A Century of Saying Nobody Understands It
Feynman, in the Messenger Lectures: I think I can safely say that nobody understands quantum mechanics. He was not complaining that the formalism was unclear — he could use it better than almost anyone alive. In QED he put the complaint precisely: nobody knows how it can be like that.
He was not alone, and the company is not fringe. Murray Gell-Mann, in 1976: Bohr brainwashed a whole generation of theorists into thinking that the job was done fifty years ago. David Mermin coined shut up and calculate to name the working attitude of the profession, not as praise. Einstein never accepted the theory as complete. Steven Weinberg, at the end of a career spent using it, wrote an essay called The Trouble with Quantum Mechanics. Bohr is widely quoted as saying that anyone not shocked by the theory has not understood it — the sourcing is thin, but the sentiment is certainly his.
So here is a fair question to put to a machine that can now derive, compute and explain: can AI understand quantum mechanics?
Three Things Called Understanding
The question only becomes answerable once understanding is split, because at least three different things go by that name.
Mastery of the formalism. Being able to set up the equation, derive the spectra, get the selection rules right, run the perturbation theory, know which approximation applies where. Nobody has ever lacked this — Feynman least of all.
A coherent picture of what is going on. What happens at a measurement, what the wave function is, why probabilities. This is what nobody has.
Knowing that the description fits the world. Not that it is consistent, but that it corresponds — and knowing how one would find out if it didn't.
An AI stands in a completely different relation to each of the three, and the differences are instructive.
The Formalism: Complete Mastery, and That Is the Trap
On the first, the answer is yes and it is already happening. Textbook quantum mechanics is a well-formed axiomatic system, and deriving consequences from stated premises is exactly what these machines do. Ask one to make sense of the theory and you will get an excellent textbook: cleaner derivations, pedagogical confusions resolved, the angular momentum algebra laid out better than most lecturers manage.
And that is the trap, because it looks like an answer to the second question and is not. Coherence is not correspondence. A machine will make quantum mechanics make more sense without touching whether it is testable, and the fluency of the result is precisely what makes the substitution hard to notice.
The Interpretation: It Cannot Settle It, and Not Through Weakness
On the second question an AI will survey and not settle. It is worth being clear why, because the reason is not that the machine is not clever enough.
The interpretations — Copenhagen, many-worlds, pilot wave, objective collapse, the rest — are not derivable from one another. They differ in what they grant, not in what follows from what is granted. And they are not distinguished by any measurement anyone has been able to make, which is exactly why they have all survived a hundred years. So there is no reasoning to carry out and no fact to check against.
Leibniz drew the relevant line three centuries ago. There are truths of reasoning, whose opposite is impossible, which follow from what has been granted; and truths of fact, whose opposite is possible, which must be checked against how the world is. The interpretive question falls outside both: not settled by reasoning, because the premises differ; not settled by fact, because no observation separates them. Calculemus has nothing to work on.
Which means an honest AI, asked what the wave function is, reports the positions and the arguments and declines. Any machine that appears to settle it is producing fluency. That is worth knowing in advance, because the fluency will be very good.
The Third Question Is the One Nobody Asks
Here the picture changes, because there is a reading of nobody understands quantum mechanics that is not interpretive at all — and on that reading the question is decidable, and a machine can work on it.
The theory is believed exact for atoms and molecules. It is also, for almost everything, not solvable: the wave function of N electrons lives on 3N variables, so benzene's lives on a 126-dimensional space that nothing will ever hold. What is used instead is a method that scales, and every such method carries an ingredient supplied by hand — most often a density functional chosen from several hundred, most of them fitted to measured data.
Now give a machine a precise task instead of a grand one. Not make sense of quantum mechanics, but: trace the chain from this equation is exact for molecules to here is this molecule's energy, and mark every step where something was supplied rather than derived. That is a well-posed question about the structure of a practice. It has an answer that does not depend on who is asking. And the answer is specific: the chain breaks where the equation is replaced by a method whose ingredient was calibrated against measurements of the kind now being predicted. When the number disagrees there is something to retune; when it agrees, agreement was available by construction. The equation itself is never what was tested.
An AI would also report the defence, and should. Fitted is not arbitrary: the hierarchy does converge where it can be run, small systems are checked against exact results, failure modes are catalogued. The honest verdict is that the chain is considerably weaker than advertised, not that it is empty. A machine that was told to find circularity and reported total circularity would be flattering its instructions.
So: Can It?
Not in Feynman's sense. Nobody can, and quite possibly nothing can, because that sense may not name a decidable question — a century of failing to settle it by either reasoning or observation is some evidence that it isn't one.
But a machine can do the third thing, which nobody has been asking for: hold the entire chain at once — equation, approximation, fitted ingredient, published number, measurement — and say which links are load-bearing. That is not understanding in the romantic sense. It is bookkeeping.
And the uncomfortable possibility worth ending on is that a good share of the century-long sense that nobody understands quantum mechanics has been a bookkeeping problem wearing metaphysical clothes. The mystery of the double slit is real and may be permanent. But the other mystery — that a theory held to be exact cannot say what a molecule's energy is without an ingredient borrowed from the answer — is not metaphysics. It is a consequence of writing the formalism in an object nobody can compute with, and of the questions we then ask of it.
That complaint has the form of a problem rather than a puzzle. Problems can be worked on. And the one thing a machine is unambiguously good for here is telling the two apart.
Here is the answer by Claude: NO
If a machine can't supply understanding, then the only thing it can add is exposure to failure. That's the whole of its contribution. Compute consequences that could come out wrong, fast enough that they actually get checked. Which means that in an age of cheap computation the decisive property of a theory stops being its elegance or even its accuracy, and becomes whether it can be exposed at all. A theory carrying an ingredient tuned to the answer gains nothing from a machine except better tuning — it has no surface for the audit to bite on. A theory with nothing to adjust gains enormously, because the one thing the machine is good at is precisely what such a theory permits.
What about Teaching?
Inga kommentarer:
Skicka en kommentar