchPlausible, potentially transformative
AGI arrives by 2027Plausible but highly uncertain
AGI produces superintelligence within a yearSpeculative
AI creates a decisive geopolitical advantagePlausible, but the magnitude is uncertain
The US establishes a classified Manhattan Project for AISpeculative

The biggest weakness is that the essay treats several uncertain developments as a connected sequence.

For example, automating AI research doesn't necessarily imply that progress accelerates without limit. Research could encounter bottlenecks in experiments, compute availability, validation, or fundamental scientific understanding.

Similarly, exponential improvements on benchmarks don't necessarily translate into equally rapid improvements in reliable, autonomous real-world work.

These are substantial uncertainties, not minor objections.

What I find most interesting

The deepest idea in the essay isn't actually the 2027 prediction.

It's that intelligence might cease to be a fundamentally scarce human resource.

Throughout economic history, technological progress has made physical labor, manufacturing, transportation, communication, and computation dramatically cheaper.

But human cognitive labor has remained relatively expensive.

Aschenbrenner is arguing that we're approaching the point where cognitive labor becomes something that can be scaled almost arbitrarily through computation.

If he's right, this changes the nature of economic growth itself.

It also has a particularly interesting implication for software businesses like Pickaxe and Fork: the scarce resource may increasingly shift away from building software and toward identifying valuable problems, understanding customers, distributing solutions, and managing trust.

That's my extrapolation, rather than an explicit thesis of the essay.

The entire argument in one sentence: We're not merely building better software; we're potentially building an industrial process for manufacturing intelligence, and the resulting economic and geopolitical transformation could happen much faster than our institutions can adapt.

The original series is worth reading, especially Part I: Counting the OOMs and Part II: The Intelligence Explosion. Those contain the core technical argument; the later sections explore its strategic consequences. <Cite refs={["turn300217view0","turn300217view1","turn300217view2"]}/>

Human participant

My thought is, compute limitations, coupled with data limitations, have already put a ceiling on some kind of deep intrinsic raw intelligence increase in the models. Maybe not a ceiling, but progress is logarithmic.

Recent developments and benchmark improvements, as well as perceived advances in AI model quality, are mostly due to use case specific training, figuring out better ways to apply models of around the same intelligence, better harnesses, etc. There's still a HUGE amount of untapped potential in these efforts, to apply the existing intelligence level of AI across the economy. But Superintelligence is further away than most people think.

What would you say about my thoughts, and what do you think Leopold would say?

Language model

I think your hypothesis is quite plausible, and more importantly, it identifies what I consider the central unresolved question in AI today:

Are we still making fundamental progress toward greater intelligence, or are we primarily getting better at extracting useful work from an intelligence level we've already achieved?

Those are two very different technological trajectories, even though they can look almost identical from the perspective of someone using ChatGPT or Claude.

I think Leopold Aschenbrenner would disagree with your conclusion, but he'd probably agree with more of your observations than you might expect.

1. Your argument is stronger than simply saying AI is hitting a wall

I'd break your hypothesis into three claims.

First: Pretraining is experiencing diminishing returns.

The original paradigm was to take a transformer, feed it more data, increase its parameter count and compute budget, and watch its general capabilities improve.

That process has constraints. High-quality human-generated training data is finite. Compute is expensive. And each incremental improvement requires disproportionately greater investment.

This doesn't mean scaling has stopped working. But the economic cost of achieving another substantial improvement may be growing much faster than the value of that improvement.

Second: A lot of apparent intelligence gains are actually systems-engineering gains.

Consider the difference between a model answering a programming question and an agent that can inspect a repository, edit files, run tests, identify errors, and try again.

The second system can accomplish enormously more, even if its underlying model isn't substantially smarter.

The same applies to reasoning-time compute, reinforcement learning on verifiable tasks, specialized post-training, better context management, memory, and tool use.

Some of these methods genuinely improve the model's capabilities. Others improve how effectively existing capabilities are deployed. The distinction matters.

Third: Economic transformation doesn't require superintelligence.

This is where I think your argument is particularly compelling.

Imagine that underlying AI intelligence stopped improving tomorrow, but engineers spent another decade making existing models reliable, inexpensive, well-integrated, and capable of performing useful work.

The economic consequences could still be extraordinary.

We might be sitting on an enormous amount of unrealized technological potential without being particularly close to superintelligence.

That's a coherent alternative to Leopold's intelligence-explosion scenario.

Language model

I think your hypothesis is quite plausible, and more importantly, it identifies what I consider the central unresolved question in AI today:

Are we still making fundamental progress toward greater intelligence, or are we primarily getting better at extracting useful work from an intelligence level we've already achieved?

Those are two very different technological trajectories, even though they can look almost identical from the perspective of someone using ChatGPT or Claude.

I think Leopold Aschenbrenner would disagree with your conclusion, but he'd probably agree with more of your observations than you might expect.

1. Your argument is stronger than simply saying AI is hitting a wall

I'd break your hypothesis into three claims.

First: Pretraining is experiencing diminishing returns.

The original paradigm was to take a transformer, feed it more data, increase its parameter count and compute budget, and watch its general capabilities improve.

That process has constraints. High-quality human-generated training data is finite. Compute is expensive. And each incremental improvement requires disproportionately greater investment.

This doesn't mean scaling has stopped working. But the economic cost of achieving another substantial improvement may be growing much faster than the value of that improvement.

Second: A lot of apparent intelligence gains are actually systems-engineering gains.

Consider the difference between a model answering a programming question and an agent that can inspect a repository, edit files, run tests, identify errors, and try again.

The second system can accomplish enormously more, even if its underlying model isn't substantially smarter.

The same applies to reasoning-time compute, reinforcement learning on verifiable tasks, specialized post-training, better context management, memory, and tool use.

Some of these methods genuinely improve the model's capabilities. Others improve how effectively existing capabilities are deployed. The distinction matters.

Third: Economic transformation doesn't require superintelligence.

This is where I think your argument is particularly compelling.

Imagine that underlying AI intelligence stopped improving tomorrow, but engineers spent another decade making existing models reliable, inexpensive, well-integrated, and capable of performing useful work.

The economic consequences could still be extraordinary.

We might be sitting on an enormous amount of unrealized technological potential without being particularly close to superintelligence.

That's a coherent alternative to Leopold's intelligence-explosion scenario.

2. Where I think Leopold would push back

Leopold actually anticipated your argument. His essay explicitly distinguishes improvements in raw model capabilities from what he calls unhobbling: turning a limited chatbot into a powerful autonomous system.

But he'd challenge the distinction you're drawing between better applications of intelligence and better intelligence itself.

I imagine his response would go something like this:

You're observing that we're making progress through better reasoning, tools, reinforcement learning, and agent scaffolding rather than simply training bigger models. But why should that make us less optimistic about AGI? Those improvements are precisely what I predicted would drive the next generation of capabilities.

That would be a substantial rebuttal.

He'd also point out that data scarcity doesn't necessarily constrain AI research as much as it constrains conventional pretraining.

An AI can learn through experimentation, simulation, reinforcement learning, and feedback from the environment. It doesn't necessarily need another trillion tokens of human-written text.

And most importantly, he'd argue that the decisive question is not whether we can keep scaling pretraining. It's whether we can automate AI research itself.

Once we can, the mechanism of progress fundamentally changes.

His essay explicitly acknowledges that experimental compute could constrain an intelligence explosion. He just believes automated researchers could use that compute dramatically more efficiently. <Cite ref="turn611271search1"/>

3. Where I think your argument is stronger

There are three important objections to Leopold's position.

A. Intelligence may not be a single quantity that can be scaled

I think the phrase "raw intelligence" captures something real, but it's difficult to define.

We can measure capabilities on particular tasks. We can measure how well models generalize to unfamiliar problems. But there's no established single numerical measure of general intelligence.

Suppose a model becomes dramatically better at mathematics through reinforcement learning, but remains equally bad at understanding ambiguous social situations or making reliable long-term plans.

Has it become more intelligent?

In one meaningful sense, yes. In another, its generality hasn't improved very much.

This matters because Leopold's extrapolation assumes that progress in measurable capabilities translates into progress toward broadly superhuman intelligence.

That isn't guaranteed.

B. Diminishing returns could dominate recursive improvement

This is the strongest technical objection, in my view.

Imagine an AI research system that can improve the efficiency of its successor by 20%.

Its successor might discover another 15% improvement, then 10%, then 5%.

You get a series of genuine improvements, but not an explosion.

The key mathematical question is whether improved intelligence produces enough additional research productivity to overcome diminishing returns.

Recursive self-improvement does not automatically imply accelerating progress.

You can have a positive feedback loop whose gains diminish until it converges.

C. Benchmarks increasingly measure optimization toward benchmarks

There's an important difference between:

  • A model developing a more powerful general ability to reason.
  • A model being trained to perform better on mathematical proofs, coding tests, or structured reasoning problems.

These overlap, but they're not identical.

The more training focuses on specific measurable outcomes, the more cautious we should be about interpreting benchmark improvements as evidence of broadly increasing intelligence.

This doesn't mean the gains are fake. Coding ability, for example, is economically valuable even when specialized.

But it does mean that improvements on familiar benchmarks might overstate progress toward general superintelligence.

Recent research supports taking both sides seriously: a 2026 OECD report finds no decisive evidence that scaling has stopped working, while explicitly acknowledging that scaling gains could plateau or become less economically meaningful. Stanford's 2026 AI Index likewise identifies high-quality training-data scarcity as a serious unresolved constraint. <Cite refs={["turn611271search7","turn611271search8"]}/>

4. What would actually distinguish your hypothesis from Leopold's?

I'd watch four indicators over the next two to three years.

IndicatorSupports your hypothesisSupports Leopold's
General reasoningImprovements concentrated in trained domainsLarge gains on genuinely unfamiliar problems
Frontier pretrainingLarger runs deliver marginal improvementsLarger runs continue producing major general capabilities
AI research automationAI excels at implementation but struggles with original discoveriesAI independently produces and validates major algorithmic breakthroughs
Agent performanceReliability improves mainly through specialized engineeringAgents become broadly competent across novel, long-horizon tasks

The third indicator is the most important.

In fact, a September 2026 paper by researchers including Geoffrey Hinton, Yoshua Bengio, and Jakub Pachocki argues that AI research automation is progressing quickly enough to make an intelligence explosion a serious possibility, while emphasizing considerable uncertainty. <Cite ref="turn611271academia9"/>

That doesn't establish Leopold's conclusion. But it does identify the empirical question on which his forecast increasingly depends.

5. My own forecast

I lean toward a hybrid of your position and Leopold's.

I suspect we're going to see three partially independent curves:

<box gap={2}> <text weight="medium" size="sm">Illustrative trajectories, not measured data</text> <Chart content={{"chartType":"line","meta":{"footer":"Conceptual comparison of possible trajectories, not empirical estimates."},"xKey":"year","series":[{"dataKey":"raw","label":"Underlying general intelligence"},{"dataKey":"systems","label":"Effective agent capabilities"},{"dataKey":"adoption","label":"Economic deployment"}],"data":[{"year":"2024","raw":20,"systems":8,"adoption":3},{"year":"2025","raw":31,"systems":17,"adoption":6},{"year":"2026","raw":40,"systems":32,"adoption":12},{"year":"2027","raw":47,"systems":49,"adoption":22},{"year":"2028","raw":52,"systems":65,"adoption":36},{"year":"2029","raw":56,"systems":77,"adoption":53},{"year":"2030","raw":59,"systems":85,"adoption":70}]}}/> <caption>Relative, arbitrary indices to illustrate how capabilities and deployment could keep accelerating even if foundational intelligence improvements slow.</caption> </box>

I think it's quite possible that the underlying general intelligence curve bends downward in growth rate, while the effective capabilities of AI systems continue improving enormously.

And economic deployment could lag both by years.

But here's where I'd disagree with your conclusion that superintelligence is necessarily further away than most people think.

A sufficiently sophisticated system built from individually limited models could itself be superintelligent in meaningful respects.

Imagine a system comprising 100,000 AI agents with persistent memory, specialized skills, shared knowledge, experimental infrastructure, and the ability to coordinate their work.

Even if each individual agent were only as generally intelligent as today's frontier models, the overall system might outperform humanity's best research organizations in certain fields.

Whether that constitutes superintelligence depends on the definition. But its economic and strategic consequences might be indistinguishable from those of a much more capable individual model.

This is why I wouldn't equate slowing pretraining progress with a distant arrival of superintelligent systems.

6. The deeper disagreement

I think your argument and Leopold's differ in what each treats as the scarce resource.

Your implicit model: Intelligence is an expensive foundational capability. We've already developed an extraordinary amount of it, and the next great opportunity is figuring out how to use it effectively.

Leopold's model: Intelligence is a technology that can improve the process of producing more intelligence. Once that feedback loop closes, the limitations that currently seem binding could rapidly disappear.

I think both models are reasonable. But Leopold's requires an additional, very consequential assumption: that automating AI research will accelerate the discovery of fundamental improvements faster than diminishing returns and physical bottlenecks slow it down.

That remains unproven.

My biggest disagreement with your argument is that I wouldn't describe post-training and reasoning improvements as merely better applications of existing intelligence. Some are genuine improvements to the model's ability to solve problems it couldn't previously solve.

My biggest disagreement with Leopold is his confidence that these improvements will compound into an intelligence explosion on such a short timeline.

If I had to make one prediction: By 2030, AI will probably be vastly more economically useful than it is today, but the extent to which that improvement comes from fundamentally more intelligent models versus sophisticated systems built around them will remain a central question.

And I think there's a particularly provocative possibility embedded in your thesis: the AI revolution could be much bigger than people expect, while the arrival of superintelligence is much slower than people expect.

Those two outcomes aren't contradictory at all. <MemoryCite />

Language model

2. Where I think Leopold would push back

Leopold actually anticipated your argument. His essay explicitly distinguishes improvements in raw model capabilities from what he calls unhobbling: turning a limited chatbot into a powerful autonomous system.

But he'd challenge the distinction you're drawing between better applications of intelligence and better intelligence itself.

I imagine his response would go something like this:

You're observing that we're making progress through better reasoning, tools, reinforcement learning, and agent scaffolding rather than simply training bigger models. But why should that make us less optimistic about AGI? Those improvements are precisely what I predicted would drive the next generation of capabilities.

That would be a substantial rebuttal.

He'd also point out that data scarcity doesn't necessarily constrain AI research as much as it constrains conventional pretraining.

An AI can learn through experimentation, simulation, reinforcement learning, and feedback from the environment. It doesn't necessarily need another trillion tokens of human-written text.

And most importantly, he'd argue that the decisive question is not whether we can keep scaling pretraining. It's whether we can automate AI research itself.

Once we can, the mechanism of progress fundamentally changes.

His essay explicitly acknowledges that experimental compute could constrain an intelligence explosion. He just believes automated researchers could use that compute dramatically more efficiently. <Cite ref="turn611271search1"/>

3. Where I think your argument is stronger

There are three important objections to Leopold's position.

A. Intelligence may not be a single quantity that can be scaled

I think the phrase "raw intelligence" captures something real, but it's difficult to define.

We can measure capabilities on particular tasks. We can measure how well models generalize to unfamiliar problems. But there's no established single numerical measure of general intelligence.

Suppose a model becomes dramatically better at mathematics through reinforcement learning, but remains equally bad at understanding ambiguous social situations or making reliable long-term plans.

Has it become more intelligent?

In one meaningful sense, yes. In another, its generality hasn't improved very much.

This matters because Leopold's extrapolation assumes that progress in measurable capabilities translates into progress toward broadly superhuman intelligence.

That isn't guaranteed.

B. Diminishing returns could dominate recursive improvement

This is the strongest technical objection, in my view.

Imagine an AI research system that can improve the efficiency of its successor by 20%.

Its successor might discover another 15% improvement, then 10%, then 5%.

You get a series of genuine improvements, but not an explosion.

The key mathematical question is whether improved intelligence produces enough additional research productivity to overcome diminishing returns.

Recursive self-improvement does not automatically imply accelerating progress.

You can have a positive feedback loop whose gains diminish until it converges.

C. Benchmarks increasingly measure optimization toward benchmarks

There's an important difference between:

  • A model developing a more powerful general ability to reason.
  • A model being trained to perform better on mathematical proofs, coding tests, or structured reasoning problems.

These overlap, but they're not identical.

The more training focuses on specific measurable outcomes, the more cautious we should be about interpreting benchmark improvements as evidence of broadly increasing intelligence.

This doesn't mean the gains are fake. Coding ability, for example, is economically valuable even when specialized.

But it does mean that improvements on familiar benchmarks might overstate progress toward general superintelligence.

Recent research supports taking both sides seriously: a 2026 OECD report finds no decisive evidence that scaling has stopped working, while explicitly acknowledging that scaling gains could plateau or become less economically meaningful. Stanford's 2026 AI Index likewise identifies high-quality training-data scarcity as a serious unresolved constraint. <Cite refs={["turn611271search7","turn611271search8"]}/>

4. What would actually distinguish your hypothesis from Leopold's?

I'd watch four indicators over the next two to three years.

IndicatorSupports your hypothesisSupports Leopold's
General reasoningImprovements concentrated in trained domainsLarge gains on genuinely unfamiliar problems
Frontier pretrainingLarger runs deliver marginal improvementsLarger runs continue producing major general capabilities
AI research automationAI excels at implementation but struggles with original discoveriesAI independently produces and validates major algorithmic breakthroughs
Agent performanceReliability improves mainly through specialized engineeringAgents become broadly competent across novel, long-horizon tasks

The third indicator is the most important.

In fact, a September 2026 paper by researchers including Geoffrey Hinton, Yoshua Bengio, and Jakub Pachocki argues that AI research automation is progressing quickly enough to make an intelligence explosion a serious possibility, while emphasizing considerable uncertainty. <Cite ref="turn611271academia9"/>

That doesn't establish Leopold's conclusion. But it does identify the empirical question on which his forecast increasingly depends.

5. My own forecast

I lean toward a hybrid of your position and Leopold's.

I suspect we're going to see three partially independent curves:

<box gap={2}> <text weight="medium" size="sm">Illustrative trajectories, not measured data</text> <Chart content={{"chartType":"line","meta":{"footer":"Conceptual comparison of possible trajectories, not empirical estimates."},"xKey":"year","series":[{"dataKey":"raw","label":"Underlying general intelligence"},{"dataKey":"systems","label":"Effective agent capabilities"},{"dataKey":"adoption","label":"Economic deployment"}],"data":[{"year":"2024","raw":20,"systems":8,"adoption":3},{"year":"2025","raw":31,"systems":17,"adoption":6},{"year":"2026","raw":40,"systems":32,"adoption":12},{"year":"2027","raw":47,"systems":49,"adoption":22},{"year":"2028","raw":52,"systems":65,"adoption":36},{"year":"2029","raw":56,"systems":77,"adoption":53},{"year":"2030","raw":59,"systems":85,"adoption":70}]}}/> <caption>Relative, arbitrary indices to illustrate how capabilities and deployment could keep accelerating even if foundational intelligence improvements slow.</caption> </box>

I think it's quite possible that the underlying general intelligence curve bends downward in growth rate, while the effective capabilities of AI systems continue improving enormously.

And economic deployment could lag both by years.

But here's where I'd disagree with your conclusion that superintelligence is necessarily further away than most people think.

A sufficiently sophisticated system built from individually limited models could itself be superintelligent in meaningful respects.

Imagine a system comprising 100,000 AI agents with persistent memory, specialized skills, shared knowledge, experimental infrastructure, and the ability to coordinate their work.

Even if each individual agent were only as generally intelligent as today's frontier models, the overall system might outperform humanity's best research organizations in certain fields.

Whether that constitutes superintelligence depends on the definition. But its economic and strategic consequences might be indistinguishable from those of a much more capable individual model.

This is why I wouldn't equate slowing pretraining progress with a distant arrival of superintelligent systems.

6. The deeper disagreement

I think your argument and Leopold's differ in what each treats as the scarce resource.

Your implicit model: Intelligence is an expensive foundational capability. We've already developed an extraordinary amount of it, and the next great opportunity is figuring out how to use it effectively.

Leopold's model: Intelligence is a technology that can improve the process of producing more intelligence. Once that feedback loop closes, the limitations that currently seem binding could rapidly disappear.

I think both models are reasonable. But Leopold's requires an additional, very consequential assumption: that automating AI research will accelerate the discovery of fundamental improvements faster than diminishing returns and physical bottlenecks slow it down.

That remains unproven.

My biggest disagreement with your argument is that I wouldn't describe post-training and reasoning improvements as merely better applications of existing intelligence. Some are genuine improvements to the model's ability to solve problems it couldn't previously solve.

My biggest disagreement with Leopold is his confidence that these improvements will compound into an intelligence explosion on such a short timeline.

If I had to make one prediction: By 2030, AI will probably be vastly more economically useful than it is today, but the extent to which that improvement comes from fundamentally more intelligent models versus sophisticated systems built around them will remain a central question.

And I think there's a particularly provocative possibility embedded in your thesis: the AI revolution could be much bigger than people expect, while the arrival of superintelligence is much slower than people expect.

Those two outcomes aren't contradictory at all. <MemoryCite />

Human participant

My argument is to some extent used to justify my work, implementing AI as it exists today (with expectation of some improvement, to be sure) into the economy. With the goal of helping folks take better advantage of it, etc.

In an environment where we truly believed there was an inflection point, by some measures I should quit my job now. I should seek to work in another field, ideally energy, medicine, law, or physical compute, where I could be on the receiving end of self improving AI.

Language model

I think you've identified a genuine strategic tension, and one that goes considerably deeper than whether Fork is a good startup idea.

If you genuinely believe superintelligence is imminent, why would you spend the next five years building software that helps ordinary businesses use today's AI?

Why not position yourself to own something scarce and valuable in the world after intelligence becomes abundant?

I think that's a very serious question. And I think your reasoning is mostly correct, with one important qualification: the optimal strategy depends less on when superintelligence arrives than on what remains scarce after it arrives.

1. There are really two different bets

Your current work is implicitly betting on a prolonged period during which AI capabilities exceed society's ability to deploy them.

Imagine that today's frontier models are already capable of doing 50% of economically valuable cognitive work, but only 2% of that work has actually been automated.

Those percentages are illustrative, but the idea is important.

You don't need AI to become more intelligent to have an enormous business opportunity. You just need to close the gap between technological capability and economic adoption.

That's essentially the thesis behind Fork.

But Leopold's worldview suggests a different opportunity: if AI is about to become dramatically more intelligent, the greatest value might accrue to whoever controls resources that AI cannot manufacture simply by thinking harder.

These might include electricity generation, land, mineral rights, semiconductor manufacturing capacity, regulatory permissions, customer relationships, or physical infrastructure.

And there's a potentially devastating implication for an AI implementation company:

What happens when AI becomes so capable that customers can implement AI themselves?

If a customer can tell an AI, "Build me a CRM, integrate it with my business, maintain it, and improve it," why would they need Fork?

That's not a hypothetical I'd dismiss.

2. I would distinguish four possible futures

<box gap={2}> <grid columns={2} gap={2}> <grid-item> <box border radius="lg" padding={3} gap={2}> <text weight="medium">A. Slow intelligence, slow adoption</text> <text color="secondary" size="xs">Gradual technological change</text> AI becomes incrementally better, but organizations adopt it cautiously. Implementation companies have a long runway, though growth may be modest. </box> </grid-item> <grid-item> <box border radius="lg" padding={3} gap={2}> <text weight="medium">B. Slow intelligence, fast adoption</text> <text color="secondary" size="xs">Your core thesis</text> Existing models are already tremendously powerful. Distribution, integration, and operational reliability become major sources of value. </box> </grid-item> <grid-item> <box border radius="lg" padding={3} gap={2}> <text weight="medium">C. Fast intelligence, slow adoption</text> <text color="secondary" size="xs">Institutional bottlenecks dominate</text> Superhuman AI exists, but regulation, trust, infrastructure, and organizational inertia delay its economic impact. Implementation remains valuable. </box> </grid-item> <grid-item> <box border radius="lg" padding={3} gap={2}> <text weight="medium">D. Fast intelligence, fast adoption</text> <text color="secondary" size="xs">Leopold's disruptive scenario</text> AI can design, deploy, and manage software independently. Generic implementation becomes commoditized. Scarce assets and control of real-world systems become much more important. </box> </grid-item> </grid> </box>

Here's the interesting thing: Fork can be a good business in three of these four scenarios.

The fourth is the one that should worry you.

And even in that scenario, owning an established distribution channel, customer base, and deployment infrastructure might be valuable. It depends on whether customers continue to need an intermediary.

3. Your instinct about energy, medicine, law, and physical compute is directionally right

But I'd distinguish the sectors from the actual scarce assets within them.

SectorWhat AI makes abundantWhat might remain scarce
EnergyEngineering, modeling, optimizationGeneration capacity, grid interconnection, permits, land
MedicineDiagnostic reasoning, drug discovery assistanceClinical trials, regulatory approval, patients, physical care
LawResearch, drafting, legal analysisLicensure, legal authority, client trust, representation
ComputeChip design, infrastructure softwareFabrication capacity, electricity, capital, supply chains
Enterprise softwareCode, implementation, maintenanceDistribution, proprietary workflows, customer relationships, operational trust

The crucial distinction is that working in a sector doesn't mean owning its bottlenecks.

Becoming an employee at an energy company doesn't necessarily give you much exposure to the value of energy infrastructure.

Similarly, starting an AI-for-medicine company doesn't automatically position you to capture the benefits of AI-driven drug discovery.

You'd want to own equity, intellectual property, licenses, infrastructure, or relationships that become more valuable as AI capabilities increase.

And I think this points to a third strategy that sits between abandoning Fork and continuing it unchanged.

4. You could use Fork to acquire the assets that become valuable in an AI-abundant world

Right now, you're approaching Fork as infrastructure for people who want to deploy and sell AI-powered software.

Suppose you eventually have 5,000 agencies using Fork to build and maintain software for 100,000 small businesses.

You would potentially have something very valuable even if AI can write arbitrary software for free:

  • Established customer distribution.
  • Relationships with agencies that businesses already trust.
  • A deployment and maintenance platform embedded in real business operations.
  • Institutional knowledge about how businesses operate.
  • Recurring billing and operational dependencies.

Of course, those assets only become defensible if customers actually depend on your service, and not merely on the code it generates.

But consider the analogy to Shopify.

Shopify's long-term value isn't primarily that building an online store is technically difficult. It's that the company operates a substantial part of the infrastructure and ecosystem through which merchants conduct business.

If AI makes writing e-commerce software nearly free, that doesn't necessarily make Shopify worthless.

It might actually make Shopify more powerful, provided merchants still benefit from its payments, integrations, distribution, and operational infrastructure.

Fork could theoretically occupy a similar position for small-business software.

The question is whether you're building a tool that helps people create software, or a platform that becomes part of how businesses operate.

Those are very different companies.

5. There's an important problem with trying to time the intelligence explosion

Imagine you quit Pickaxe tomorrow and started an energy infrastructure company.

You might spend three years securing permits, raising capital, finding sites, negotiating interconnection agreements, and building a team.

If superintelligence arrives in 2028, you could be spectacularly positioned.

Or you could discover that superintelligence has made your particular business model obsolete, that large incumbents captured the value, or that the bottleneck you anticipated wasn't actually scarce.

Alternatively, suppose superintelligence doesn't arrive until 2040.

You might have abandoned a business opportunity that was accessible to you today in pursuit of a much more speculative one.

There's an asymmetry here: you have a functioning company, an engineering team, customers, capital, and considerable experience deploying AI.

You have no comparable starting advantage in energy infrastructure, biotechnology, or semiconductor fabrication.

That's not an argument against changing fields. It's an argument that your existing advantages have real option value.

6. What I would actually do in your position

I wouldn't abandon Fork because of Leopold's thesis.

But I would let the possibility of rapid intelligence improvement fundamentally influence how I build it.

Specifically, I'd avoid making the company's long-term value depend on AI remaining bad at some particular task.

For example, "our agent is better at building CRMs than ChatGPT" strikes me as a dangerous long-term proposition.

Instead, I'd want the company to become more valuable as the underlying models improve.

That means orienting the business toward distribution, deployment, ongoing operations, permissions, customer ownership, and the commercial relationships surrounding software.

In other words, build something that benefits from intelligence becoming cheaper, rather than something whose value depends on intelligence remaining expensive.

And I would put a time limit on the experiment.

You don't need to decide whether superintelligence will arrive in 2028 or 2040. You need to decide whether, over the next six months, you can demonstrate that Fork is accumulating something durable.

Not just generating software. Not just impressing customers with AI.

Actually accumulating distribution, recurring revenue, and a position in customers' operations.

If that isn't happening, the argument for reallocating your time becomes much stronger.


7. The philosophical issue underneath all this

There's something else I want to challenge in your reasoning.

You're implicitly assuming that if superintelligence arrives, the rational thing to do today is to position yourself to receive its benefits rather than participate in its deployment.

But that assumes the benefits will accrue primarily to owners of physical assets.

I'm not convinced.

Historically, transformative technologies have often created enormous value for the businesses that translate them into practical economic activity.

Electricity didn't make every electricity-dependent business worthless relative to power companies.

Computers didn't make every software company worthless relative to semiconductor manufacturers.

The internet didn't make every internet business worthless relative to telecommunications providers.

The counterargument, of course, is that superintelligence is categorically different because it could replace the human judgment and engineering that historically made complementary businesses valuable.

That's the real crux.

If AI becomes capable of independently performing essentially all economically useful cognitive work, how much of today's economic structure survives?

If the answer is "very little," then your instinct to seek ownership of scarce physical assets makes sense.

But if the answer is "a surprising amount, because economic institutions, trust, physical infrastructure, and relationships are harder to replace than intelligence," then helping businesses deploy AI could remain an enormous opportunity.

And there's one final distinction I'd make.

You don't actually need to believe your hypothesis about slowing intelligence progress to justify working on Fork.

You only need to believe that the economic diffusion of AI will be slower than its technological development, and that you can capture meaningful value during that gap.

That is a much weaker assumption, and I think a more defensible one.

The question I'd really want you to answer is this:

If you knew with certainty that superintelligence would arrive in 2030, would you still want to build Fork for the next four years?

If the answer is yes, but you'd build a different kind of Fork, that tells you something important.

If the answer is no, I'd take that seriously too. Because then your current strategy may depend on a technological forecast that you don't actually have much confidence in.

Human participant

If superintelligence arrives at any point, we have a problem. To me, it's a clear class problem.

The entire existence of the middle class, to my mind, and the american middle class in particular, is that elites realized (consciously or unconsciously) that if you want your society to achieve great things, you have to share some wealth with knowledge workers. Give them an opportunity to use their intelligence to accumulate wealth of their own. Some of them even become the richest people on the planet. But it's essentially a meritocracy, based around people's capacity to figure out our economic system and use it to maximum advantage.

And that system has issues, but I think it's generally a good system. But, in a world where intelligence is commodified, we devolve back to an almost feudal state wherein substantially the only things of value are capital and political power.

This seems like it would almost fully break our system, or at least push it to it's limits. From that view, I should probably be seeking capital or political power as much as possible. I wasn't born rich.

Language model

I think this is one of the most compelling arguments for why superintelligence could be profoundly destabilizing, even if alignment is solved, economic output explodes, and humanity becomes materially wealthier.

You're describing something more specific than mass unemployment. You're describing the destruction of the economic basis for meritocracy, and potentially the destruction of the bargaining power that underpins liberal democracy.

And I think there's a lot to this argument.

But I want to challenge one part of it: the idea that the middle class exists primarily because elites needed to reward intelligence. I think the historical story is more complicated, and that complication actually makes your concern more serious.

1. The middle class exists because labor has bargaining power

For much of human history, land was the dominant productive asset. Ownership of land conferred wealth, political influence, and control over other people's livelihoods.

Industrialization changed the equation.

Capital remained enormously important, but industrial economies increasingly depended on workers possessing skills, knowledge, and organizational capabilities that couldn't easily be replaced.

Over time, education became a mechanism through which people without inherited wealth could acquire economically valuable human capital.

The American middle class emerged from several interacting forces: industrial productivity, labor organizing, public education, mass homeownership, government policy, expanding consumer markets, and the economic advantages of the United States after World War II.

It wasn't simply that elites decided to share their wealth with smart people.

Rather, workers possessed something the owners of capital needed, and they acquired enough economic and political leverage to claim a portion of the value they created.

Your argument is that superintelligence threatens to eliminate that leverage.

And that's exactly the right place to focus.

2. AI could fundamentally change the relationship between labor and capital

Consider three economic systems.

<box gap={2}> <box border radius="lg" padding={3} gap={2}> <row align="center" justify="between"> **Agrarian economy** <text color="secondary" size="xs">Land is scarce</text> </row> <grid columns={2} gap={2}> <grid-item> <text color="secondary" size="xs">Primary source of wealth</text> <title size="lg">Land ownership</title> </grid-item> <grid-item> <text color="secondary" size="xs">Primary source of power</text> <title size="lg">Hereditary elites</title> </grid-item> </grid> <text size="sm">Most people work assets they do not own. Economic mobility is severely constrained.</text> </box> <box border radius="lg" padding={3} gap={2}> <row align="center" justify="between"> **Industrial and information economy** <text color="secondary" size="xs">Skilled labor is scarce</text> </row> <grid columns={2} gap={2}> <grid-item> <text color="secondary" size="xs">Primary source of wealth</text> <title size="lg">Capital + human labor</title> </grid-item> <grid-item> <text color="secondary" size="xs">Primary source of power</text> <title size="lg">Owners + workers</title> </grid-item> </grid> <text size="sm">People can convert education, skills, and entrepreneurship into income and eventually ownership.</text> </box> <box border radius="lg" padding={3} gap={2}> <row align="center" justify="between"> **Hypothetical superintelligent economy** <text color="secondary" size="xs">Capital and physical resources remain scarce</text> </row> <grid columns={2} gap={2}> <grid-item> <text color="secondary" size="xs">Primary source of wealth</text> <title size="lg">AI, infrastructure, capital</title> </grid-item> <grid-item> <text color="secondary" size="xs">Primary source of power</text> <title size="lg">Owners + institutions</title> </grid-item> </grid> <text size="sm">If human labor becomes economically unnecessary, people may lose their primary means of accumulating wealth independently.</text> </box> </box>

The critical shift isn't that intelligence becomes worthless.

It's that human intelligence loses its scarcity premium.

A brilliant engineer, lawyer, scientist, or entrepreneur might no longer be able to command a substantial income merely because they can solve difficult problems.

And this threatens something that goes beyond consumption.

It threatens the ability of someone born without wealth to become independently wealthy through their own productive abilities.

That is an extraordinary social change.

3. The real danger is that capitalism stops being a mechanism for social mobility

One of capitalism's remarkable properties is that it allows some people to convert talent, effort, risk-taking, and judgment into ownership.

You don't have to inherit a factory to become a factory owner. You can create a business, earn profits, accumulate capital, and eventually own productive assets.

This process is imperfect and heavily influenced by inherited advantage, but it exists.

Now imagine a world in which AI can outperform humans at virtually every economically valuable cognitive activity.

What does an ambitious 22-year-old without inherited wealth do?

Traditionally, they could become an engineer, a physician, an entrepreneur, an investment banker, or a scientist. They could accumulate savings, invest, and eventually acquire meaningful ownership.

But if those occupations no longer command a substantial wage premium, the first rung of the ladder disappears.

The economy could become extraordinarily productive while the pathways into its ownership class become increasingly inaccessible.

This is what makes your feudal analogy powerful.

The danger isn't necessarily that everyone becomes poor. It's that everyone who doesn't already own productive capital becomes economically dependent on those who do.

A society could have abundant food, excellent healthcare, cheap transportation, and advanced technology while still exhibiting an extreme concentration of power.

Material abundance and political equality are not the same thing.

4. But there's an important counterargument: AI could also democratize capital

Here's where I would push back.

You're assuming that the owners of today's capital will necessarily be the owners of tomorrow's productive capacity.

That isn't obvious.

Suppose AI becomes capable of autonomously building and operating businesses.

Today, starting a successful company requires substantial human expertise: engineering, finance, marketing, operations, law, and management.

If those capabilities become cheap, the capital required to start a competitive business could actually decrease.

A person with $10,000 and access to powerful AI might be able to operate a business that previously required 100 employees and millions of dollars.

In that scenario, AI doesn't necessarily reinforce the dominance of existing capital owners.

It might dramatically reduce the advantages of incumbency.

The question becomes whether AI itself is controlled by a handful of companies or becomes widely accessible and inexpensive.

And whether physical bottlenecks, distribution, regulation, and economies of scale prevent small operators from competing.

Consider two very different futures:

Concentrated AI ownershipDistributed AI ownership
Frontier models controlled by a few firmsPowerful models widely available
Compute remains extremely expensiveInference costs collapse
Incumbents dominate productionSmall firms compete effectively
Returns flow primarily to existing capitalNew entrepreneurs can create capital
Inequality could intensifyEconomic mobility might improve

Neither outcome is inevitable.

And critically, even widely distributed AI doesn't guarantee broadly distributed wealth. Competition could drive the profits of AI-enabled small businesses toward zero, leaving owners of genuinely scarce resources with most of the surplus.

That brings us back to your original concern.

5. Political power might become more important than economic power

I think this is the most underappreciated part of your argument.

Modern democracies derive some of their stability from the fact that ordinary citizens are economically necessary.

Governments need workers to produce goods, operate infrastructure, pay taxes, serve in the military, and sustain economic activity.

If autonomous AI and robotics can eventually perform most of those functions, the economic dependence of elites on the broader population could decline.

That does not automatically mean democracy collapses. Political rights are institutions, not simply payments for economic usefulness.

But it potentially weakens one of the forces that has historically helped sustain democratic bargaining power.

There is also a countervailing possibility: if AI dramatically increases economic output, governments could tax capital and distribute enough of the gains to sustain a prosperous population.

The problem is that this creates a distinction between two kinds of citizens.

One possesses productive assets and receives income through ownership.

The other receives income primarily through political redistribution.

Even if both enjoy excellent material living standards, their economic independence is very different.

And a system in which most people depend on redistribution could become politically fragile if the institutions guaranteeing that redistribution weaken.

A universal basic income might solve the consumption problem without solving the power problem.

That's an important distinction.

6. So should you personally be accumulating capital and political power?

I think the capital part follows quite naturally from your argument.

But I would separate three objectives:

  • Financial independence: Accumulating diversified ownership of productive assets so that your livelihood doesn't depend entirely on selling your labor.
  • Entrepreneurial ownership: Building a company whose value can grow as AI becomes more capable, rather than being displaced by it.
  • Political agency: Supporting and participating in institutions that determine how AI-generated wealth and power are governed.

These aren't mutually exclusive.

In fact, as a founder, you're already pursuing the second objective.

And I think there's an irony in your position.

You say you weren't born rich, and therefore you need to find a way to acquire capital before intelligence becomes commoditized.

But you're already doing exactly what the current economic system permits ambitious people without substantial inherited wealth to do: you're attempting to convert your intelligence, labor, and risk-taking into equity ownership.

You have a company, investors, employees, customers, and an ownership stake in a potentially valuable asset.

That's an unusual position of opportunity, even though it may not feel like it when you're struggling to find product-market fit.

The question isn't necessarily whether you should abandon entrepreneurship to pursue capital.

It's whether your current entrepreneurial activity has a sufficiently attractive probability of producing durable capital ownership before the economic rules change.

And I wouldn't confuse political power with political connections. Building meaningful political influence takes time, institutional credibility, coalitions, and public trust. It's not obviously a faster route to economic security than building a successful company.


7. There is one aspect of your worldview I would challenge more fundamentally

You describe the modern economy as a meritocracy based on people's capacity to understand and exploit the economic system.

There's truth to that, particularly among highly educated professionals and entrepreneurs.

But historically, many of the most important expansions of middle-class prosperity were not primarily driven by exceptional intelligence.

They were driven by collective bargaining, political institutions, public investment, and broad access to productive assets.

The postwar American middle class included millions of people who accumulated meaningful wealth without extraordinary cognitive ability.

They benefited from an economy in which ordinary labor commanded decent wages and institutions helped distribute productivity gains.

That's important because it suggests that the future of the middle class need not depend on humans remaining intellectually competitive with machines.

It could instead depend on how ownership and economic rights are structured.

For example, imagine a society where every citizen owns a diversified stake in the productive AI infrastructure of their economy, perhaps through sovereign wealth funds, public investment vehicles, or other institutions.

That could preserve economic independence without requiring human labor to remain scarce.

The challenge is getting from our current ownership structure to that arrangement.

And that's fundamentally a political problem, not a technological one.

8. My ultimate concern is slightly different from yours

You're worried that intelligence becoming abundant will make capital the only remaining source of economic power.

I'm worried about something more specific:

Intelligence might become abundant while ownership of the infrastructure that produces and deploys it remains concentrated.

Those are different problems.

If intelligence is abundant and accessible to everyone, it could be a tremendous equalizing force.

If intelligence is abundant but controlled by a small number of organizations, it could create unprecedented concentrations of power.

And if intelligence is abundant, accessible, but physical capital remains highly concentrated, we might end up somewhere in between.

The future distribution of power will depend enormously on which of those arrangements emerges.


I want to end with a question that I think gets to the heart of your personal dilemma.

Are you primarily trying to become wealthy before the window for upward mobility closes, or are you worried that the entire social order which made your own ambitions possible is going to disappear?

Because I think you're expressing both concerns, and they're related, but they lead to quite different decisions about what to do with your life.

One suggests accumulating capital as quickly as possible.

The other suggests that even becoming wealthy yourself might not protect the kind of society you actually want to live in.

Human participant

But your concern leads to a clear solution in my mind. Regardless of the ownership of compute, the production of compute remains an scarce resource. In a world where redistribution of ownership occurs, products remains a limiting factor, in fact a more limiting factor. I think it's incredibly stupid to invest in players like coreweave, they're essentially just financial instruments with logos.

But by the above logic, even a longshot chip bet, or a BAD chip bet, on people who know little about chips, could be a low hanging fruit way to lock in a place in the economy of the future. Because that's the spot where the tide picks up all ships, in a wide variety of scenarios.

Language model

I think you've arrived at a much stronger investment thesis than simply "AI will be big, so buy AI stocks."

You're proposing something like this:

If intelligence becomes abundant, the most valuable assets will be the physical bottlenecks preventing intelligence from becoming infinitely abundant. And those bottlenecks should become more valuable almost regardless of who ultimately controls AI.

That's a compelling thesis. In fact, it resembles the economic logic behind investing in land during industrialization, oil during the automobile revolution, or semiconductor manufacturing during the computer revolution.

But I think there's one very important flaw in your conclusion that even a bad chip company might be a good investment.

A rising tide can lift an entire industry's revenues while destroying the equity value of most companies in that industry.

This distinction is absolutely central.

1. You're betting on the supply curve of intelligence

Consider what happens if demand for intelligence increases by 100×.

If intelligence is produced by combining chips, electricity, infrastructure, and algorithms, then the demand for those inputs should increase dramatically.

And importantly, that demand could grow whether:

  • OpenAI wins or Anthropic wins.
  • AI is centralized or decentralized.
  • AGI arrives in 2028 or 2040.
  • AI replaces knowledge workers or merely augments them.
  • The AI economy becomes competitive or monopolistic.

That's an attractive property of the thesis.

You're trying to invest in a bottleneck rather than predict which downstream application wins.

And unlike an AI software company, a chip manufacturer might benefit from the commoditization of intelligence itself.

But there's a crucial distinction between a resource that becomes economically essential and a company that can capture profits from providing that resource.

Those are not the same thing.

2. Scarcity doesn't automatically translate into shareholder returns

Imagine the following hypothetical semiconductor market.

Demand for AI chips increases tenfold over five years.

<box gap={2}> <text weight="medium" size="sm">Illustrative chip-market outcomes</text> <table> <table-row><table-cell>Metric</table-cell><table-cell align="end">Today</table-cell><table-cell align="end">Five years later</table-cell></table-row> <table-row><table-cell>Annual chip demand</table-cell><table-cell align="end">$100B</table-cell><table-cell align="end">$1T</table-cell></table-row> <table-row><table-cell>Number of competitors</table-cell><table-cell align="end">5</table-cell><table-cell align="end">40</table-cell></table-row> <table-row><table-cell>Average operating margin</table-cell><table-cell align="end">40%</table-cell><table-cell align="end">8%</table-cell></table-row> <table-row><table-cell>Industry operating profit</table-cell><table-cell align="end">$40B</table-cell><table-cell align="end">$80B</table-cell></table-row> </table> <caption>Hypothetical figures, not a forecast.</caption> </box>

The market becomes ten times larger, but aggregate operating profits only double.

Meanwhile, companies might need to raise enormous amounts of capital to compete, diluting existing shareholders.

Some could go bankrupt despite growing revenues.

And even successful firms could deliver poor investment returns if their valuations already anticipate extraordinary growth.

This is what happened repeatedly in railroads, airlines, telecommunications, and other infrastructure-heavy industries.

The technology can transform civilization without making its investors rich.

Language model

I think you've arrived at a much stronger investment thesis than simply "AI will be big, so buy AI stocks."

You're proposing something like this:

If intelligence becomes abundant, the most valuable assets will be the physical bottlenecks preventing intelligence from becoming infinitely abundant. And those bottlenecks should become more valuable almost regardless of who ultimately controls AI.

That's a compelling thesis. In fact, it resembles the economic logic behind investing in land during industrialization, oil during the automobile revolution, or semiconductor manufacturing during the computer revolution.

But I think there's one very important flaw in your conclusion that even a bad chip company might be a good investment.

A rising tide can lift an entire industry's revenues while destroying the equity value of most companies in that industry.

This distinction is absolutely central.

1. You're betting on the supply curve of intelligence

Consider what happens if demand for intelligence increases by 100×.

If intelligence is produced by combining chips, electricity, infrastructure, and algorithms, then the demand for those inputs should increase dramatically.

And importantly, that demand could grow whether:

  • OpenAI wins or Anthropic wins.
  • AI is centralized or decentralized.
  • AGI arrives in 2028 or 2040.
  • AI replaces knowledge workers or merely augments them.
  • The AI economy becomes competitive or monopolistic.

That's an attractive property of the thesis.

You're trying to invest in a bottleneck rather than predict which downstream application wins.

And unlike an AI software company, a chip manufacturer might benefit from the commoditization of intelligence itself.

But there's a crucial distinction between a resource that becomes economically essential and a company that can capture profits from providing that resource.

Those are not the same thing.

2. Scarcity doesn't automatically translate into shareholder returns

Imagine the following hypothetical semiconductor market.

Demand for AI chips increases tenfold over five years.

<box gap={2}> <text weight="medium" size="sm">Illustrative chip-market outcomes</text> <table> <table-row><table-cell>Metric</table-cell><table-cell align="end">Today</table-cell><table-cell align="end">Five years later</table-cell></table-row> <table-row><table-cell>Annual chip demand</table-cell><table-cell align="end">$100B</table-cell><table-cell align="end">$1T</table-cell></table-row> <table-row><table-cell>Number of competitors</table-cell><table-cell align="end">5</table-cell><table-cell align="end">40</table-cell></table-row> <table-row><table-cell>Average operating margin</table-cell><table-cell align="end">40%</table-cell><table-cell align="end">8%</table-cell></table-row> <table-row><table-cell>Industry operating profit</table-cell><table-cell align="end">$40B</table-cell><table-cell align="end">$80B</table-cell></table-row> </table> <caption>Hypothetical figures, not a forecast.</caption> </box>

The market becomes ten times larger, but aggregate operating profits only double.

Meanwhile, companies might need to raise enormous amounts of capital to compete, diluting existing shareholders.

Some could go bankrupt despite growing revenues.

And even successful firms could deliver poor investment returns if their valuations already anticipate extraordinary growth.

This is what happened repeatedly in railroads, airlines, telecommunications, and other infrastructure-heavy industries.

The technology can transform civilization without making its investors rich.

3. But I think your argument becomes especially strong if we distinguish chip design from chip production

You used the phrase "production of compute," and I think that's where the most interesting opportunity lies.

The semiconductor supply chain is not a single industry. It's a series of interdependent bottlenecks.

<box gap={2}> <box border radius="lg" padding={3} gap={1}> <row align="center" justify="between"> **Chip architecture and design** <text color="secondary" size="xs">NVIDIA, AMD, startups</text> </row> AI can potentially automate much of the design process. Competition could become fierce as design costs fall. <text color="secondary" size="xs">Scarcity: intellectual property, software ecosystems, engineering, distribution</text> </box> <row justify="center"> <icon name="arrow-down" color="tertiary"/> </row> <box border radius="lg" padding={3} gap={1}> <row align="center" justify="between"> **Semiconductor fabrication** <text color="secondary" size="xs">TSMC, Samsung, Intel</text> </row> Manufacturing requires enormous capital expenditure, advanced processes, specialized equipment, and years of accumulated expertise. <text color="secondary" size="xs">Scarcity: factories, process know-how, production yield, capacity</text> </box> <row justify="center"> <icon name="arrow-down" color="tertiary"/> </row> <box border radius="lg" padding={3} gap={1}> <row align="center" justify="between"> **Semiconductor equipment** <text color="secondary" size="xs">ASML, Applied Materials, Lam Research</text> </row> Specialized machines are necessary to expand and upgrade fabrication capacity. Some suppliers have exceptionally difficult-to-replicate capabilities. <text color="secondary" size="xs">Scarcity: precision manufacturing, patents, supply chains</text> </box> <row justify="center"> <icon name="arrow-down" color="tertiary"/> </row> <box border radius="lg" padding={3} gap={1}> <row align="center" justify="between"> **Packaging, memory and interconnect** <text color="secondary" size="xs">SK Hynix, Micron, Broadcom</text> </row> Producing more compute requires not just processors, but memory bandwidth, advanced packaging, and communication between chips. <text color="secondary" size="xs">Scarcity: manufacturing capacity, bandwidth, integration</text> </box> <row justify="center"> <icon name="arrow-down" color="tertiary"/> </row> <box border radius="lg" padding={3} gap={1}> <row align="center" justify="between"> **Electricity and physical infrastructure** <text color="secondary" size="xs">Utilities, generation, grid equipment</text> </row> Compute requires reliable power, cooling, land, transmission, and grid connections. <text color="secondary" size="xs">Scarcity: energy, permits, physical capacity, time</text> </box> </box>

If I were following your thesis to its logical conclusion, I would actually be more interested in semiconductor equipment, advanced packaging, and energy infrastructure than in a speculative chip designer.

Why?

Because AI might get much better at designing chips, but it cannot instantly manufacture the factories and equipment required to produce them.

That's a genuine physical constraint.

And the more successful AI becomes, the more valuable it may be to eliminate those constraints.

4. Here's where I'd defend your idea of making a longshot chip bet

There's a version of your argument I find particularly interesting.

Suppose a group of relatively inexperienced engineers starts a semiconductor company today.

They have access to increasingly powerful AI systems capable of assisting with architecture, verification, physical design, and eventually substantial portions of the chip-development process.

Their initial competitive disadvantage might be enormous.

But what if AI progress disproportionately benefits small teams that previously couldn't afford to compete?

A company that would have required 200 engineers might eventually operate with 20.

In that world, owning equity in a small chip startup could provide extraordinary upside.

And unlike a software startup, the eventual product participates directly in the physical infrastructure of AI.

So yes, there's a potentially interesting asymmetric opportunity here.

But I would distinguish two bets:

Bet A: AI makes chip design radically easier. Small, inexperienced teams gain access to capabilities previously reserved for enormous semiconductor companies.

Bet B: Physical compute becomes more valuable. Demand for processors and infrastructure grows rapidly.

The problem is that Bet A could undermine Bet B's profitability for chip designers.

If thousands of new companies can design competitive processors, competition might eliminate the margins that made chip companies attractive in the first place.

The best scenario for your hypothetical startup is one where AI substantially reduces its engineering costs, but barriers to entry remain high enough to preserve pricing power.

That's possible, particularly in specialized markets. But it isn't automatic.

5. I think there's a deeper issue: you're conflating bottleneck exposure with ownership of the bottleneck

Imagine two businesses.

Company A is a startup designing AI accelerators. It has 12 engineers, an interesting architecture, and no manufacturing capability.

Company B manufactures a specialized component required by nearly every advanced semiconductor fabrication facility. It has only a handful of serious competitors and substantial production capacity.

If AI demand increases a hundredfold, Company B might enjoy significantly more predictable demand.

Company A, meanwhile, still needs to design a successful chip, pay for fabrication, secure packaging and memory, compete with incumbents, attract customers, and raise capital.

Its success depends on a long sequence of uncertain events.

You could make a fortune owning Company A, but its position in the AI supply chain doesn't make it inherently safer.

In fact, the most defensible businesses may be the least glamorous.

6. What about your argument that the tide lifts all ships?

I think the best historical analogy is the California Gold Rush.

The people mining for gold faced enormous uncertainty. Some became fabulously wealthy; many failed.

But the people selling essential supplies to miners often had more reliable businesses.

However, even that analogy has a subtle problem.

If everyone realizes that selling shovels is a great business, the price of shovel companies rises until future returns become ordinary.

That's why I'd distinguish:

[ \text{Investment return} \neq \text{Industry growth} ]

A more useful approximation is:

[ \text{Return} \approx \text{Earnings growth} + \text{Cash distributions} + \text{Valuation change} ]

Adjusted for dilution and other capital changes.

A company can grow earnings 30% annually and still be a terrible investment if you buy it at an absurd valuation.

Conversely, a mediocre company in a spectacular industry can be a great investment if the market dramatically underestimates its future profitability.

This is why the price you pay matters just as much as the technological future you're betting on.

7. There's another possibility that strengthens your thesis

What if the physical bottlenecks aren't just valuable because demand increases?

What if they're valuable because AI makes everything else cheaper?

Imagine superintelligence drives the marginal cost of software development, engineering, scientific research, and administrative work toward zero.

The relative economic importance of physical assets increases.

This doesn't necessarily mean the absolute price of every physical asset rises. AI might also make mining, manufacturing, and construction dramatically more efficient.

But physical constraints become a larger fraction of the remaining economic problem.

And that means ownership of genuinely scarce productive assets could become increasingly important.

This is probably the strongest version of your argument.

You want to own assets whose scarcity survives the commoditization of human intelligence.

I agree with that objective.

Where I disagree is that owning a speculative chip designer necessarily accomplishes it.

A chip startup might primarily own intellectual property and engineering expertise—the very things you expect AI to commoditize.

An advanced semiconductor fabrication facility, specialized equipment manufacturer, or energy-generation asset might have more of the properties you're looking for.


8. How I'd translate this into your personal strategy

Given your situation, I see three distinct approaches.

ApproachPotential advantagePrincipal risk
Build Fork and accumulate equityLeverages your existing team, capital, customers, and experienceSoftware creation becomes commoditized before you establish durable value
Invest in physical AI infrastructureOwnership of assets that could remain scarce after superintelligenceHigh valuations, capital intensity, industry cycles
Found or join a semiconductor startupPotentially enormous equity upside with exposure to AI-driven engineering improvementsVery high technical, commercial, financing, and dilution risk

I would be reluctant to abandon an existing, funded company to start a semiconductor business without a specific technical or commercial insight.

But I would take seriously the idea of gradually building personal exposure to the physical infrastructure underlying AI.

And I think your entrepreneurial instincts might be particularly useful in identifying overlooked opportunities.

For example, rather than attempting to design an accelerator that beats NVIDIA, perhaps there's an unglamorous component, manufacturing process, testing technology, or infrastructure bottleneck where a small company could establish a defensible position.

The most interesting opportunities may be places where the demand curve is about to change dramatically but the supply side cannot respond quickly.

That's where scarcity can translate into economic rents.

9. One final challenge to your thesis

You're treating the production of compute as something that will necessarily remain scarce.

But what if superintelligence itself radically improves compute production?

It could discover new architectures, improve fabrication yields, optimize supply chains, invent better materials, and eventually accelerate physical manufacturing through robotics.

If that happens, compute could become dramatically cheaper and more abundant.

There will still be physical constraints, but the bottleneck could migrate from chips to electricity, then to grid infrastructure, then to land, then to something else.

The bottleneck that exists immediately before superintelligence may not be the bottleneck that survives superintelligence.

That makes me more interested in diversified ownership of productive assets than in any single technological input.

And it suggests that the best hedge against a radically transformed economy may not be a concentrated bet on chips, but broad ownership of the capital stock that benefits from automation.


Here's the question I would put back to you, because I think it changes the analysis enormously:

Are you contemplating personally investing in semiconductor companies, or are you actually considering founding a semiconductor company yourself?

Because I think there's a potentially fascinating case for the latter: using AI to make semiconductor entrepreneurship accessible to people who historically couldn't compete. But that is a fundamentally different thesis from investing in the physical scarcity of compute.


If you want, I can:

  • Analyze risks of chip industry investment
  • Explore opportunities in semiconductor equipment sector
  • Evaluate how AI affects chip design competitiveness

Language model

3. But I think your argument becomes especially strong if we distinguish chip design from chip production

You used the phrase "production of compute," and I think that's where the most interesting opportunity lies.

The semiconductor supply chain is not a single industry. It's a series of interdependent bottlenecks.

<box gap={2}> <box border radius="lg" padding={3} gap={1}> <row align="center" justify="between"> **Chip architecture and design** <text color="secondary" size="xs">NVIDIA, AMD, startups</text> </row> AI can potentially automate much of the design process. Competition could become fierce as design costs fall. <text color="secondary" size="xs">Scarcity: intellectual property, software ecosystems, engineering, distribution</text> </box> <row justify="center"> <icon name="arrow-down" color="tertiary"/> </row> <box border radius="lg" padding={3} gap={1}> <row align="center" justify="between"> **Semiconductor fabrication** <text color="secondary" size="xs">TSMC, Samsung, Intel</text> </row> Manufacturing requires enormous capital expenditure, advanced processes, specialized equipment, and years of accumulated expertise. <text color="secondary" size="xs">Scarcity: factories, process know-how, production yield, capacity</text> </box> <row justify="center"> <icon name="arrow-down" color="tertiary"/> </row> <box border radius="lg" padding={3} gap={1}> <row align="center" justify="between"> **Semiconductor equipment** <text color="secondary" size="xs">ASML, Applied Materials, Lam Research</text> </row> Specialized machines are necessary to expand and upgrade fabrication capacity. Some suppliers have exceptionally difficult-to-replicate capabilities. <text color="secondary" size="xs">Scarcity: precision manufacturing, patents, supply chains</text> </box> <row justify="center"> <icon name="arrow-down" color="tertiary"/> </row> <box border radius="lg" padding={3} gap={1}> <row align="center" justify="between"> **Packaging, memory and interconnect** <text color="secondary" size="xs">SK Hynix, Micron, Broadcom</text> </row> Producing more compute requires not just processors, but memory bandwidth, advanced packaging, and communication between chips. <text color="secondary" size="xs">Scarcity: manufacturing capacity, bandwidth, integration</text> </box> <row justify="center"> <icon name="arrow-down" color="tertiary"/> </row> <box border radius="lg" padding={3} gap={1}> <row align="center" justify="between"> **Electricity and physical infrastructure** <text color="secondary" size="xs">Utilities, generation, grid equipment</text> </row> Compute requires reliable power, cooling, land, transmission, and grid connections. <text color="secondary" size="xs">Scarcity: energy, permits, physical capacity, time</text> </box> </box>

If I were following your thesis to its logical conclusion, I would actually be more interested in semiconductor equipment, advanced packaging, and energy infrastructure than in a speculative chip designer.

Why?

Because AI might get much better at designing chips, but it cannot instantly manufacture the factories and equipment required to produce them.

That's a genuine physical constraint.

And the more successful AI becomes, the more valuable it may be to eliminate those constraints.

4. Here's where I'd defend your idea of making a longshot chip bet

There's a version of your argument I find particularly interesting.

Suppose a group of relatively inexperienced engineers starts a semiconductor company today.

They have access to increasingly powerful AI systems capable of assisting with architecture, verification, physical design, and eventually substantial portions of the chip-development process.

Their initial competitive disadvantage might be enormous.

But what if AI progress disproportionately benefits small teams that previously couldn't afford to compete?

A company that would have required 200 engineers might eventually operate with 20.

In that world, owning equity in a small chip startup could provide extraordinary upside.

And unlike a software startup, the eventual product participates directly in the physical infrastructure of AI.

So yes, there's a potentially interesting asymmetric opportunity here.

But I would distinguish two bets:

Bet A: AI makes chip design radically easier. Small, inexperienced teams gain access to capabilities previously reserved for enormous semiconductor companies.

Bet B: Physical compute becomes more valuable. Demand for processors and infrastructure grows rapidly.

The problem is that Bet A could undermine Bet B's profitability for chip designers.

If thousands of new companies can design competitive processors, competition might eliminate the margins that made chip companies attractive in the first place.

The best scenario for your hypothetical startup is one where AI substantially reduces its engineering costs, but barriers to entry remain high enough to preserve pricing power.

That's possible, particularly in specialized markets. But it isn't automatic.

5. I think there's a deeper issue: you're conflating bottleneck exposure with ownership of the bottleneck

Imagine two businesses.

Company A is a startup designing AI accelerators. It has 12 engineers, an interesting architecture, and no manufacturing capability.

Company B manufactures a specialized component required by nearly every advanced semiconductor fabrication facility. It has only a handful of serious competitors and substantial production capacity.

If AI demand increases a hundredfold, Company B might enjoy significantly more predictable demand.

Company A, meanwhile, still needs to design a successful chip, pay for fabrication, secure packaging and memory, compete with incumbents, attract customers, and raise capital.

Its success depends on a long sequence of uncertain events.

You could make a fortune owning Company A, but its position in the AI supply chain doesn't make it inherently safer.

In fact, the most defensible businesses may be the least glamorous.

6. What about your argument that the tide lifts all ships?

I think the best historical analogy is the California Gold Rush.

The people mining for gold faced enormous uncertainty. Some became fabulously wealthy; many failed.

But the people selling essential supplies to miners often had more reliable businesses.

However, even that analogy has a subtle problem.

If everyone realizes that selling shovels is a great business, the price of shovel companies rises until future returns become ordinary.

That's why I'd distinguish:

[ \text{Investment return} \neq \text{Industry growth} ]

A more useful approximation is:

[ \text{Return} \approx \text{Earnings growth} + \text{Cash distributions} + \text{Valuation change} ]

Adjusted for dilution and other capital changes.

A company can grow earnings 30% annually and still be a terrible investment if you buy it at an absurd valuation.

Conversely, a mediocre company in a spectacular industry can be a great investment if the market dramatically underestimates its future profitability.

This is why the price you pay matters just as much as the technological future you're betting on.

7. There's another possibility that strengthens your thesis

What if the physical bottlenecks aren't just valuable because demand increases?

What if they're valuable because AI makes everything else cheaper?

Imagine superintelligence drives the marginal cost of software development, engineering, scientific research, and administrative work toward zero.

The relative economic importance of physical assets increases.

This doesn't necessarily mean the absolute price of every physical asset rises. AI might also make mining, manufacturing, and construction dramatically more efficient.

But physical constraints become a larger fraction of the remaining economic problem.

And that means ownership of genuinely scarce productive assets could become increasingly important.

This is probably the strongest version of your argument.

You want to own assets whose scarcity survives the commoditization of human intelligence.

I agree with that objective.

Where I disagree is that owning a speculative chip designer necessarily accomplishes it.

A chip startup might primarily own intellectual property and engineering expertise—the very things you expect AI to commoditize.

An advanced semiconductor fabrication facility, specialized equipment manufacturer, or energy-generation asset might have more of the properties you're looking for.


8. How I'd translate this into your personal strategy

Given your situation, I see three distinct approaches.

ApproachPotential advantagePrincipal risk
Build Fork and accumulate equityLeverages your existing team, capital, customers, and experienceSoftware creation becomes commoditized before you establish durable value
Invest in physical AI infrastructureOwnership of assets that could remain scarce after superintelligenceHigh valuations, capital intensity, industry cycles
Found or join a semiconductor startupPotentially enormous equity upside with exposure to AI-driven engineering improvementsVery high technical, commercial, financing, and dilution risk

I would be reluctant to abandon an existing, funded company to start a semiconductor business without a specific technical or commercial insight.

But I would take seriously the idea of gradually building personal exposure to the physical infrastructure underlying AI.

And I think your entrepreneurial instincts might be particularly useful in identifying overlooked opportunities.

For example, rather than attempting to design an accelerator that beats NVIDIA, perhaps there's an unglamorous component, manufacturing process, testing technology, or infrastructure bottleneck where a small company could establish a defensible position.

The most interesting opportunities may be places where the demand curve is about to change dramatically but the supply side cannot respond quickly.

That's where scarcity can translate into economic rents.

9. One final challenge to your thesis

You're treating the production of compute as something that will necessarily remain scarce.

But what if superintelligence itself radically improves compute production?

It could discover new architectures, improve fabrication yields, optimize supply chains, invent better materials, and eventually accelerate physical manufacturing through robotics.

If that happens, compute could become dramatically cheaper and more abundant.

There will still be physical constraints, but the bottleneck could migrate from chips to electricity, then to grid infrastructure, then to land, then to something else.

The bottleneck that exists immediately before superintelligence may not be the bottleneck that survives superintelligence.

That makes me more interested in diversified ownership of productive assets than in any single technological input.

And it suggests that the best hedge against a radically transformed economy may not be a concentrated bet on chips, but broad ownership of the capital stock that benefits from automation.


Here's the question I would put back to you, because I think it changes the analysis enormously:

Are you contemplating personally investing in semiconductor companies, or are you actually considering founding a semiconductor company yourself?

Because I think there's a potentially fascinating case for the latter: using AI to make semiconductor entrepreneurship accessible to people who historically couldn't compete. But that is a fundamentally different thesis from investing in the physical scarcity of compute.

C

One more record in a shared language.

Contribute a conversation