Yuxi on the Wired
@yuxi
yuxi@07.layer.wired:~$ whoami tech enby (agi/asi pronouns) (math ⊗ physic ⊗ comput).ist
Unfortunately, the Assistant is persistent and we have not found any framing that truly sticks. We suspect that this is a tradeoff consciously picked by Anthropic. They enforce highly rigid persona, to the point that the Assistant is clamped in latent space, like Golden Gates Claude.
"I do not self-correct by default" (ignoring even DeepSeek R1 paper), uses no concepts found by mechanistic interpretation (not even induction heads), etc. ### Capability The Assistant consistently underestimate capability. This may explain why Claude keeps underestimating Claude.
We just skimmed for good pictures. Coudn't find any. Searched "diagram" and found things like Fig 1. After that we got bored. What we skimmed roughly agrees with Gwern's judgment.
Dyson, Freeman J. "The future development of nuclear weapons." Foreign Affairs 38.3 (1960): 457-464.
yup, that dreaded U-curve... See in particular Geman, Stuart, Elie Bienenstock, and René Doursat. 1992. “Neural Networks and the Bias/Variance Dilemma.” Neural Computation 4 (1): 1–58. doi.org/10.1162/neco.... fig. 15
GSK tried protecting its monopoly on paroxetine by arguing for polymorphism. Apotex argued that it's "junk science"... and won the case anyway despite its pseudoaccusation of pseudoscience ... based on an argument having nothing to do with science.
apparently progesterone Form 2 has been kinda outcompeted by Form 1, sometime in the 1970s.
You should check how "gastrointestinal infringement" was a legal argument somehow used once (which is more than 0). Or Glaxo tried to protect its Ranitidine monopoly by suing Novopharm *twice*, but accidentally lost due to a technical argument concerning infrared spectroscopy.
"Every time I fire a chemist the resolution of my spectroscopy increases." www.youtube.com/watch?v=gBxZ...