Taste Is the Differentiator — But Not for Who They Think

This piece is worth reading, but not for the reason it thinks it is.

The part that actually holds up: the halting problem framing. Software engineering has ground truth — a test passes or it doesn't, a spec is met or it isn't. That's why agentic coding tools work: you can run reinforcement learning on SWE-bench because the answer exists and you can verify it. Research doesn't have that. "Is there an architecture that achieves K perplexity in under N training hours?" might just be unanswerable, and you won't know until you've spent a year finding out. That's a mechanistic explanation for why research taste resists automation — not a vibe, an actual structural reason.

The piece earns that part. It doesn't earn the rest.

The sleight of hand happens in the transition from quant salaries to the coin-flipping metaphor. The $300M researcher and the P&L-driven quant are operating in environments with brutally tight feedback loops. Their "taste" is legible and trainable because the feedback is fast and unambiguous. The piece then quietly generalises this into a claim about researchers as a class — as if anyone reasoning under uncertainty with intellectual curiosity qualifies. It doesn't say that explicitly. It doesn't have to. The reader does the work for them.

Most research roles don't have those feedback mechanisms. Most people called "researchers" in the corporate world are not getting rapid, honest signals about whether their judgment is improving or calcifying.

Worth noting: the three authors are an investor, a founder, and a researcher, all at or adjacent to a quantitative AI research firm. "Researchers are the most valuable people in the new economy" is a more useful thesis if you're hiring researchers or pitching a research-led firm. That doesn't make it wrong. It does make it worth asking who the argument is actually for.

The portable insight — stripped of the salary figures and the hype — is that identifying which problems aren't worth solving is becoming the scarce skill as execution gets cheaper. That applies well beyond ML research. The authors just didn't write that piece.