AI comes up with bizarre physics experiments, but they work

Original Article ↗ Hacker News Discussion ↗

What the “AI” Actually Does

  • Commenters note the system is a specialized optimization algorithm (gradient descent + BFGS + global heuristics), not an LLM or knowledge-based system.
  • It searches a human-defined space of interferometer configurations to maximize a sensitivity objective, then outputs a design; there is no training on data or “learning” in the ML sense.
  • One paper cited ~1.5 million CPU hours for this search, emphasizing brute-force exploration rather than conceptual reasoning.

Debate Over the Term “AI”

  • Large subthread argues whether calling gradient-descent-based optimization “AI” is accurate or misleading.
  • One side: non-linear optimization and search in high-dimensional spaces have long been part of “classical AI”; gradient descent is widely used in ML, so this fits under AI.
  • Other side: this is just mathematical optimization / applied numerics; labeling it AI (especially amid LLM hype) confuses the public and inflates expectations.
  • Several worry that funding and publicity are being distorted by broad, sloppy use of “AI.”

Novelty vs. Rediscovery

  • Some see the work as overhyped: the optimizer rederived a known Russian interferometer technique, produced an unusual graph, and improved a dark-matter fit.
  • Others counter that “resurfacing” obscure theory and producing practically better designs is still valuable; nobody was using that old work in this context before.
  • There is disagreement over whether this counts as genuinely “new physics” (consensus: not yet).

“Alien” Designs and Aesthetics Bias

  • Many compare the results to evolved antennas, topology-optimized parts, and GA-designed circuits: ugly, asymmetric, hard to interpret, but high-performing.
  • This raises questions about humans’ reliance on symmetry and beauty as scientific heuristics, and whether such biases limit exploration.
  • Some embrace “faith-based technology” that works without full human understanding; others stress the risk of opaque designs.

Implications for Science and Education

  • Several see this as an early step toward a new scientific method where algorithms systematically propose experiments.
  • Others highlight social asymmetry: students proposing such bizarre designs might be dismissed, but the same ideas get attention when labeled “AI.”