The Question Is the Hard Part - Jens Ernstberger
Jens Ernstberger
SubscribeSign in
The Question Is the Hard Part<br>How to think about research when finding answers is cheap.
Jens Ernstberger<br>Jul 26, 2026
Share
I always pictured human knowledge as a sphere<br>Vast, but tightly bounded at its edge. Research is the act of making a tiny bump against that boundary, pushing it outward in a direction worth pursuing for humanity’s sake. There’s a kind of sanctity in achieving that bump. Not because it’s useful, but because it’s scarce. You can’t buy a proof with money; you can only earn it. There’s something selfless in that, and something timeless too. A theorem discovered a thousand years ago still lives with us today.<br>Thanks for reading Isar Valley Builder Notes! Subscribe for free to receive new posts and support my work.
Subscribe
Right now, a lot of researchers seem to be quietly running the same experiment: take a hard open question you’ve been stuck on for years, feed it into a frontier model, and see what happens. For a surprisingly large fraction of these questions, the model simply solves it.<br>I don’t think most people have absorbed what this means. The two oldest arguments for why humans still matter in research, that machines don’t remember, and that machines can’t check their own work, are no longer reliably true. Memory can be built. Verification can be delegated to another instance of the same mind.<br>There’s a concept in physics called the threshold theorem, proven in the late 1990s for fault-tolerant quantum computation. A single physical qubit fails at some rate. Encode it into a larger code, and you get a logical qubit that fails less often, but only if the physical error rate starts below a fixed threshold. Below that threshold, each additional layer of encoding squares the error rate down. A qubit failing at some rate, encoded once, fails much less. Encoded again, it falls further, doubly exponentially with each added layer. Above the threshold, the same process runs in reverse. in summary, if the base failure rate is too high, that overhead costs more than the correction saves.<br>It is the same logic now showing up in how models check each other. A generator model gets a hard problem right most of the time, wrong some percentage of the time, call that its error rate. A second model checks the first one’s work, looking for mistakes, sending them back to be redone. Loop this a few times. Whether this loop helps or hurts depends entirely on which side of the threshold you started on, exactly as it does for qubits. Below the threshold, each round of checking compounds the improvement. Above it, the checker is no longer catching rare mistakes, it is catching noise, and its own errors start dominating instead of helping.<br>What feels new, what should feel unsettling, is that a lot of today’s models seem to have crossed that threshold sometime this year. Verification loops that used to make outputs worse are now making them reliably better, round after round. Maybe as a result, the past few months have brought a wave of stories about scientists, especially those in fundamental research, quietly questioning their sense of purpose.<br>Surprisingly, I believe this will lead to an era of research that is more focused than ever on the questions that actually matter. Richard Hamming gave a talk in 1986 that is not only a good orientation for research, but for life in general.<br>“You and Your Research”<br>The fundamental question he poses is why so few scientists make significant contributions and so many are forgotten in the long run. And similarly, in this era of machine dependent thinking, we will find ourselves reaching for the same lessons.<br>Hamming watched people with more mathematical ability than most of his colleagues produce nothing lasting, and he watched people others dismissed as inarticulate, Bill Pfann, Clogston, become the ones who collected the prizes. What separated them was courage to work on problems that mattered rather than problems that were merely tractable, and the discipline to generalize a result so that others could build on it rather than merely solve the isolated case in front of them.<br>If a frontier model can now close the isolated case in an afternoon, the isolated case was never where the courage lived anyway. Hamming’s advice was already, decades ago, a warning against the wrong kind of research: chasing a solvable problem because it is solvable, rather than an important one because it is important. Machines are extremely good at the former. They have no opinion about the latter. Deciding which questions are worth a model’s afternoon, or a decade of a life, remains a human act of judgment, and it is the only part of Hamming’s talk that a machine cannot yet stand in for.<br>I don’t know yet whether this leads to a future where fewer questions are pose, or more of them. I suspect it depends entirely on what we’re willing to admit we don’t know out loud, and to whom.<br>But I keep coming back...