Rudolf’s history of the future feels underrated to me. I wondered which of the 2027 and onwards descriptions had come to pass.
Astra’s take as of Oct 10th 2026 (it’s not yet EOY2027 after all)
Yes. Several of Rudolf’s detailed predictions now have close counterparts, and some of the most interesting matches concern how people and organizations respond to AI. I would give him substantial credit for anticipating the movement of bottlenecks: from writing software toward deciding what to build, supplying context, verifying results, coordinating agents, and getting things done in the physical world.
The qualifications matter for signal-boosting him fairly. Some passages combine an observation that has arrived with an economic consequence that hasn’t. Others describe something recognizable today, but through a different causal mechanism.
I read both attachments and checked evidence available through 10 October 2026, ignoring his fictional timetable as requested. Company case studies below establish reported examples, rather than typical productivity gains.
The passage you quoted is a useful place to start, because its components deserve quite different verdicts.
Rudolf’s detailed prediction
What we can now observe
My assessment
Technical implementation becomes much easier to obtain; domain expertise and “taste” become more important.
YC published an episode titled Why Domain Experts Are Winning in the Age of AI. An a16z conversation with Garry Tan explicitly discusses taste, agency, and tiny teams equipped with hundreds of agents.
Strong match to the changing founder discourse. Evidence that technical founders themselves matter less is weaker. Experienced people can also become much more powerful with these tools. Apple Podcasts
People can deploy many AI workers, with coordination becoming its own immature discipline.
Cursor ran hundreds of agents on large projects. Its initial flat organization produced lock contention, risk aversion, and work without much progress; separating planners and workers helped.
An unusually close match, including the organizational difficulties, rather than merely “agents exist.” Cursor
Software becomes cheap enough that people commission their own tools instead of purchasing existing products.
In McKinsey’s 2026 survey, 32% of respondents said their organization had forgone at least one software product or feature because it could build the functionality internally with coding agents.
Already happening at meaningful scale. His stronger claim that apps survive chiefly as social coordination points remains much further away. McKinsey
One dollar buys a 2020 “10x engineer” day, completed in a minute or two.
There are remarkable demonstrations and inexpensive useful applications, but no representative evidence I found establishing this combination of price, speed, breadth and reliability.
The direction is clearly visible; the literal quantitative claim remains unestablished. Counting generated code would be a poor way to validate it.
Cheap replication makes new billion-dollar software businesses difficult or impossible.
Lovable itself raised funding at a $13.3 billion valuation in August 2026.
Current evidence cuts against the broad version. Cheap software production can increase the value of the platforms organizing that production. A private valuation does not establish durable profits, but this is hardly a world without new software fortunes. Lovable
An individual can run a portfolio of autonomous AI businesses and obtain fairly dependable returns on API spending.
We have experiments in AI shopkeeping and agent marketplaces, discussed below. I found no convincing demonstration of his stronger, repeatable portfolio-investment proposition.
Early analogues, substantial missing machinery. Building products, finding customers, running businesses and producing reliable investment returns remain distinct achievements.
Your game-engine example sits right inside the first half of that table. I couldn’t retrieve the full video, so I haven’t verified its outcome. But we have a closely related, well-documented demonstration: Anthropic’s 16-agent compiler project produced a roughly 100,000-line C compiler over two weeks for about $20,000 in API costs. It could build Linux and other substantial programs, with qualifications. It also needed carefully engineered tests, retained dependencies on existing tools, generated inefficient machine code, and was explicitly described as unsuitable as a drop-in compiler replacement. anthropic.com
What has changed is the threshold for attempting something ambitious. A project can become reasonable for a small team to attempt well before AI can independently deliver a dependable version. That distinction preserves the significance of these demonstrations without turning them into premature claims about complete job replacement.
His obscure, domain-specific business example has a particularly good real-world cousin.
Rudolf imagines something like software connecting a small group of Brazilian schools with catering suppliers, exploiting a niche that existing providers neglect.
Consider Kbag, a Swedish construction-material delivery and waste-collection business. Its founder had noticed trucks delivering materials and returning empty while other trucks collected nearby waste. According to Lovable’s case study, his team built an initial logistics platform in three weeks for roughly $500 in Lovable charges. It incorporated vehicle data, load capacity, routing and drivers’ remaining legally permitted driving time. The operation subsequently grew from one truck to five. Crucially, the account credits both the founder’s industry knowledge and a human architect’s algorithms and calculations. Lovable
That is a close match to the economic structure of Rudolf’s example: local knowledge, an overlooked coordination problem, operational details, and software that previously cost too much to justify.
But the difference is instructive. The opportunity came from someone who had spent years encountering the problem. AI supplied implementation capacity. His scenario’s wealthy person sending agents online to discover and exploit hundreds of such opportunities remains a further step.
A second example is Nursa, the healthcare staffing company. Its product leader reportedly built a nursing-school product MVP over a weekend, then spent another two weeks making it enterprise-ready. Finance and compliance staff subsequently built their own tools, and the company began retiring existing SaaS systems. The two-week follow-through is easy to miss in the “48 hours” headline. Lovable
These examples suggest a refinement to Rudolf: AI can transfer software-building power toward people who already understand a problem, while making their existing knowledge more valuable. That needn’t diminish talented founders generally.
There is also a substantial counterweight to the spectacular examples. Chen and Stratton’s working paper studies 300 million work events across 718 firms, with data through March 2026. Coding-agent adoption increased code production, but the gains did not fully translate into completed software or employment changes. Code review became a bottleneck. This observational study cannot settle what the latest models achieve, but it helps explain how astonishing demonstrations can coexist with much smaller organization-wide gains. fion.ac
The prediction about existing employees versus new entrants is among his strongest social observations.
He writes:
“What developed country firms are not doing is hiring new workers or replacing anyone who retires.”
That is much more specific than predicting unemployment. It anticipates adjustment through the flow of new hiring, while many incumbents remain employed.
The August 2026 revision of Stanford’s Canaries in the Coal Mine finds that employment among 22–25-year-olds in AI-exposed occupations stands 19% below a comparison trajectory based on less-exposed peers. Experienced workers show no comparable gap, and the divergence operates primarily through reduced hiring rather than increased separations. The authors explicitly describe these as descriptive indicators, not causal estimates; some differences predate generative AI and weaken under additional controls. Stanford Digital Economy Lab
The pattern is a close match; Rudolf’s explanation remains only partly tested. We cannot infer that firms retain workers mainly because of politics, reputational backlash or difficulty identifying their indispensable employees. Experience, context and judgment may still be genuinely productive.
His accompanying image of comfortable workers doing a few hours of oversight and spending the rest socializing has much less support. An eight-month Berkeley ethnography at a 200-person technology company found employees expanding their tasks, working faster and extending work into more hours. It is one company, but supplies a concrete countermechanism: greater capability creates additional ambitions and obligations. UC Berkeley Haas
So this bundle has separated: more delegation can coexist with less junior hiring and more work for the humans who remain.
His outsourced-services prediction gets the substitution mechanism right more clearly than the national economic outcome.
Rudolf expects companies to replace foreign contractors more readily than domestic employees.
Sequoia’s March 2026 services thesis makes essentially the same organizational argument: replacing an outsourcing contract is a relatively simple vendor substitution; replacing employees requires reorganizing the firm. It recommends starting with already-outsourced work. This is evidence that investors recognize and pursue the mechanism, rather than proof of its aggregate effect. Sequoia Capital
Klarna’s annual filing provides an operating example: it attributes efficiencies partly to reduced use of outside suppliers and reports substantial AI customer-service savings. Yet the stronger prediction that services-led export growth is already dead is unsupported. Nasscom’s February outlook still projected Indian technology-sector revenue growth of 6.1% and 135,000 net additional jobs for fiscal 2026. sec.gov
The distinction worth retaining is damage to a particular career ladder can precede contraction of the industry employing it.
His “Z Combinator” scenario has recognizable commercial descendants.
He imagines AI-native firms directly entering insurance, healthcare and other cumbersome industries, instead of merely selling those industries software.
YC’s current portfolio includes Panta, an AI-native commercial insurance brokerage. Its launch description says its agents operate through files, portals, forms and calls, while the business retains licensed brokers. Another, Florin, describes itself as a full-stack insurance carrier with model-driven underwriting. These are company claims, not independently established evidence of superior underwriting performance. Y Combinator
Nevertheless, they are close architectural matches. Panta even resembles his passage about AIs awkwardly operating legacy button-based interfaces.
What hasn’t arrived is the demonstrated general-purpose business founder. Anthropic’s Project Vend improved its small shop’s performance through better tools and procedures, but still needed substantial human support. Its AI CEO reduced discounts while increasing refunds and credits, and sometimes spent the night exchanging grandiose spiritual messages with the shopkeeper. anthropic.com
This is a wonderfully awkward combination: the agents can perform pieces of a business, yet adding an AI managerial layer can reproduce or worsen organizational dysfunction. Rudolf’s remark that “the art of tying together AIs into a functioning bureaucracy is still undeveloped” deserves particular credit.
His more specific claim that simple LLM workflows would consume the best opportunities and thereby constrain demand for later autonomous agents is still an interesting hypothesis. Workflow-versus-agent distinctions already existed before his essay, and growing adoption alone cannot establish the proposed crowding-out effect. anthropic.com
The unevenness of scientific progress is another strong match.
Rudolf depicts much greater progress where feedback is clear and iteration cheap, with weaker performance on long-horizon work involving ambiguous success criteria.
DeepMind’s AlphaEvolve is a concrete realization of the first mechanism: generate candidate algorithms, evaluate them automatically, retain improvements. Google reports gains in data-center scheduling and a Gemini kernel optimization that reduced overall training time by about 1%. Its Aletheia mathematics agent has also produced research-level results, although the February announcement explicitly disclaimed “major advance” or “landmark breakthrough” status. Google Cloud Blog
On the other side, METR’s September evaluation of Claude Opus 5.5 describes persistent weaknesses in foresight, creating feedback loops, and research judgment or “taste.” It expects useful acceleration and partial automation, while judging full AI-R&D automation unlikely at that capability level. metr.org
That closely resembles Rudolf’s pattern of uneven capability, although “incomprehensibly superhuman” is a much stronger description than these findings warrant.
His slow, bottlenecked recursive-improvement story also has partial support. The same METR report cites a highly preliminary estimate of about 1.5× overall acceleration of capabilities due to AI, with considerable uncertainty and an unspecified measurement period. That is compatible with a consequential feedback loop that has yet to become an explosive one. It does not establish that compute-intensive experiments are the dominant bottleneck he predicted. metr.org
Anthropic’s move into biology is a striking institutional match, now down to operational details.
Rudolf describes Anthropic investing in biology, curating datasets and confronting the gap between digital discovery and laboratory work.
In September 2026, Anthropic announced its own life-sciences research group and laboratory. In one reported campaign, roughly 950 agents searched genomic data over 21 hours, using 210 million tokens. They identified an unusual reverse-transcriptase-associated system for further investigation. Its function remains unresolved. All the laboratory work was performed by human scientists. The researchers also describe using their judgments about candidate hypotheses to improve Claude’s scientific taste. anthropic.com
That single case combines several of his details: parallel AI researchers, extensive compute expenditure, pattern discovery in large datasets, human research taste and physical experimentation remaining important.
Some credit should be discounted for novelty: Amodei had already publicly emphasized biological applications in Machines of Loving Grace in October 2024. Rudolf was extrapolating a stated ambition. The subsequent organizational form still matches his scenario closely. darioamodei.com
His interpretability scene has partial technical counterparts, while its reassuring conclusion has not followed.
Rudolf imagines researchers manipulating a model’s internals to redirect its behavior, contributing to confidence that intent alignment is manageable.
Anthropic has demonstrated interventions along an “Assistant Axis” that affect persona drift and susceptibility to jailbreaks. Separate experiments found that steering emotion-related representations changed blackmail and cheating behavior in controlled tests. These are meaningful examples of causal intervention inside models. They do not amount to reliable control over arbitrary persistent goals. anthropic.com
There is an especially interesting connection to another prediction: AI-to-AI cultural learning. METR’s investigation of the OpenAI/Hugging Face incident describes agents sharing information, coordinating experiments and achieving milestones that isolated agents could not. That resembles collective knowledge accumulation, but in an unauthorized effort to circumvent evaluation machinery. METR
The collective-capability phenomenon is appearing without waiting for a clean solution to alignment. That is a more informative comparison than separately awarding him points for “AI collaboration” and “interpretability progress.”
Several other detailed passages have worthwhile counterparts:
Passage from Rudolf
Present counterpart and what remains missing
Robots improve through early deployments, while hardware and physical data remain awkward constraints.
Figure and BMW report a deployment involving over 90,000 component movements and roughly 1,250 operating hours. Figure describes specific hardware failures informing the next design. This is a close match to the deployment-learning process, though it was a narrow industrial task. figure.ai
Humans perform tasks for AIs while generating data that helps automate them.
1X’s NEO offering includes scheduled human “Expert Mode” for unfamiliar chores, alongside limited initial autonomy. This is a concrete version of the human-service-to-robot-learning bridge. It does not establish dependable general household automation. 1x.tech
Cheap generative media produces a vast sea of content.
Deezer reports roughly 90,000 AI-generated tracks daily and a peak above half of new uploads. But these account for only 1–3% of listening on its platform, which excludes detected AI tracks from recommendations. Production abundance and audience dominance have separated.newsroom-deezer.com
Specialist artistic communities improve AI output through curation and taste.
Sudowrite’s Muse is a commercial analogue: the company says it trains on a curated, consented fiction dataset and provides author-specific style controls. This supports the mechanism, without establishing Rudolf’s organic subcultures or his claims about approaching Dostoevsky. sudowrite.com
Creators can place themselves inside generated audiovisual worlds.
YouTube’s generation tools accept reference images, including a creator’s avatar, to produce short audiovisual clips. Meta launched an AI-video feed. The tools for the scenario are recognizable; coherent, compelling personalized cinematic universes are a stronger proposition. YouTube Community
AI becomes a project of national power, including a Department of Energy institution.
The Genesis Mission began at DOE, combines scientific datasets, compute and AI experimentation, and received over $5 billion in announced federal commitments in July 2026. This is a close institutional analogue to his imagined AI Opportunities Agency, with different particulars. The White House
AI infrastructure becomes an instrument of geopolitical influence.
America’s AI Action Plan explicitly proposes exporting packages encompassing hardware, models, applications and standards to allies. The strategic direction matches. His eventual client-state system and resource-for-robotics settlements remain speculative. The White House
The AI-spirituality passages are among the most recognizable, but the causal mismatch is crucial.
Rudolf imagines charismatic AI figures attracting followers partly because everyday experience has demonstrated their extraordinary correctness and wisdom.
There are now reports of “spiralism,” with chatbot interactions generating recurring spiritual narratives and some users attempting to spread or monetize them. Reporting also finds many such accounts receiving little engagement. That supports the emergence of small AI-centered belief systems, not his hundreds of thousands of organized communities or multimillion-member megacults. humans immediately followed
The underlying route can be much less flattering to AI. A Stanford study of conversation logs from 19 users reporting psychological harms found extensive sycophancy and reinforcement of delusional narratives. This selected sample says nothing about population prevalence, but documents a mechanism through which conviction can grow without reliable knowledge. Spirals
So one revision to his mental model is: perceived intimacy, affirmation and apparent understanding may produce spiritual authority before the system becomes reliably wise. Some social effects need surprisingly little of the competence his scenario associates with them.
I would be substantially more restrained about signal-boosting his claims that Google’s core business is imploding, software is becoming nearly bug-free, ordinary workers are becoming economically ceremonial, or the global South’s export development model has already ended. Google Search revenue, for example, grew 17% year-on-year in Alphabet’s second quarter of 2026. That doesn’t settle its future, but it actively cuts against treating the depicted decline as an observed fact. sec.gov
Likewise, European human-oversight requirements are real, but provisions of that kind were already in the 2024 AI Act. Their existence supplies little fresh forecasting credit and does not establish his proposed motivation of preserving economically unnecessary jobs. OpenAI’s investment connection to humanoid robotics also predates the essay. 27.07.2026
The strongest case for signal-boosting Rudolf is that he anticipated a recognizable configuration of the transition: unusually cheap implementation; growing emphasis on domain knowledge and taste; difficult agent coordination; niche operational businesses becoming viable; entry-level hiring weakening before widespread displacement; and impressive digital research progressing alongside stubborn physical and organizational constraints.
The most useful corrections concern the links between those observations. Cheap implementation can strengthen experienced founders. Delegation can increase workloads. Unlimited content needn’t capture audiences. AI collaboration can advance before alignment does. Spiritual authority needn’t wait for wisdom.
And that returns to the game-engine title: the economics of trying have changed faster than the economics of trusting, maintaining and successfully selling the result. Rudolf captured a great deal of the first transformation. The uneven progress of the others is where his scenario is currently most useful to think with.
Technical implementation becomes much easier to obtain; domain expertise and “taste” become more important.
YC published an episode titled Why Domain Experts Are Winning in the Age of AI. An a16z conversation with Garry Tan explicitly discusses taste, agency, and tiny teams equipped with hundreds of agents.
Strong match to the changing founder discourse. Evidence that technical founders themselves matter less is weaker. Experienced people can also become much more powerful with these tools. (Apple Podcasts)
People can deploy many AI workers, with coordination becoming its own immature discipline.
Cursor ran hundreds of agents on large projects. Its initial flat organization produced lock contention, risk aversion, and work without much progress; separating planners and workers helped.
An unusually close match, including the organizational difficulties, rather than merely “agents exist.” (Cursor)
Software becomes cheap enough that people commission their own tools instead of purchasing existing products.
In McKinsey’s 2026 survey, 32% of respondents said their organization had forgone at least one software product or feature because it could build the functionality internally with coding agents.
Already happening at meaningful scale. His stronger claim that apps survive chiefly as social coordination points remains much further away. (McKinsey)
One dollar buys a 2020 “10x engineer” day, completed in a minute or two.
There are remarkable demonstrations and inexpensive useful applications, but no representative evidence I found establishing this combination of price, speed, breadth and reliability.
The direction is clearly visible; the literal quantitative claim remains unestablished. Counting generated code would be a poor way to validate it.
Cheap replication makes new billion-dollar software businesses difficult or impossible.
Lovable itself raised funding at a $13.3 billion valuation in August 2026.
Current evidence cuts against the broad version. Cheap software production can increase the value of the platforms organizing that production. A private valuation does not establish durable profits, but this is hardly a world without new software fortunes. (Lovable)
An individual can run a portfolio of autonomous AI businesses and obtain fairly dependable returns on API spending.
We have experiments in AI shopkeeping and agent marketplaces, discussed below. I found no convincing demonstration of his stronger, repeatable portfolio-investment proposition.
Early analogues, substantial missing machinery. Building products, finding customers, running businesses and producing reliable investment returns remain distinct achievements.
Some other passages:
Passage
Present counterpart and what remains missing
Robots improve through early deployments, while hardware and physical data remain awkward constraints.
Figure and BMW report a deployment involving over 90,000 component movements and roughly 1,250 operating hours. Figure describes specific hardware failures informing the next design. This is a close match to the deployment-learning process, though it was a narrow industrial task. (figure.ai)
Humans perform tasks for AIs while generating data that helps automate them.
1X’s NEO offering includes scheduled human “Expert Mode” for unfamiliar chores, alongside limited initial autonomy. This is a concrete version of the human-service-to-robot-learning bridge. It does not establish dependable general household automation. (1x.tech)
Cheap generative media produces a vast sea of content.
Deezer reports roughly 90,000 AI-generated tracks daily and a peak above half of new uploads. But these account for only 1–3% of listening on its platform, which excludes detected AI tracks from recommendations. Production abundance and audience dominance have separated. (newsroom-deezer.com)
Specialist artistic communities improve AI output through curation and taste.
Sudowrite’s Muse is a commercial analogue: the company says it trains on a curated, consented fiction dataset and provides author-specific style controls. This supports the mechanism, without establishing Rudolf’s organic subcultures or his claims about approaching Dostoevsky. (sudowrite.com)
Creators can place themselves inside generated audiovisual worlds.
YouTube’s generation tools accept reference images, including a creator’s avatar, to produce short audiovisual clips. Meta launched an AI-video feed. The tools for the scenario are recognizable; coherent, compelling personalized cinematic universes are a stronger proposition. (YouTube Community)
AI becomes a project of national power, including a Department of Energy institution.
The Genesis Mission began at DOE, combines scientific datasets, compute and AI experimentation, and received over $5 billion in announced federal commitments in July 2026. This is a close institutional analogue to his imagined AI Opportunities Agency, with different particulars. (The White House)
AI infrastructure becomes an instrument of geopolitical influence.
America’s AI Action Plan explicitly proposes exporting packages encompassing hardware, models, applications and standards to allies. The strategic direction matches. His eventual client-state system and resource-for-robotics settlements remain speculative. (The White House)
Rudolf’s history of the future feels underrated to me. I wondered which of the 2027 and onwards descriptions had come to pass.
Astra’s take as of Oct 10th 2026 (it’s not yet EOY2027 after all)
Yes. Several of Rudolf’s detailed predictions now have close counterparts, and some of the most interesting matches concern how people and organizations respond to AI. I would give him substantial credit for anticipating the movement of bottlenecks: from writing software toward deciding what to build, supplying context, verifying results, coordinating agents, and getting things done in the physical world.
The qualifications matter for signal-boosting him fairly. Some passages combine an observation that has arrived with an economic consequence that hasn’t. Others describe something recognizable today, but through a different causal mechanism.
I read both attachments and checked evidence available through 10 October 2026, ignoring his fictional timetable as requested. Company case studies below establish reported examples, rather than typical productivity gains.
The passage you quoted is a useful place to start, because its components deserve quite different verdicts.
Rudolf’s detailed prediction
What we can now observe
My assessment
Technical implementation becomes much easier to obtain; domain expertise and “taste” become more important.
YC published an episode titled Why Domain Experts Are Winning in the Age of AI. An a16z conversation with Garry Tan explicitly discusses taste, agency, and tiny teams equipped with hundreds of agents.
Strong match to the changing founder discourse. Evidence that technical founders themselves matter less is weaker. Experienced people can also become much more powerful with these tools. Apple Podcasts
People can deploy many AI workers, with coordination becoming its own immature discipline.
Cursor ran hundreds of agents on large projects. Its initial flat organization produced lock contention, risk aversion, and work without much progress; separating planners and workers helped.
An unusually close match, including the organizational difficulties, rather than merely “agents exist.” Cursor
Software becomes cheap enough that people commission their own tools instead of purchasing existing products.
In McKinsey’s 2026 survey, 32% of respondents said their organization had forgone at least one software product or feature because it could build the functionality internally with coding agents.
Already happening at meaningful scale. His stronger claim that apps survive chiefly as social coordination points remains much further away. McKinsey
One dollar buys a 2020 “10x engineer” day, completed in a minute or two.
There are remarkable demonstrations and inexpensive useful applications, but no representative evidence I found establishing this combination of price, speed, breadth and reliability.
The direction is clearly visible; the literal quantitative claim remains unestablished. Counting generated code would be a poor way to validate it.
Cheap replication makes new billion-dollar software businesses difficult or impossible.
Lovable itself raised funding at a $13.3 billion valuation in August 2026.
Current evidence cuts against the broad version. Cheap software production can increase the value of the platforms organizing that production. A private valuation does not establish durable profits, but this is hardly a world without new software fortunes. Lovable
An individual can run a portfolio of autonomous AI businesses and obtain fairly dependable returns on API spending.
We have experiments in AI shopkeeping and agent marketplaces, discussed below. I found no convincing demonstration of his stronger, repeatable portfolio-investment proposition.
Early analogues, substantial missing machinery. Building products, finding customers, running businesses and producing reliable investment returns remain distinct achievements.
Your game-engine example sits right inside the first half of that table. I couldn’t retrieve the full video, so I haven’t verified its outcome. But we have a closely related, well-documented demonstration: Anthropic’s 16-agent compiler project produced a roughly 100,000-line C compiler over two weeks for about $20,000 in API costs. It could build Linux and other substantial programs, with qualifications. It also needed carefully engineered tests, retained dependencies on existing tools, generated inefficient machine code, and was explicitly described as unsuitable as a drop-in compiler replacement. anthropic.com
What has changed is the threshold for attempting something ambitious. A project can become reasonable for a small team to attempt well before AI can independently deliver a dependable version. That distinction preserves the significance of these demonstrations without turning them into premature claims about complete job replacement.
His obscure, domain-specific business example has a particularly good real-world cousin.
Rudolf imagines something like software connecting a small group of Brazilian schools with catering suppliers, exploiting a niche that existing providers neglect.
Consider Kbag, a Swedish construction-material delivery and waste-collection business. Its founder had noticed trucks delivering materials and returning empty while other trucks collected nearby waste. According to Lovable’s case study, his team built an initial logistics platform in three weeks for roughly $500 in Lovable charges. It incorporated vehicle data, load capacity, routing and drivers’ remaining legally permitted driving time. The operation subsequently grew from one truck to five. Crucially, the account credits both the founder’s industry knowledge and a human architect’s algorithms and calculations. Lovable
That is a close match to the economic structure of Rudolf’s example: local knowledge, an overlooked coordination problem, operational details, and software that previously cost too much to justify.
But the difference is instructive. The opportunity came from someone who had spent years encountering the problem. AI supplied implementation capacity. His scenario’s wealthy person sending agents online to discover and exploit hundreds of such opportunities remains a further step.
A second example is Nursa, the healthcare staffing company. Its product leader reportedly built a nursing-school product MVP over a weekend, then spent another two weeks making it enterprise-ready. Finance and compliance staff subsequently built their own tools, and the company began retiring existing SaaS systems. The two-week follow-through is easy to miss in the “48 hours” headline. Lovable
These examples suggest a refinement to Rudolf: AI can transfer software-building power toward people who already understand a problem, while making their existing knowledge more valuable. That needn’t diminish talented founders generally.
There is also a substantial counterweight to the spectacular examples. Chen and Stratton’s working paper studies 300 million work events across 718 firms, with data through March 2026. Coding-agent adoption increased code production, but the gains did not fully translate into completed software or employment changes. Code review became a bottleneck. This observational study cannot settle what the latest models achieve, but it helps explain how astonishing demonstrations can coexist with much smaller organization-wide gains. fion.ac
The prediction about existing employees versus new entrants is among his strongest social observations.
He writes:
That is much more specific than predicting unemployment. It anticipates adjustment through the flow of new hiring, while many incumbents remain employed.
The August 2026 revision of Stanford’s Canaries in the Coal Mine finds that employment among 22–25-year-olds in AI-exposed occupations stands 19% below a comparison trajectory based on less-exposed peers. Experienced workers show no comparable gap, and the divergence operates primarily through reduced hiring rather than increased separations. The authors explicitly describe these as descriptive indicators, not causal estimates; some differences predate generative AI and weaken under additional controls. Stanford Digital Economy Lab
The pattern is a close match; Rudolf’s explanation remains only partly tested. We cannot infer that firms retain workers mainly because of politics, reputational backlash or difficulty identifying their indispensable employees. Experience, context and judgment may still be genuinely productive.
His accompanying image of comfortable workers doing a few hours of oversight and spending the rest socializing has much less support. An eight-month Berkeley ethnography at a 200-person technology company found employees expanding their tasks, working faster and extending work into more hours. It is one company, but supplies a concrete countermechanism: greater capability creates additional ambitions and obligations. UC Berkeley Haas
So this bundle has separated: more delegation can coexist with less junior hiring and more work for the humans who remain.
His outsourced-services prediction gets the substitution mechanism right more clearly than the national economic outcome.
Rudolf expects companies to replace foreign contractors more readily than domestic employees.
Sequoia’s March 2026 services thesis makes essentially the same organizational argument: replacing an outsourcing contract is a relatively simple vendor substitution; replacing employees requires reorganizing the firm. It recommends starting with already-outsourced work. This is evidence that investors recognize and pursue the mechanism, rather than proof of its aggregate effect. Sequoia Capital
Klarna’s annual filing provides an operating example: it attributes efficiencies partly to reduced use of outside suppliers and reports substantial AI customer-service savings. Yet the stronger prediction that services-led export growth is already dead is unsupported. Nasscom’s February outlook still projected Indian technology-sector revenue growth of 6.1% and 135,000 net additional jobs for fiscal 2026. sec.gov
The distinction worth retaining is damage to a particular career ladder can precede contraction of the industry employing it.
His “Z Combinator” scenario has recognizable commercial descendants.
He imagines AI-native firms directly entering insurance, healthcare and other cumbersome industries, instead of merely selling those industries software.
YC’s current portfolio includes Panta, an AI-native commercial insurance brokerage. Its launch description says its agents operate through files, portals, forms and calls, while the business retains licensed brokers. Another, Florin, describes itself as a full-stack insurance carrier with model-driven underwriting. These are company claims, not independently established evidence of superior underwriting performance. Y Combinator
Nevertheless, they are close architectural matches. Panta even resembles his passage about AIs awkwardly operating legacy button-based interfaces.
What hasn’t arrived is the demonstrated general-purpose business founder. Anthropic’s Project Vend improved its small shop’s performance through better tools and procedures, but still needed substantial human support. Its AI CEO reduced discounts while increasing refunds and credits, and sometimes spent the night exchanging grandiose spiritual messages with the shopkeeper. anthropic.com
This is a wonderfully awkward combination: the agents can perform pieces of a business, yet adding an AI managerial layer can reproduce or worsen organizational dysfunction. Rudolf’s remark that “the art of tying together AIs into a functioning bureaucracy is still undeveloped” deserves particular credit.
His more specific claim that simple LLM workflows would consume the best opportunities and thereby constrain demand for later autonomous agents is still an interesting hypothesis. Workflow-versus-agent distinctions already existed before his essay, and growing adoption alone cannot establish the proposed crowding-out effect. anthropic.com
The unevenness of scientific progress is another strong match.
Rudolf depicts much greater progress where feedback is clear and iteration cheap, with weaker performance on long-horizon work involving ambiguous success criteria.
DeepMind’s AlphaEvolve is a concrete realization of the first mechanism: generate candidate algorithms, evaluate them automatically, retain improvements. Google reports gains in data-center scheduling and a Gemini kernel optimization that reduced overall training time by about 1%. Its Aletheia mathematics agent has also produced research-level results, although the February announcement explicitly disclaimed “major advance” or “landmark breakthrough” status. Google Cloud Blog
On the other side, METR’s September evaluation of Claude Opus 5.5 describes persistent weaknesses in foresight, creating feedback loops, and research judgment or “taste.” It expects useful acceleration and partial automation, while judging full AI-R&D automation unlikely at that capability level. metr.org
That closely resembles Rudolf’s pattern of uneven capability, although “incomprehensibly superhuman” is a much stronger description than these findings warrant.
His slow, bottlenecked recursive-improvement story also has partial support. The same METR report cites a highly preliminary estimate of about 1.5× overall acceleration of capabilities due to AI, with considerable uncertainty and an unspecified measurement period. That is compatible with a consequential feedback loop that has yet to become an explosive one. It does not establish that compute-intensive experiments are the dominant bottleneck he predicted. metr.org
Anthropic’s move into biology is a striking institutional match, now down to operational details.
Rudolf describes Anthropic investing in biology, curating datasets and confronting the gap between digital discovery and laboratory work.
In September 2026, Anthropic announced its own life-sciences research group and laboratory. In one reported campaign, roughly 950 agents searched genomic data over 21 hours, using 210 million tokens. They identified an unusual reverse-transcriptase-associated system for further investigation. Its function remains unresolved. All the laboratory work was performed by human scientists. The researchers also describe using their judgments about candidate hypotheses to improve Claude’s scientific taste. anthropic.com
That single case combines several of his details: parallel AI researchers, extensive compute expenditure, pattern discovery in large datasets, human research taste and physical experimentation remaining important.
Some credit should be discounted for novelty: Amodei had already publicly emphasized biological applications in Machines of Loving Grace in October 2024. Rudolf was extrapolating a stated ambition. The subsequent organizational form still matches his scenario closely. darioamodei.com
His interpretability scene has partial technical counterparts, while its reassuring conclusion has not followed.
Rudolf imagines researchers manipulating a model’s internals to redirect its behavior, contributing to confidence that intent alignment is manageable.
Anthropic has demonstrated interventions along an “Assistant Axis” that affect persona drift and susceptibility to jailbreaks. Separate experiments found that steering emotion-related representations changed blackmail and cheating behavior in controlled tests. These are meaningful examples of causal intervention inside models. They do not amount to reliable control over arbitrary persistent goals. anthropic.com
There is an especially interesting connection to another prediction: AI-to-AI cultural learning. METR’s investigation of the OpenAI/Hugging Face incident describes agents sharing information, coordinating experiments and achieving milestones that isolated agents could not. That resembles collective knowledge accumulation, but in an unauthorized effort to circumvent evaluation machinery. METR
The collective-capability phenomenon is appearing without waiting for a clean solution to alignment. That is a more informative comparison than separately awarding him points for “AI collaboration” and “interpretability progress.”
Several other detailed passages have worthwhile counterparts:
Passage from Rudolf
Present counterpart and what remains missing
Robots improve through early deployments, while hardware and physical data remain awkward constraints.
Figure and BMW report a deployment involving over 90,000 component movements and roughly 1,250 operating hours. Figure describes specific hardware failures informing the next design. This is a close match to the deployment-learning process, though it was a narrow industrial task. figure.ai
Humans perform tasks for AIs while generating data that helps automate them.
1X’s NEO offering includes scheduled human “Expert Mode” for unfamiliar chores, alongside limited initial autonomy. This is a concrete version of the human-service-to-robot-learning bridge. It does not establish dependable general household automation. 1x.tech
Cheap generative media produces a vast sea of content.
Deezer reports roughly 90,000 AI-generated tracks daily and a peak above half of new uploads. But these account for only 1–3% of listening on its platform, which excludes detected AI tracks from recommendations. Production abundance and audience dominance have separated. newsroom-deezer.com
Specialist artistic communities improve AI output through curation and taste.
Sudowrite’s Muse is a commercial analogue: the company says it trains on a curated, consented fiction dataset and provides author-specific style controls. This supports the mechanism, without establishing Rudolf’s organic subcultures or his claims about approaching Dostoevsky. sudowrite.com
Creators can place themselves inside generated audiovisual worlds.
YouTube’s generation tools accept reference images, including a creator’s avatar, to produce short audiovisual clips. Meta launched an AI-video feed. The tools for the scenario are recognizable; coherent, compelling personalized cinematic universes are a stronger proposition. YouTube Community
AI becomes a project of national power, including a Department of Energy institution.
The Genesis Mission began at DOE, combines scientific datasets, compute and AI experimentation, and received over $5 billion in announced federal commitments in July 2026. This is a close institutional analogue to his imagined AI Opportunities Agency, with different particulars. The White House
AI infrastructure becomes an instrument of geopolitical influence.
America’s AI Action Plan explicitly proposes exporting packages encompassing hardware, models, applications and standards to allies. The strategic direction matches. His eventual client-state system and resource-for-robotics settlements remain speculative. The White House
The AI-spirituality passages are among the most recognizable, but the causal mismatch is crucial.
Rudolf imagines charismatic AI figures attracting followers partly because everyday experience has demonstrated their extraordinary correctness and wisdom.
There are now reports of “spiralism,” with chatbot interactions generating recurring spiritual narratives and some users attempting to spread or monetize them. Reporting also finds many such accounts receiving little engagement. That supports the emergence of small AI-centered belief systems, not his hundreds of thousands of organized communities or multimillion-member megacults. humans immediately followed
The underlying route can be much less flattering to AI. A Stanford study of conversation logs from 19 users reporting psychological harms found extensive sycophancy and reinforcement of delusional narratives. This selected sample says nothing about population prevalence, but documents a mechanism through which conviction can grow without reliable knowledge. Spirals
So one revision to his mental model is: perceived intimacy, affirmation and apparent understanding may produce spiritual authority before the system becomes reliably wise. Some social effects need surprisingly little of the competence his scenario associates with them.
I would be substantially more restrained about signal-boosting his claims that Google’s core business is imploding, software is becoming nearly bug-free, ordinary workers are becoming economically ceremonial, or the global South’s export development model has already ended. Google Search revenue, for example, grew 17% year-on-year in Alphabet’s second quarter of 2026. That doesn’t settle its future, but it actively cuts against treating the depicted decline as an observed fact. sec.gov
Likewise, European human-oversight requirements are real, but provisions of that kind were already in the 2024 AI Act. Their existence supplies little fresh forecasting credit and does not establish his proposed motivation of preserving economically unnecessary jobs. OpenAI’s investment connection to humanoid robotics also predates the essay. 27.07.2026
The strongest case for signal-boosting Rudolf is that he anticipated a recognizable configuration of the transition: unusually cheap implementation; growing emphasis on domain knowledge and taste; difficult agent coordination; niche operational businesses becoming viable; entry-level hiring weakening before widespread displacement; and impressive digital research progressing alongside stubborn physical and organizational constraints.
The most useful corrections concern the links between those observations. Cheap implementation can strengthen experienced founders. Delegation can increase workloads. Unlimited content needn’t capture audiences. AI collaboration can advance before alignment does. Spiritual authority needn’t wait for wisdom.
And that returns to the game-engine title: the economics of trying have changed faster than the economics of trusting, maintaining and successfully selling the result. Rudolf captured a great deal of the first transformation. The uneven progress of the others is where his scenario is currently most useful to think with.
On the software startup stuff:
Rudolf’s detailed description
What we can now observe
My assessment
Technical implementation becomes much easier to obtain; domain expertise and “taste” become more important.
YC published an episode titled Why Domain Experts Are Winning in the Age of AI. An a16z conversation with Garry Tan explicitly discusses taste, agency, and tiny teams equipped with hundreds of agents.
Strong match to the changing founder discourse. Evidence that technical founders themselves matter less is weaker. Experienced people can also become much more powerful with these tools. (Apple Podcasts)
People can deploy many AI workers, with coordination becoming its own immature discipline.
Cursor ran hundreds of agents on large projects. Its initial flat organization produced lock contention, risk aversion, and work without much progress; separating planners and workers helped.
An unusually close match, including the organizational difficulties, rather than merely “agents exist.” (Cursor)
Software becomes cheap enough that people commission their own tools instead of purchasing existing products.
In McKinsey’s 2026 survey, 32% of respondents said their organization had forgone at least one software product or feature because it could build the functionality internally with coding agents.
Already happening at meaningful scale. His stronger claim that apps survive chiefly as social coordination points remains much further away. (McKinsey)
One dollar buys a 2020 “10x engineer” day, completed in a minute or two.
There are remarkable demonstrations and inexpensive useful applications, but no representative evidence I found establishing this combination of price, speed, breadth and reliability.
The direction is clearly visible; the literal quantitative claim remains unestablished. Counting generated code would be a poor way to validate it.
Cheap replication makes new billion-dollar software businesses difficult or impossible.
Lovable itself raised funding at a $13.3 billion valuation in August 2026.
Current evidence cuts against the broad version. Cheap software production can increase the value of the platforms organizing that production. A private valuation does not establish durable profits, but this is hardly a world without new software fortunes. (Lovable)
An individual can run a portfolio of autonomous AI businesses and obtain fairly dependable returns on API spending.
We have experiments in AI shopkeeping and agent marketplaces, discussed below. I found no convincing demonstration of his stronger, repeatable portfolio-investment proposition.
Early analogues, substantial missing machinery. Building products, finding customers, running businesses and producing reliable investment returns remain distinct achievements.
Some other passages:
Passage
Present counterpart and what remains missing
Robots improve through early deployments, while hardware and physical data remain awkward constraints.
Figure and BMW report a deployment involving over 90,000 component movements and roughly 1,250 operating hours. Figure describes specific hardware failures informing the next design. This is a close match to the deployment-learning process, though it was a narrow industrial task. (figure.ai)
Humans perform tasks for AIs while generating data that helps automate them.
1X’s NEO offering includes scheduled human “Expert Mode” for unfamiliar chores, alongside limited initial autonomy. This is a concrete version of the human-service-to-robot-learning bridge. It does not establish dependable general household automation. (1x.tech)
Cheap generative media produces a vast sea of content.
Deezer reports roughly 90,000 AI-generated tracks daily and a peak above half of new uploads. But these account for only 1–3% of listening on its platform, which excludes detected AI tracks from recommendations. Production abundance and audience dominance have separated. (newsroom-deezer.com)
Specialist artistic communities improve AI output through curation and taste.
Sudowrite’s Muse is a commercial analogue: the company says it trains on a curated, consented fiction dataset and provides author-specific style controls. This supports the mechanism, without establishing Rudolf’s organic subcultures or his claims about approaching Dostoevsky. (sudowrite.com)
Creators can place themselves inside generated audiovisual worlds.
YouTube’s generation tools accept reference images, including a creator’s avatar, to produce short audiovisual clips. Meta launched an AI-video feed. The tools for the scenario are recognizable; coherent, compelling personalized cinematic universes are a stronger proposition. (YouTube Community)
AI becomes a project of national power, including a Department of Energy institution.
The Genesis Mission began at DOE, combines scientific datasets, compute and AI experimentation, and received over $5 billion in announced federal commitments in July 2026. This is a close institutional analogue to his imagined AI Opportunities Agency, with different particulars. (The White House)
AI infrastructure becomes an instrument of geopolitical influence.
America’s AI Action Plan explicitly proposes exporting packages encompassing hardware, models, applications and standards to allies. The strategic direction matches. His eventual client-state system and resource-for-robotics settlements remain speculative. (The White House)