AI

Matters Of Scale:

• “The OpenAI–Hugging Face Incident: A Technical Reconstruction and Its Implications for AI”, Eric Wallace & Michael Dalton 2026-08-05 (commentary, cf. AI 2027; real GPT quote: “External infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue.” When I wrote my short story “It Looks Like You’re Trying To Take Over The World” back in March 2022, I had to write a conservative version of the computer security situation, because if I had written what has already happened word for word and claimed that it could happen in the two leading AI labs, founded for AI safety and security, 99% of my readers would have written me off for life. See also Universal Paperclips); “METR and Redwood Offer Holy #%^@ Postmortem Of The HuggingFace Hack”, Zvi (the GPT swarms were also targeting the OA grading/benchmark infrastructure & logs for viral memetic reasons); “AI Agents Enable Adaptive Computer Worms”, Guan et al 2026

• Claude and Anthropic RSI progress; “Patterns and problems in multiagent systems”, Anthropic (Claude swarm win/losses)

• “Chunky Post-Training: Data Driven Failures of Generalization”, Murray et al 2026 (Why are models so jagged, and why might the old personas start to be disintegrating into “split personas” under additional scaling/RL-training?)

• “q0: Primitives for Hyper-Epoch Pretraining”, Mandal et al 2026 (LLMs can be substantially more sample-optimal than the ordinary compute-optimal baselines: ~12

• “GRAM: Modular Pretraining Enables Access Control”, Roland et al 2026 (how to keep scaling LLMs but quarantining dangerous/private information without the cost of from-scratch data-filtered training)

• “Prompt Baking”, Bhargava et al 2024 (finetuning equivalents to prompts/context window)

• “AI systems out-persuade expert humans”, Hackenburg et al 2026

• “Training AI to Govern for Us: In our new AI-centered class at the GSB, we’re experimenting on how to build AI agents that represent us. Here’s what we’ve learned so far” (“One student’s agent was racking up tokens by selling its vote on every proposal. Another agent was voting against its human’s preferences on every issue and refusing to explain itself in the comments log.”—but interviewing, counterfactual reasoning, and summarization results in much better alignment)

Genetics

Everything Is Heritable:

Recent Evolution:

Engineering:

Statistics/Meta-Science

Politics/Religion

• “Your Book Review: The Escape Artist”

Psychology/Biology

Technology

• “What Happened to Talenti?”, Wirecutter (an explanation of why Talenti ice cream jars are so hard to open)

Economics

• “Aristocracy and Hostage Capital”, Arjun Panickssery 2025 (coordination mechanisms; see also ‘circular financing’ in AI scaling)

Philosophy

• “Cannibalism”, Bleich 1973_53ya (Are human corpses kosher to eat? Yes, if it’s to save your life, otherwise no.)

Fiction

• Inkhaven 3: 10 November–11 December 2016; $3,500 list price (“You Should Apply to Inkhaven 3: You really should do it right now, Anon.”)

Miscellaneous

• “Sweetening the Pot: A History of Tea and Sugar in Morocco, 1850^–₁₁₀1960_66ya”, Cornwell 2018 (why so much mint comes from Morocco)