Coding Is Where the AI Money Actually Is, and What Falls Next

AI Analysis

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.

Bargo Research

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.

What falls first vs. what stays stuck (illustrative)

That yields a rough order of dominoes after coding.

Domain<br>Falls when<br>Why

Formal math<br>Now, alongside coding<br>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)<br>Near term<br>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<br>Near to mid term<br>Deterministic simulators give verifiable metrics and are replayable.

Low branching theoretical science (theoretical physics, theoretical CS)<br>Mid term<br>Few internally consistent theories, so parallel reasoning can search the tree with little experiment.

Domains with imperfect simulators (protein and molecule design, robotics)<br>Slower<br>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<br>Stuck<br>Verifiable but not grindable; outcomes take months or years and cannot be re-run in parallel.

Experiment heavy science (biology, materials, condensed matter)<br>Stuck<br>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...