Coding Is Where the AI Money Actually Is, and What Falls Next
Software is the first place AI became real economic value. Where it spreads next, and who actually pockets the money, is where the sharpest voices disagree.
Software is the first place AI turned into real, large economic value. Where that value spreads next, and who actually pockets it, is the part the field has not settled. What follows is a map built from the most useful recent framework, the podcaster Dwarkesh Patel's idea of "grindability," stress tested against the economists, investors, and builders who see the same evidence and draw different conclusions.
Why coding fell first, the one thing almost everyone agrees on
Patel's sharpest recent point is that being "verifiable," meaning you can automatically check an answer, does not by itself explain where AI improves fastest. A task also has to be grindable: you can run thousands of attempts in parallel against a cheap, replayable simulator and keep what works. Software is the rare domain that is both. Code either passes its tests or it does not, and you can hand a thousand agents "an identical copy of the container" of a repo and let them attack the same bug. Contrast the live world: "You can't have a thousand agents go try the same checkout flow on Amazon," Patel wrote, "because Andy Jassy will find and detect your bots and shut your ass down."
Investors are converging on the same insight from the money side. The venture firm Bessemer describes reinforcement learning as its own emerging layer: "RL grounds AI in experience: environment building, RL-as-a-service, and platform infra are becoming their own stack." The AI writer Rohan Paul, summarizing a survey of more than 500 papers on agentic RL, puts the limitation plainly: "common LLM training rewards a single answer once, then stops learning," which is exactly why single shot, checkable tasks like coding fell before messy multi step ones.
The coding job did not vanish, it moved to verification
The near term reality for software is not a headcount cliff but a change of task. As Rohan Paul framed a recent Futurism piece: "From creation to verification. Software engineers now face a harder job: managing code they did not write." The developer community is openly split on how far that goes; the influential engineer swyx amplified a deliberately divisive take asking whether AI engineers should even "read code anymore in 2026." MIT labor economist David Autor rejects the idea of a software jobs collapse outright, arguing that past computing waves killed specific tasks but raised the value of judgment, expertise, and trust. The prediction that squares these: coding demand shifts from writing toward reviewing, architecting, and verifying, and the leverage moves to whoever is best at steering and checking machine output.
What falls next
Run every task through the same filter, verifiable and grindable, plus the physicist Adam Brown's idea of a "branching fraction," meaning how much real world experiment you need to prune the tree of possible answers. Plotted, the intuition is simple: the lower right falls first, the upper left stays stuck.
That yields a rough order of dominoes after coding.
| Domain | Falls when | Why |
|---|---|---|
| Formal math | Now, alongside coding | Proof checkers make it verifiable and proof search parallelizes. The recent AI disproof of an Erdős conjecture is the tell. |
| Symbolic desk work (SQL, data analysis, spreadsheet and financial modeling) | Near term | Checkable against expected output and cheap to replay once someone builds the environment. Patel cites teaching models to use Excel as a live example. |
| Optimization, logistics, formal verification, parts of chip design | Near to mid term | Deterministic simulators give verifiable metrics and are replayable. |
| Low branching theoretical science (theoretical physics, theoretical CS) | Mid term | Few internally consistent theories, so parallel reasoning can search the tree with little experiment. |
| Domains with imperfect simulators (protein and molecule design, robotics) | Slower | Good partial simulators exist, but final checks need a wet lab or a physical robot, which raises the branching fraction. |
| Live world agents, running a business, litigation, trading, elections | Stuck | Verifiable but not grindable; outcomes take months or years and cannot be re-run in parallel. |
| Experiment heavy science (biology, materials, condensed matter) | Stuck | High branching fraction; you have to build the experiment to know which theory is right. |
The order is not fixed by the tasks alone. The near term answer to "what falls next" is really "whatever gets a good synthetic environment built for it," which is why both Patel and Bessemer point at the RL environment industry as the leading indicator. Watching where that industry aims next is watching the next domain fall in real time.
Where the money actually lands, the contested part
Here the voices genuinely diverge. Patel's own framing is a paradox: if a model automates a job worth $200,000, its value is "both far more and far less" than that, because competition can drive the price of one more copy down toward "the cost of the GPUs and energy" even as the total value created balloons. In July he added the escape hatch: today's roughly three way race between comparable labs is what "competes their margins away," and if one lab opens a durable lead it "could probably get away with charging a lot more."
Almost everyone agrees the model layer is the squeezed middle. They disagree on who keeps the surplus. The equity research account Pequity Research estimates the application layer's share of AI revenue climbed while the model layer's share slipped over the past year.
That is evidence value is "increasingly pooling at the application end." A widely shared essay argues the opposite end wins: "AI's Biggest Winners Have the Lowest Margins," meaning manufacturers, trucking carriers, distributors, and field service operators, not the software rich firms. The bear case, from the account TheShortBear, is that nobody keeps it yet: models are being sold "at a loss," moats look thinner than they are priced, and this is happening "within the most cyclical businesses out there via semis and memory." Wharton's Ethan Mollick adds the labor version of the same force: "by leveling performance, AI also commoditizes contract labor."
The skeptics' discount
Set against the "everything falls" energy is a serious camp that expects far less, or far slower. In a June 28, 2026 Wall Street Journal piece, "Is an AI Jobs Apocalypse Coming? Three Economists Square Off," three economists split cleanly. MIT's David Autor is the conditional optimist: "Technology automates, it complements and it creates new expertise and new work. I don't think we're headed into a new world where human judgment, moral reasoning, empathy and know-how have no economic value." The University of Virginia's Anton Korinek is the alarmed one, warning that if the quest for artificial general intelligence succeeds, "labor itself becomes optional for the economy." MIT's Daron Acemoglu is the patient skeptic, arguing that neither economic theory nor the data support the exuberant forecasts and that change will unfold over decades, not quarters. The independent forecaster Gary Marcus lands near Acemoglu, holding that "less than 10% of the work force will be replaced by AI, probably less than 5%," with the net near term effect "not huge." Even a bull like Citadel's Ken Griffin frames it as timing, not doubt: white collar displacement is "real and coming fast," but the political backlash is a cycle away.
Underneath all of it sits the capability Patel thinks gates the economy: continual learning. Today's models are frozen after training and cannot get better at your specific job by doing it. He calls the popular "AI just takes time to diffuse" explanation "cope" for that missing ability, and notes coding is the conspicuous exception where value has already landed. If he is right, coding stays a lonely peak longer than the "agents do everything" narrative assumes, because most work is neither grindable nor low branching, and none of it is learnable on the job yet.
What to watch
Five signals. The RL environment startups, the real leading indicator of the next domain to fall. The size and stickiness of the quality gap between top models, the variable that turns AI from a thin margin commodity into a pricing power business. Where the value layer settles, applications, infrastructure, or the low margin operators who deploy it. The political timing window Griffin flags, which caps how fast deployment can run regardless of capability. And the first credible sign of learning on the job, the moment the value pool finally spreads beyond software.
Sources
- @dwarkesh_sp. Grindability versus verifiability, the Amazon checkout example. Jun 26, 2026.
- @dwarkesh_sp. The $200k job value paradox. Jun 26, 2026.
- @dwarkesh_sp. Model pricing power. Jul 8, 2026.
- @dwarkesh_sp. The Erdős unit distance disproof. Jul 2, 2026.
- Dwarkesh Podcast. Adam Brown on branching fractions. Jul 10, 2026.
- @BessemerVP. RL environments as an emerging stack. Jul 8, 2026.
- @rohanpaul_ai. Coding shifts from creation to verification. Jul 7, 2026.
- @pequityresearch. Value shifting to the application layer. Jun 26, 2026.
- Wall Street Journal, "Is an AI Jobs Apocalypse Coming? Three Economists Square Off," Jun 28, 2026. Confirmed via MIT News clip; fuller three view framing in the Carnegie Endowment writeup. Surfaced via @rohanpaul_ai.
- @GaryMarcus. The low displacement forecast. Jun 4, 2026.
- @PodcastAlphaX. Ken Griffin on displacement timing. Jun 22, 2026.
- @emollick. AI commoditizing contract labor. Jun 20, 2026.
More research at bargo.ai/research.