Links for January 2025
AI Education: Harvard ran a study comparing physics students who work with an AI tutor against a human-led, active learning classroom and saw extremely promising results
“Not only did the AI tutor seem to help students learn more material, the students also self-reported significantly more engagement and motivation to learn when working with AI.”
AI Research: Packy McCormick shared an interesting model for what research might look like in the age of AI. Marcel Grossmann played a material role in Albert Einstein’s discovery of General Relativity by providing him with the necessary mathematical framework to make the whole theory work. AI will not take over Einstein’s role anytime soon but very well might be able to do what Grossmann did.
”In other words, the cheaper Grossmann becomes, the more valuable Einstein becomes.”
Revisiting Einstein: Sabine Hossenfelder wrote an interesting post about Einstein’s Other Theory of Everything. One idea Einstein worked on after his big breakthroughs were attempts to geometrize everything including electromagnetism and elementary particles. Together with Nathan Rosen he wrote a paper speculating if wormholes might be interpreted as elementary particles. (Modern examples of “geometric” elementary particles are solitons, sphalerons, and instantons.) Sabine is wondering if this idea “isn’t worth revisiting”.
Another interesting idea Einstein worked on later in his career is tele-parallelism. Manifolds are characterized by torsion, curvature, and non-metricity. General relativity is a theory of manifolds with non-zero curvature but zero torsion and no non-metricity. A weird feature of general relativity is that regarding all observational predictions it seems to work equally well if we consider it as a theory non-zero torsion (and zero curvature and zero non-metricity) OR non-zero non-metricity (and zero curvature and zero torsion). This seems to raise serious questions about the standard interpretation of general relativity as a model of gravity in terms of spacetime curvature. Moreover, a much richer class of theories emerges once you relax the requirements and allow, for example, as Einstein did, non-zero curvature and non-zero torsion.
Con-Artist Entertainment: Recently started listening to the The Pirate of Prague podcast and really enjoyed the first few episodes. It’s a crazy story about a czech grifter pulling off one insane scam after another.
AI News: OpenAI released Operator an AI agent that can control a web browser. Definitely a cool idea but I very much doubt there will be serious use cases anytime soon. This is one of these instances where you can produce cool demos with something that 80% works but is virtually unusable for real life use cases unless you fix the remaining 20%. Most websites have strong guardrails against automated access. The web in its current form is far too diverse in terms of technologies that are used. Most websites are quite unstable and have tons of dark patterns that regular mislead humans.
Danish Utopia: David Heinemeier Hanson shared some thoughts on why things work better in Denmark. The gist of his theory is that Denmark takes the broken window theory far more serious and rigorously punishes small antisocial behavior to prevent bigger problems. (As a German living in Denmark I very much doubt this has a meaningful impact. All examples he describes, e.g. police ticketing bicyclists, happen exactly the same way in Germany and yet the country looks very different.)
On X, he shared a second theory: immigration policy.
“Sweden and Denmark ran the world's greatest A-B test on mass immigration and the results are in.”