links 5/19/25: https://roamresearch.com/#/app/srcpublic/page/05-19-2025

• https://arxiv.org/html/2502.15631 diminishing performance returns to reasoning token number.
  • more difficult questions get allocated more reasoning tokens.
  • higher reasoning "level" doesn't
• https://arxiv.org/html/2503.04697  there is an an optimal amount of reasoning to do at inference-time and doing more degrades performance. ("overthinking"). you can design models to estimate this for themselves.
  • see also:
    • https://arxiv.org/html/2504.05185
    • https://arxiv.org/html/2504.21370
    • https://arxiv.org/abs/2412.18547
• more on "overthinking" degrading performance:
  • https://arxiv.org/html/2410.21333
  • https://arxiv.org/abs/2502.07266
• https://www.paulgraham.com/safe.html  this is an analytical argument by Paul Graham; you can be "nice" (not capture the most value, in zero-sum contexts) and it won't affect your revenue/profits much at all compared to your growth rate.
• https://en.m.wikipedia.org/wiki/Monte_Cristo_sandwich  -- it's dipped in powdered sugar (!)
• things I looked up while reading about the French Revolution
  • https://en.m.wikipedia.org/wiki/Anne_Robert_Jacques_Turgot physiocrat of my heart!
  • https://en.m.wikipedia.org/wiki/Charles_Alexandre_de_Calonne spendiest minister ever
  • https://en.m.wikipedia.org/wiki/Guillaume-Chr%C3%A9tien_de_Lamoignon_de_Malesherbes critic of royal absolutism
  • https://en.m.wikipedia.org/wiki/Montesquieu
    • https://en.m.wikipedia.org/wiki/The_Spirit_of_Law
  •  
  • https://en.m.wikipedia.org/wiki/Germaine_de_Sta%C3%ABl political theorist & salonniere
  •