links 5/13/2025: https://roamresearch.com/#/app/srcpublic/page/05-13-2025

• https://www.cnn.com/2025/05/12/asia/india-pakistan-kashmir-ceasefire-intl-hnk  India/Pakistan ceasefire
• drug price controls executive order
  • https://www.cato.org/blog/trump-attempts-price-controls-prescription-drugs
  • https://www.whitehouse.gov/presidential-actions/2025/05/delivering-most-favored-nation-prescription-drug-pricing-to-american-patients/
• https://blog.sentinel-team.org/p/global-risks-weekly-roundup-202025
  • US/China trade deal reduces tariffs
  • cuts in US military officers, federal research funding
  • Stephen Miller, homeland security advisor and White House deputy chief of staff, says they're "actively looking at" suspending habeas corpus for immigrants
• https://languagehat.com/why-adjectives/ not all languages have adjectives! read the comment thread too, it's fascinating.
• https://classics.mit.edu/Aristotle/interpretation.2.2.html  "On Interpretation"
• https://dynomight.net/titles/ how to title blog posts
• Claude Sonnet 3.7 benchmark & compute references
  • https://www.anthropic.com/news/claude-3-7-sonnet
  • https://wandb.ai/byyoung3/Generative-AI/reports/Evaluating-Claude-3-7-Sonnet-Performance-reasoning-and-cost-optimization--VmlldzoxMTYzNDEzNQ
  • https://www.interconnects.ai/p/claude-3-7-thonks
  • https://www.vals.ai/benchmarks/mmlu_pro-03-13-2025
  • https://smythos.com/news/claude-3-7-sonnet-an-in-depth-analysis/
• Grok 3 benchmark & compute references
  • https://x.ai/news/grok-3
  • https://www.helicone.ai/blog/grok-3-benchmark-comparison
• GPT4 benchmark & compute references
  • https://www.linkedin.com/pulse/supposed-leak-gpt4-architecture-alvaro-duran-tovar/
• Llama 2 benchmark & compute references
  • https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/
  • https://www.cursor.com/en/blog/llama-inference
• GPT-3.5-turbo benchmark & compute references
  • https://www.vellum.ai/blog/gpt-4o-mini-v-s-claude-3-haiku-v-s-gpt-3-5-turbo-a-comparison
• calculating inference FLOPs and other compute measurements from model architecture, in transformers
  • https://kipp.ly/transformer-inference-arithmetic/ a now-classic reference post, explains the kv cache
  • https://www.adamcasson.com/transformer-flops.pdf
  • https://jax-ml.github.io/scaling-book/transformers/
  • https://www.artfintel.com/p/where-do-llms-spend-their-flops
    • inference time (and hence FLOPs on a given machine)scales linearly with output length if you use a KV cache
  • https://www.baseten.co/blog/llm-transformer-inference-guide/
  • https://www.lesswrong.com/posts/g7H2sSGHAeYxCHzrz/how-much-ai-inference-can-we-do

• https://marginalrevolution.com/marginalrevolution/2025/05/who-wants-impartial-news.html  who prefers impartial news over news that shares their point of view? apparently old, rich, male, non-ideological people. (I did not read the paper & thus do not have an opinion on its credibility.)