Dear Magicians,

Why I’m Not Worried About My Job

I signed up for Claude Science the week it shipped. I went in chasing Lorentz invariance, maybe a crack at the theory of everything that eluded Einstein. What I found was protein folding. Genomics. PubMed, NIH, cheminformatics. I clicked through every offering and, when I looked, nothing was physics. It was biology, biochemistry, pharmacology, top to bottom.

That is not an accident. It is a confession.

Here is the mechanism, and it’s one I feel in my own work. Biology is a supervised-learning problem: millions of genomes, billions of protein sequences, decades of labeled examples. A neural network eats that. Physics is not. There is no training corpus of a million quantum-gravity examples. Nature supplied exactly one universe, and experimental cosmology proceeds one observation at a time, each one costing a multinational collaboration a decade and a billion dollars. A model cannot interpolate between examples that do not exist. So it is fluent in molecular biology and merely ordinary at the vacuum.

Ask Sam Altman or Dario Amodei what a scientist needs and the honest answer is a genomics stack and a protein-folder. To them, “science” means biology. AlphaFold earned Hassabis a Nobel, and it deserved it, but notice what it was: enormous intelligence applied to a problem already well posed. That is brilliant, but is it truly universal, like the laws of physics purport to be?

This is also why the labs are quietly most afraid of bioweapons. The dual-use is not a side effect, it’s arithmetic: the algorithm that finds a beneficial mutation can find a harmful one. And the barrier is gone. Want proof? Iran is still trying to produce a fission device 80 years after the USA deployed one in war. A pathogen has no such floor. Annie Jacobsen’s sources told her a kid in a basement, with an LLM, now approaches what once took a state. When she got a model to surface something it shouldn’t, it was fixed fast, because she had the CEO’s ear. Most people don’t. Safety becomes a race against diffusion, and diffusion usually wins.

Progress amplifies human intentions. It does not improve them.

Which is exactly why I sleep fine about my own trade. These systems master what has already been done and merely needs horsepower. Give one my Brian Keating-Einstein test and it’s a brilliant undergraduate and a hopeless Einstein: it will fold a known protein, but it cannot tell you which unasked question about the vacuum is worth your life. Wisdom is still scarce. So is physics. I’ll keep my job.

Until next time, have a M.A.G.I.C. Week,

Brian