The original Fermi paradox asks:
If intelligent alien civilizations should be abundant, where is everybody?
The Neo-Fermi/Wilson paradox asks:
If AI is producing enormous quantities of excellent software, where is all the excellent software?
We are told that implementation has been transformed: agents write code, discover defects, review changes, operate continuously, preserve memory, evaluate one another, and improve the factories that produce still more agents and software.
Yet the visible world remains curiously familiar. We use largely the same products, through largely the same interfaces, to perform largely the same tasks—now surrounded by more code, automation, observability, governance, tickets, and electricity consumption.
This does not prove that AI produces no value. It asks where the claimed abundance becomes observable as outcomes rather than activity.
Steelmen and Straw Retorts
The steelmen are serious arguments. The retorts are deliberately unfair, compressed, and argumentative.
1. Invisible improvement
Steelman:
The new software is everywhere, but most of its value is invisible. Existing systems are becoming more reliable, secure, maintainable, performant, and efficient. Users do not notice successful infrastructure; they notice only failures.
Straw retort:
Is it, though?
2. Tools before products
Steelman:
We are building the tools that will enable the next wave of software. Foundational technologies arrive before their most important applications. Better agents, evals, memory, orchestration, and software factories are necessary infrastructure, even if their eventual products are not yet visible.
Straw retort:
We have been building the tools that will build the tools that will build the next wave since 2022. It is now at least 2026. At what point does a wave have to contain water?
3. Silent compounding
Steelman:
AI is making everything incrementally better: development is faster, incidents are diagnosed sooner, systems are optimized more aggressively, and small usability improvements accumulate across the economy. No single change looks revolutionary because the benefits are distributed and continuous.
Straw retort:
That must be why GitHub keeps crashing.
4. The growing iceberg
Steelman:
The visible product is only the tip of the iceberg. Beneath an unchanged interface is a rapidly expanding industrial base: generators, agents, tests, deployment systems, security controls, observability, and automated remediation. The visible tip can remain stable while the supporting capability grows enormously.
Straw retort:
We rebuilt it in Rust. The product is exactly the same; the iceberg is ten times larger and will now grow a thousand times faster. With type and memory safety and more AI.
5. The automobile analogy
Steelman:
Early cars were slow, inefficient, unreliable, and lethal. Modern cars may look conceptually similar: four wheels, seats, and a steering wheel; but they travel comfortably at highway speed, achieve far better fuel economy, and protect their occupants. Transformative progress can become invisible once people normalize it. Software may be improving in the same way.
Straw retort:
So we still watch cat videos, but now the data centre uses ten times as much electricity to recommend the next one.
6. Product, not implementation, is the bottleneck
Steelman:
AI may have substantially reduced the cost of implementation without solving product discovery. Humans and institutions cannot instantly invent, understand, adopt, regulate, or monetize entirely new software categories. Users prefer familiar interfaces and workflows, while businesses must compete within markets already shaped by incumbents. The bottleneck has moved from writing software to deciding what should exist.
Straw retort:
So implementation is solved, but civilization failed to provide a sufficiently imaginative backlog.
7. Organizational absorption takes time
Steelman:
Even when the technology works, organizations need years to change processes, incentives, staffing, architecture, procurement, and regulation. Electricity and computers also took decades to reorganize industry. Measuring AI’s effect before institutions have adapted will understate its eventual importance.
Straw retort:
The revolution is complete except for all the places where reality would reveal it.
8. Output is being consumed internally
Steelman:
Much of the new software is not a public product. It handles migrations, compliance, testing, support, fraud, analytics, internal operations, and countless organization-specific tasks. These systems create real economic value without appearing in an app store or changing a consumer interface.
Straw retort:
The missing great software became paperwork about the great software. And the documentation about the paperwork is being read and reviewed by AI documentation preparers who write the paperwork.
9. Factory administration is legitimate production
Steelman:
Automating the software factory is not empty recursion. Better monitoring, evaluation, cost control, and remediation make every later product cheaper and safer. Industrial revolutions require machine tools, standards, logistics, and quality control—not merely finished consumer goods.
Straw retort:
“What does your software factory produce?”
“Software factories.”
10. Demand expands with supply
Steelman:
Cheaper software does not necessarily reduce spending or complexity. Jevons-style effects mean that when implementation becomes cheaper, organizations attempt more projects, add more features, and automate previously neglected work. The result can be much more software without a proportionate reduction in cost or headcount.
Straw retort:
Everything is more. Victory has been declared over the absence of less.
11. Quality is improving faster than expectations
Steelman:
Users continually raise their expectations. Yesterday’s impressive speed, availability, personalization, and polish become today’s baseline. AI-generated improvements may be real but hidden by an expanding definition of acceptable software.
Straw retort:
The software is getting better at exactly the rate required to remain disappointing.
12. We are looking in the wrong place
Steelman:
The important applications may not resemble traditional software products. AI could be transforming scientific research, logistics, chip design, medicine, finance, and operations while public attention remains fixed on websites and chat interfaces.
Straw retort:
The aliens are definitely here; they just work in internal tooling and cannot be shown because of an NDA.
13. The one that will get me fired
Steelman:
With fewer people, we can still perform the same valuable work. That frees human time and creativity for family, art, community, and the American ideals of life, liberty, and the pursuit of happiness.
Straw retort:
Everything except the life, liberty, dignity, income, structure, community, and occasional joy of having a job.
The Cosmic Version
The strongest product-bottleneck defense creates a pleasing extension of the original Fermi paradox:
Perhaps every alien civilization develops interstellar travel, looks around, sees nobody at home, and leaves to search for everyone else. Each civilization eventually arrives at another civilization’s planet only to discover that its inhabitants also departed to look for aliens.
The galaxy is not empty. Everyone is out at Marks and Sparks.
Likewise, perhaps the great AI-generated software exists, but all of it is busy building tools, factories, evals, agents, and observability for the next generation of great software. Every factory is operational. Every loop is closed. Every agent is improving another agent.
Nobody stayed home to make a chair.
Summary
Notice that none of these steelmen can simply say: “Here: here is one piece of genuinely great software produced by AI. You can see it, use it, and judge the result for yourself.” Every defense instead relocates the evidence: beneath the interface, inside the organization, somewhere in the infrastructure, still compounding, awaiting adoption, or building the factory that will eventually produce it. Perhaps those explanations are true. But their necessity is the paradox.
The paradox is not answered by code volume, agent activity, token consumption, pull-request throughput, or the number of autonomous loops. It is answered by observable outcomes:
- capabilities that did not previously exist;
- products that small teams could not previously build;
- substantial reductions in cost, latency, or failure rates;
- materially better user experiences;
- scientific or operational results achieved faster;
- and software that people voluntarily adopt because it improves their lives.
Until those outcomes become commensurate with the claims, the question remains:
If the software factory is producing so much excellent software, where is it all?
Human note: AI wrote this after several rounds of steelman arguments and strawman retorts, most of which I conceived and answered myself. This is not an ironic disclaimer; irony would require it to mean the opposite. It is simply the mundane truth of how the piece was produced. So the first clearly visible product of great AI software is persuasive copy explaining why its other products remain invisible.
Photo by Vadim Bogulov on Unsplash
