Running list of papers (and a few blog posts) I’ve read, newest first. Not sorted by year — just appended.

  • Use of machine learning to enhance detection of transient astronomical phenomena in historical observatory images (Bruehl, Doherty, Streblyanska, Villarroel; Sci Rep 16:31890, 2026) · thread
    • pre-Sputnik sky plates have star-like points that appear then vanish in <30 min. Claim: they thin out in Earth’s shadow (“shadow deficit”) and spike within ±1 day of nuclear tests (strongest ~24h before a test).
    • an ML classifier (AUC 0.81) trained on 250 eye-reviewed real-vs-plate-defect pairs, run over 107,875 catalogued transients; the correlations get stronger among the highest-P(real) ones (nuclear window p<0.0001, shadow deficit effect p=0.003).
    • pretty extraordinary claim. could the classifier be detecting something emitted from the nuclear tests? but very hard to falsify. fun, but not certain imo.
  • Cryptanalytic Extraction of Neural Network Models (Carlini, Jagielski, Mironov, 2020)
    • recovering a trained neural net’s exact weights from query access, framed as cryptanalysis — differential-style attacks to peel off layers.
    • crypto × ML; the model-stealing threat model.
  • First-in-human quantum entanglement imaging
    • PET scan observing entanglement on a 68Ga - observed annihilation photons
  • EML: All elementary functions from a single binary operator
    • eml(x, y) = e^x - ln(y)
    • gives you all the cos, sin, tan
    • cute. practical? who knows!
    • i sketched a bit what making circuits with this would look like.
  • Intel Skylake hyperthreading bug
    • Skylake series of Intel processors had a hardware bug when using certain registers in tight loops, with hyperthreading enabled.
    • Manifested as crashes in OCaml programs.