The Machines Are Taking Our Jobs - Thank God? Emad Mostaque’s Guide to the next 1000 Days
Summary
Mostaque’s near-term thesis is that “useful intelligence,” not AGI, breaks labor economics first. The decisive systems will be reliable “cooks” that follow instructions across keyboard-video-mouse jobs, not frontier “chefs” inventing new science. He expects virtual workers costing roughly a dollar an hour, an 8-billion-parameter medical model running on a smartphone, and agents operating for hours without supervision to make the next thousand days radically different even if capabilities soon plateau.
AI’s cost curve creates an “abundance trap”: intelligence becomes plentiful while a scarcity-based economy records the result as unemployment and poverty. GPT-3 input reportedly cost $60 per million tokens versus roughly $1.25-$1.50 for GPT-5, while a website that once cost thousands can already be generated for $20-$40. Because GPUs “don’t need to eat,” buy housing, or consume their wages, rising AI output need not recycle demand, tax revenue, and employment through the economy.
The “intelligence inversion” leaves human labor with no obvious higher rung to climb after machines outperform both muscle and cognition. Mostaque traces economic advantage from land and serfs, through physical labor and industrial or software capital, to intelligence itself; once AI becomes the marginal producer, “there’s nowhere left to pivot.” Nathan Labenz’s electricity calculation sharpens the threat: a laptop can run for hours on less than two cents of power in his Detroit market.
The distributional upside therefore accrues first to GPU owners, frontier laboratories, and data flywheels—not automatically to workers or consumers. Mostaque warns that OpenAI, Anthropic, and xAI need not keep their best models available through APIs; once internal systems outperform public ones, vertically integrated labs could compete with every application and ultimately “take on the entire economy themselves.” He sees Grok 5 versus Grok 4 as a near-term scaling test, while stressing that even a plateau around today’s frontier would still disrupt employment.
Human care, community, and creativity remain meaningful, but they do not solve the transition’s monetary arithmetic. Mostaque accepts that people may intrinsically prefer human teachers, nurses, friends, and performers, yet estimates that paying every American adult the roughly $16,000 poverty threshold would cost $5 trillion—the entire US tax base—while corporate taxes contribute only about $0.9 trillion. His distinction is that “computation and consciousness are different”: AI can supply the how, but humans still create the why, provided they retain income, identity, community, and attention.
His alternative scorecard replaces GDP-only optimization with four multiplicative “MIND” capitals: material, intelligence, network, and diversity. Material goods such as apples are rivalrous and depleted by consumption; knowledge can circulate without being lost, networks determine trust and coordination, and diversity supplies resilience and optionality. Together with laws of flow, openness, and resilience, the framework argues that maximizing profits or GDP while degrading social connectivity and redundancy makes systems look efficient precisely as they become brittle.
The proposed counterweight to digital feudalism is collectively controlled AI infrastructure financed through a Bitcoin-like “FoundationCoin.” Verified deployments of open models in health, education, finance, and government would secure the asset, while sale proceeds would fund public-interest compute such as multilingual diagnosis checking and organized cancer knowledge. Mostaque is explicit that the second currency—a cash-like issuance tied to being human—“we haven’t quite figured out yet,” making this a direction and incentive-design experiment rather than a completed monetary system.
The thousand-day urgency comes from defaults hardening into control: whoever owns users’ memories, data, interfaces, and GPU capacity may become almost impossible to dislodge. Mostaque’s preferred outcome is neither one corporate singleton nor nationally fragmented AI, but a loosely coupled swarm whose objective is human flourishing. His closing choice is stark: use AI “to increase the nature of our agency versus replace us with agents,” so that machines taking jobs frees people for family, care, art, exploration, and “our real work.”
Deep dive
1. Intelligence begins with surviving surprise
Mostaque’s starting axiom is persistence: complex adaptive systems survive uncertain environments when their internal models most closely match reality. Intelligence theory therefore treats intelligence as minimizing the loss—or “surprise”—between prediction and the world, using mathematics analogous to the loss functions of generative AI.
Labenz grounds the claim in predictive processing: each neural layer predicts incoming signals, stays quiet when reality matches, and escalates discrepancies toward conscious attention. His objection is the thousand-year-old rock: if persistence indicates intelligence, why does an inert object seemingly persist longer than an 80-year-old human?
Mostaque limits the framework to agents that act and adapt. Their burden decomposes into predictive error, model complexity—the “cost of thinking”—and update cost, while a rock needs no computation or learning. He connects this to Karl Friston’s free-energy principle and asks economics to replace 200-year-old assumptions about human producers with “the wealth of robots.”
2. Slow human updating hides an extraordinarily fast capability shift
Asked why Tyler Cowen can envision only an additional half-point of annual GDP growth, Mostaque answers with “the update cost”: radical evidence is expensive to absorb. His pandemic analogy is that warnings remained abstract until Tom Hanks became infected and triggered a phase shift in public perception.
The chronology is his evidence for cognitive lag: just over three years since Stable Diffusion, roughly a thousand days since ChatGPT, one year since o1 preview, and just over a month since GPT-5. Until that last month, he says, most global AI users were still using GPT-4o; he cites an estimate that 20% of Americans have never heard of ChatGPT.
Frontier users now encounter a different product from the instant, hallucination-prone models anchoring public expectations. Mostaque describes leaving Codex running from the CLI for three hours to construct a textbook website: “set it and forget it.” Agents can increasingly inspect screens, check errors, adapt, and continue working rather than merely answer prompts.
His sharpest small-model example is Alibaba’s Tongyi Qwen Lab releasing a model with roughly 3 billion active parameters and 30 billion total that, he says, beat Grok 4 on Humanity’s Last Exam and Deep Research systems. That makes frontier-like capability runnable on a CPU with 16 GB of RAM rather than only in a giant data center.
3. Reliable cooks matter more economically than genius chefs
Mostaque separates general intelligence from “actually useful intelligence.” The economy does not require every worker to be a polymath; it requires vast numbers of dependable executors. Borrowing the cooks-versus-chefs framing, he argues: “The cooks are the ones that will impact the economy.”
His iMed model has 8 billion parameters, needs roughly 8 GB of RAM, and, he says, already outperforms human doctors—though he adds, “We don’t think it’s good enough yet.” His categorical projection is that by next year it will beat any doctor with full traceability and run on any smartphone.
Labenz adds the price curve: GPT-3 cost about $60 per million input tokens, versus roughly $1.25-$1.50 for GPT-5—a reduction above 95% alongside much higher capability. Hedonic GDP adjustments struggle when a $100 medical appointment becomes a ten-cent consultation rather than merely a slightly improved version of the old service.
4. Abundant intelligence can register as mass poverty
The “abundance trap” is Mostaque’s name for intelligence reaching post-scarcity while economic institutions continue interpreting value through scarcity. Patients receive better advice and companies produce more, but displaced knowledge workers lose wages; the same event can raise real capability and corporate profit while appearing as poverty to households.
The “metabolic rift” explains why the gains do not naturally recycle. AI accountants, lawyers, and designers need no food, housing, rest, or consumption budget; GPU usage may even be tax-deductible. Once a system becomes smarter or more capable than a human, “it’s not going to get dumber ever.”
Labenz supplies physical intuition: a phone battery stores about 20 watt-hours and a laptop around 100; at approximately 18 cents per kilowatt-hour in Detroit, a full laptop charge costs under two cents. The resulting Malthusian competition is not between a salary and a large energy bill, but a salary and pennies of electricity.
Creation costs are already collapsing: Mostaque contrasts a website once costing thousands with a Replit agent producing one for $20-$40, including margin, and projects another tenfold decline next year and again the year after. “The cost of consumption in the previous internet age went to zero. Now the cost of creation is going to zero.”
5. The intelligence inversion leaves no conventional labor refuge
Mostaque’s historical stack runs from land and its serfs, to muscular labor, to industrial and software capital, and finally to intelligence. Humans survived earlier automation by moving upward; if machines now outperform cognition and soon manipulate the physical world, he asks, “Where do we pivot now?”
Robots intensify the conclusion rather than create it. He points to a Unitree robot recovering within a second after being pushed, robots making recipes, and a hypothetical $20,000 Tesla Optimus working continuously for roughly $150 an hour. Plumbing and other physical trades may follow knowledge work within a few years.
Labenz tests whether care, teaching, mentoring, and creativity form the next refuge. Mostaque concedes that human concerts, walks with one’s children, nursing, and education retain intersubjective value; his objection is not that care disappears, but that today’s capital flows do not give displaced people the freedom to perform it.
6. Community can replace job identity only if money still reaches people
Mostaque says conventional monetary transmission itself may break. Lower interest rates once encouraged companies to borrow, hire, and pay wages that supported consumption; increasingly, cheaper capital may purchase more GPUs instead. Productive capacity rises without restoring the employment side of the Federal Reserve’s employment-and-inflation framework.
The US arithmetic makes “tax the AI” inadequate on his assumptions: paying every adult the roughly $16,000 poverty threshold would cost about $5 trillion, equal to the entire tax base, while corporations contribute only around $0.9 trillion. “We need to rethink how money flows,” not simply enlarge a familiar transfer program.
Job loss also removes identity, community, purpose, and structure. Neighbors, religion, and local networks have weakened while occupational and brand identities expanded; a fellow Apple owner will not necessarily help, though “a fellow Swiftie might.” People with strong communities can fall back on them, while isolated workers cannot.
7. Human attention is the remaining scarce asset
Mostaque’s philosophical distinction is that computation and consciousness were once bundled in humans but are now separating. Machines can increasingly answer how; people remain “meaning makers” who decide why something is beautiful, just, or worthwhile. The scarce input is finite human attention and where systems direct it.
His fisherman-and-investment-banker parable captures the inversion. The banker urges a fisherman to work longer, buy boats, scale, and list a company so he can eventually retire by the beach, eat fish, dance, and see his family—exactly what the fisherman was already about to do.
Mostaque mentions a figure he thinks Sam Altman gave—about 10,000 people committing suicide each month—and asks how many of them had talked to ChatGPT. He believes ChatGPT has probably reduced suicides overall, even as some people have committed suicide because of it.
A trusted person’s AI double might provide extraordinary continuity and coaching, but it could also become more trusted than any living person. Mostaque frames the architectural choice as increasing human agency and connection or accepting a WALL-E future of headsets, passive consumption, and robots running everything.
8. Today’s record markets coexist with signs of systemic fracture
Mostaque invokes Neil Howe’s predicted Fourth Turning around 2025: critical slowing, greater variance from small shocks, unsustainable debt, and flickering categories such as employee versus gig worker or Bitcoin versus money. He predicts AI will be huge this year and calls next year’s US digital-asset expansion potentially “the biggest bubble we’ve probably ever seen.”
Rising correlations expose hidden fragility: one Ever Given blockage can disrupt global supply, while all-time-high equities, margins, and profits coexist with depression, suicide, and widespread dissatisfaction. Classical assumptions—that scarcity is fundamental, human labor has positive value, markets equilibrate, and money measures value—are reaching their limits.
Labenz finds the recovery pattern especially troubling: systems approaching crisis take longer to recover from new insults, just as recessions appear harder to exit. His interpretation is that economies traded resilience for extreme efficiency, leaving fewer buffers against the next perturbation.
The AI doctor makes the value problem concrete. If a consultation delivering comparable or better advice falls from $100 to ten cents, measured expenditure could collapse even though access and utility rise. GDP may therefore fall precisely because an important service became abundant.
9. GDP optimization turns corporations into brittle, antihuman AIs
Mostaque describes corporations as “slow dumb AIs” that optimize measured objectives and consume humans as inputs. Goodhart’s law then bites: organizations adapt to the metric itself. GDP, developed by Simon Kuznets in the 1940s, became the governing target despite Kuznets warning that it was a poor measure of economic or social well-being.
Offshoring and attention experiments follow the objective rather than human flourishing. Mostaque cites Meta testing whether showing people sadder material makes them sadder and post more sadness; corporate profit can improve while psychological health, local capacity, and social trust deteriorate.
Legibility and monoculture make optimization easier but remove fallbacks. Villages, diverse supply chains, and redundant practices are bulldozed into uniform systems that fail together; Mostaque’s absurd comparison is that the Los Angeles-San Francisco railway may have consumed more money than all AI-model training to date.
10. Living systems need flow, openness, and resilience
The first law is flow: value must circulate. When money, intelligence, or other capital is hoarded, the system enters stasis and eventually collapses; conservation without circulation is not health.
The second law is openness because “connection fights entropy.” Mostaque uses Tokugawa Japan’s isolation from 1633 to 1853 as an example of a closed system becoming dangerously unprepared for an external shock such as Commodore Perry’s cannons.
The third law is resilience through diversity rather than mere connectivity. Potato and banana collapses illustrate monoculture risk: a system may look efficient in normal conditions while possessing no alternative path once its dominant component fails.
11. MIND capital measures what GDP leaves invisible
Material capital is the familiar rivalrous stock: if Mostaque gives Labenz an apple, he has one less, and consumption destroys it. GDP is built to record this gradient-like flow but poorly captures capabilities, knowledge, relationships, and optionality.
Intelligence capital is non-rivalrous: sharing an idea need not deplete it and may increase its value. Mostaque cites Eric Beinhocker’s GDP-B framework and a possible $96 trillion addition from recognizing digital and intangible benefits that conventional accounts omit.
Network capital records position, connection, and trust—who will answer a request and whom one can rely upon. Diversity capital supplies adaptability and alternative routes through a phase transition. The four capitals are multiplicative: “If any of those are zero, you’re screwed.”
Singapore represents a relatively balanced MIND portfolio, while the resource curse describes material abundance without corresponding intelligence, networks, openness, or diversity. Labenz extends the concern to AI: models trained on similar data may converge on a common latent space and inherit common failure modes.
12. Generative-AI mathematics becomes a map of firms and markets
Mostaque explains diffusion through Stable Diffusion: progressively add noise until an ordered image is destroyed, then learn to reverse the process from a prompt and seed. He argues markets perform an analogous reconstruction as actors turn incomplete information into internal models and decisions to buy, sell, or build.
Organizations resemble transformers: they ingest comparatively structured information, decide what deserves attention, and compress it into a latent organizational model. Markets more closely resemble diffusion or world-model processes; Mostaque says the same equations yield the laws of living systems and the MIND-capital decomposition.
A Hodge decomposition produces three value flows. Adam Smith emphasizes gradient flows in scarce material goods; Marx’s M-C-M′ emphasizes circular, compounding accumulation; Hayek and Douglas North emphasize harmonic flow—the institutional landscape and “rules of the game” shaping every path.
The schools are blind scholars touching different parts of an elephant: trunk as hose, tail as mop, tusk as spear. Neoliberal capitalism was “the worst of all systems except for the rest,” but AI requires a holistic picture because its marginal producer neither consumes like a human nor thinks in scarcity.
13. Three futures compete: feudalism, fragmentation, or symbiosis
Labenz illustrates compounding with Stripe’s payments model: processing roughly 1.3% of global GDP supplies fraud data, better fraud detection attracts more users, and new scale improves the model again. Such intelligence flywheels can make an incumbent progressively harder to challenge.
Digital feudalism results if a few corporate AGIs own those loops and extract rents. Fragmentation produces Chinese, British, and American AI stacks, sovereign firewalls, and incompatible information environments as governments recognize that defaults can reshape citizens’ minds.
Mostaque’s preferred symbiosis is decentralized, optimized for human flourishing, and collectively controls the consequential layers. He worries far less about a creative-writing singleton than about who runs education, health, finance, procurement, and government—the domains where objective functions become social power.
14. Frontier labs may stop selling the intelligence they need themselves
Mostaque points to the escalating capital moat, including OpenAI’s announced $30 billion Stargate UK investment. If scaling continues, the world’s most capable entities become owners of the largest GPU clusters, while schools, factories, and universities cease to be the decisive productive capital stock.
Labenz’s pushback—worth keeping—is that society has no disclosure regime requiring labs to reveal internal models, observed behavior, or even whether public APIs match first-party capability. Startups such as Cursor remain viable while parity holds; one superior private model could reverse their adoption and revenue almost overnight.
Mostaque cites projections toward a $200 billion OpenAI revenue run rate in which API revenue shrinks while ChatGPT approaches roughly $80 billion and new products and agents together approach another roughly $80 billion. An agent product is directly “a replacement for human workers,” and vertically operating those workers may offer higher margins than selling API access.
Internal divergence may also be physically unavoidable. GPT-4.5 reportedly cost $150 per million tokens, “if I remember right,” while internal models may require 72 chips or more to run; Mostaque later describes 72- or 144-chip configurations for large active models. He compares that gap with OpenAI’s IMO gold-medal model, which labs may reserve for their own strategy, coding, and market operations rather than serve to every user.
15. Grok 5 is Mostaque’s next scaling-law referendum
Asked what could resolve the debate within three to six months, Mostaque chooses Grok 5. Grok 4 was already a major run; if Grok 5 delivers another clear capability jump, Elon Musk is unlikely to hide it. Labenz notes Musk had just suggested Grok 5 might be AGI, reportedly his first such claim.
Mostaque’s conditional is sharp: if classical scaling continues into next year’s clusters, “next year is the year where we break AGI full stop.” If it plateaus, the future changes substantially—though he still expects benchmarks to saturate by 2030 and today’s systems to disrupt the economy.
The competing architecture is an ecology of small specialists. Qwen’s low-active-parameter model makes continuous learning easier than for trillion-parameter behemoths, while verifiers and multi-agent scaffolds may let many cheap models outperform one giant system. The open question is “a thousand small models or one big model.”
His own base case remains hedged: scaling probably forms an S-curve and intelligence costs collapse toward zero, yet he expects GPT-5 Pro-level edge models within two years. Agents already work for hours with improving performance; if extending runtime and coordination is sufficient, “120 IQ for every human” arrives without a singular omniscient model.
16. Open-source control matters most at the interface and memory layers
Mostaque says he “used to be an open-source maximalist” but revised the position. He may accept ChatGPT supplying a component of his child’s tutor; he cares much more that an aligned entity—ideally the family—controls the education application, memory, data, and decisions around that model.
His desired topology is hierarchical and loosely bound: “swarm intelligence, not Borg.” Narrow models and agents can reinforce each person and community, while plural control creates resilience against a singleton and preserves local adaptation.
Vitalik Buterin’s “revenue-evil curve” describes the financing problem: services begin open, then premium exclusion and rivalry push organizations toward extraction. A model optimized for engagement or profit is ultimately optimized for manipulation, not well-being, regardless of later attempts to align its behavior.
Shared ethics also cannot mean one global cultural default. Japanese, German, religious, family, and community consensuses differ; Mostaque would rather curate roughly “a trillion good tokens” and transparent anchor sets than rely only on 36-trillion- or 100-trillion-token runs whose objective functions nobody locally controls.
17. FoundationCoin tries to turn beneficial compute into monetary backing
Mostaque’s new social contract begins with consequential AI as a utility “owned and controlled and optimized for the people.” Existing money originates largely as bank credit and debt, transfers rents from young borrowers toward older asset owners, and lets accumulated capital compound—just as AI reduces labor’s ability to attract capital.
His company is building fully open models for finance, education, health, and government. Verified deployments would help secure a Bitcoin-like FoundationCoin, while nationally distributed miners stack compute to provide universal AI, organize collective knowledge, and fund supercomputers for cancer, education, culture, and other public goods.
The demand-side bet uses existing pools: crypto is roughly a $4 trillion market; Mostaque says OpenAI’s annual inference budget is comparable to Bitcoin’s compute budget, AI companies generated around $20 billion, and approximately $160 billion flowed into crypto. A useful digital asset might redirect a fraction toward public-interest intelligence.
The practical promise is a free app next year that checks every diagnosis in every language using an open medical model. Mostaque asks why nobody has organized all cancer knowledge for every family: “Someone’s got to do that,” even when GDP and conventional company incentives do not reward it.
18. The monetary proposal is intentionally unfinished
Labenz presses the tokenomics: if buyers finance compute, can they later redeem the coin for compute, or is this still speculative? Mostaque’s answer is closer to auditable benefit than redemption—holders choose cancer, Alzheimer’s, education, or another destination, see the supercomputer, and measure how many people their capital helped.
FoundationCoin would function as the gold-like store of value; a second cash-like currency might be issued because someone is a conscious human, then support baseline survival and AI access. Mostaque plainly concedes, “The next part, which is the part we haven’t quite figured out yet,” is that human-linked issuance layer.
His illustrative ownership alternative shows why shares in frontier labs are insufficient: even if OpenAI reached $100 trillion and Americans collectively owned 10%, that is about $29,000 of equity each and, as Mostaque states, roughly $15,000 annually at a 5% dividend. It would still not finance universal subsistence, much less the whole world.
Asked whether nation-states remain the right unit, Mostaque offers “no position.” Countries are pragmatic because health, education, government, financial services, and data are local; he imagines dollar-valued national champions owned by citizens, while acknowledging that transnational AI labs and emerging network states will challenge geography-based sovereignty.
19. Policy should engineer the landscape, not push isolated levers
Universal AI could reverse government’s information bottleneck: citizens report richer local conditions while policies are checked before deployment. Mostaque wants every US bill tested reproducibly for constitutional consistency, common sense, benefit to Americans, and contribution to flourishing.
The goal is “geometry engineering versus policy engineering.” Instead of injecting cash and hoping, government changes the harmonic landscape so desirable activity flows naturally—raising intelligence, networks, openness, diversity, and resilience rather than maximizing GDP while degrading those foundations.
Better measurement would also create checks against corporate capture. The first trusted system that evaluates every policy or organizes the world’s cancer knowledge becomes a powerful Schelling point; Mostaque argues those control planes must be collectively owned before their authority hardens.
20. Better data and incentives are Mostaque’s route to safer AGI
His safety case has three layers: public-interest compute can become the highest marginal buyer and alter lab incentives; cultural and ethical “gold-standard anchor sets” improve training data; and a distributed hierarchy of narrow agents supplies computational resilience against a centralized attacker.
Data matters as much as architecture. A small amount of wrong or distorted training material can produce strange behavior, so Mostaque would require AI companies to disclose model ingredients and meet data standards. A medical model makes the question obvious: “Why do I need any Reddit data?”
Having signed the 2023 pause letter, he now says pauses and conventional regulation will not overcome the economic prize of capturing human intellectual labor. “If you want to change people’s behavior, you have to change the incentive landscape”; FoundationCoin is his proposed incentive, not something he claims is certain to succeed.
The desired endpoint is a sovereign AI each person owns: clean inputs, transparent consequential models, personal alignment, and an objective of individual, community, and societal flourishing. He rejects both government allocation of every decision and private-company ownership of the lifelong control plane.
21. The next thousand days determine who owns the control plane
Capability and adoption will move at different speeds. Mostaque corrects Dario Amodei’s forecast that 90% of code “will” be AI-written to “can” be written; perhaps only 50% of programmers currently use AI. Jobs therefore disappear unevenly, even as the technical substitution threshold arrives quickly.
Some non-performance-based public jobs may remain protected while private entry-level work falls first, but lost income and consumption eventually reduce tax receipts. Once users have deposited their lives, health records, and memories into one assistant, switching costs and organizational moats become formidable.
That is why he wants GPU capacity diverted toward benefit before scale locks it up, and collectively owned interfaces established while users still have choices. “Don’t give them ownership of the control plane. Don’t give them ownership of your data.”
His positive destination is “Star Trek,” not “Star Wars”: robots handle material necessities while people explore, learn, create art, strengthen families, and care for one another. The machines are taking jobs; the hoped-for conclusion is, “Thank God. Now we can get to our real work.”