In late July, a conversation at a private gathering made its way onto X — and set off a week-long argument between Silicon Valley and Wall Street.
Google engineer Patrick Toulme relayed what two engineers from a top New York quant fund had told him: they still hand-write all of their code, they find frontier-model output "poor quality," and they genuinely dislike agentic coding. Toulme added his own verdict: he was starting to believe there's a skill gap in using frontier models — one that not everyone has.
That's not what you'd expect to hear from Wall Street in 2026. Over the past year, the street has been adopting AI at a pace that borders on aggressive: Morgan Stanley's internal platform has rewritten roughly 16 million lines of legacy code as of February. Goldman Sachs began deploying Devin, an autonomous coding agent, in mid-2025 — starting with hundreds of instances and planning for thousands. Jane Street went further and trained its own code models.
So the interesting question isn't who's right. It's why the two sides keep talking past each other.
They're not answering the same question
Read the replies closely and you'll notice the debate has two different subjects.
Silicon Valley is answering: can AI write code? That answer improves every month.
The quant engineers are answering something else entirely: in an industry where a mistake is priced in real dollars, and where the core codebase has never appeared in any training set — what are the preconditions for handing code to an AI?
The most-endorsed reply in the thread, from a Meta ML engineer, laid out why quant shops are extreme outliers: their code is hyper-optimized, often written in niche languages (Jane Street runs its entire stack on OCaml), built on proprietary APIs, and radically closed. By Jane Street's own public account, its internal OCaml codebase may exceed all the OCaml publicly available on the internet — combined. LLMs excel at pattern generation over things they've seen many times. A quant firm's most valuable code is precisely the code no model has ever seen.
Then there's the cost asymmetry. One engineer in the thread described his workflow bluntly: he barely reads code anymore — he reviews ER diagrams and sequence charts, and "if there's a bug, so what? It's easy to fix." That's a perfectly rational stance when errors are cheap. In trading systems, a subtle, plausible-looking logic error doesn't crash a page; it loses money — and spotting it demands the most senior judgment in the building. Same AI, two error-cost regimes, two rational conclusions.
The real bottleneck isn't generation. It's verification.
Generation is getting cheaper by the month. Verification isn't. That widening gap is the actual engine of this argument — and it's not unique to coding.
Former OpenAI scientist Jason Wei calls it Verifier's Law: the ease of training AI to solve a task is proportional to how verifiable the task is. Everything that can be checked quickly, objectively, and at scale eventually gets solved. Sudoku falls fast. Fact-checking an argument doesn't. Ten years of deep learning is, roughly, the history of machines conquering whatever could be graded instantly.
The overlooked corollary: verifiability is not a fixed property of a task. It can be engineered. Reviewing code line-by-line is slow; a comprehensive test suite grades any submission instantly. Same exam — someone just made it machine-gradable.
Which reframes the opening anecdote. Jane Street didn't stop at "models haven't seen OCaml." It built training data from workspace snapshots capturing real bug-fixing sessions, and required every RL-generated patch to pass parsing, type checks, and existing tests. Goldman redefined the engineer's job as "describe the problem, supervise the agent." Morgan Stanley decomposed decades of legacy systems into structured blueprints so every rewrite could be checked item by item.
"Wall Street rejects AI coding" is an incomplete picture. On the same street, teams without verification infrastructure are rationally refusing — while teams that built it are adopting at scale.
Why quant might get there first
Here's the twist: quantitative investing may be one of the few fields where verifiability is easiest to engineer — because it comes with a built-in referee.
In most software, quality is fuzzy: UX, maintainability, architecture, taste. Quant research has run a brutal evaluation loop for decades: hypothesis, implementation, backtest, out-of-sample validation, live market feedback. The market doesn't care whether your code is elegant. It cares whether your prediction holds. Against Verifier's Law's criteria — objective, fast, scalable, low-noise, rankable — quant research checks nearly every box.
So the question that matters is shifting. Not "can AI write a factor?" but "can AI keep proposing hypotheses — inside a system that reliably kills the wrong ones?" That's where AI-native investment research is heading: not more generation, but a closed loop of generate → validate → eliminate → remember → regenerate, where failed paths become memory and every new search starts from higher ground. Humans don't disappear in that loop; they move — from signing off on every line to designing the rules of verification and judging the gray areas the rules can't reach.
One boundary worth stating: most industries don't have a referee as merciless as the market. Where verification still depends on human eyes, those two New York engineers are still right to be careful.
But the direction is clear. Generation is becoming a commodity — every firm calling the same models at the same price. And this industry has a ready-made verdict for capabilities anyone can buy: Generation is becoming beta. Verification is the new alpha.
Which is also why this debate will never produce a public winner. The first rule of alpha is that you don't share it. Next time this argument flares up on X, the real signal won't be who argues loudest — it'll be which firms have quietly left the thread.
Sources
Patrick Toulme's thread on X (July 2026); Jason Wei, "Asymmetry of Verification and Verifier's Law" (2025); John Crepezzi, "AI Engineering at Jane Street" (AI Engineer conference); CNBC interview with Goldman Sachs CIO Marco Argenti (July 2025); American Banker on Morgan Stanley's DevGen.AI (June 2026).