< Home

Favorites

  • Strategy as a Wicked Problem (Camillus, 2008)
  • Why You Can’t Just Do Things - Octopusyarn
  • Strategic Taste
  • How to win a best paper award (or, an opinionated take on how to do important research)
  • How I've run major projects | benkuhn.net

Everything I've read

  • What We Learned from Briefing 140+ Lawmakers on the Threat from AI — LessWrong
  • The Best PR Advice You’ve Never Heard - from Facebook’s Head of Tech Communications
  • MIRI 2024 Communications Strategy - Machine Intelligence Research Institute
  • We Must Remember That Our World Contains Hell — LessWrong
  • RL creates split personas — LessWrong
  • Q2.5 2026 Timelines Update: Uplift and Revenue
  • A Spillway for Agent Coordination — LessWrong
  • Arguments for P — LessWrong
  • Why do models task game? — LessWrong
  • User awareness in frontier models | Transluce AI
  • This A.I. Just Created Viruses Not Found in Nature - The New York Times
  • How to pace the US frontier
  • The next chapter of our AI momentum
  • Returning to ARC — LessWrong
  • [r]AI energy use: its impact on prices, climate, and more | Epoch AI
  • Beware the Permanent Periphery—Asterisk
  • Do your capabilities homework — LessWrong
  • AGI Ruin: A List of Lethalities - LessWrong
  • A Safe Path to Open Weights - Thinking Machines Lab
  • Investigating three real-world incidents in our cybersecurity evaluations \ Anthropic
  • Design tips | Malaika Aiyar
  • How independent researchers could investigate AI propensities after misalignment incidents - METR
  • Untrusted advice for AI control: Short, strong advice significantly uplifts weak LLMs
  • Our position on open-weights models \ Anthropic
  • Introducing MAI-Cyber-1-Flash inside MDASH | Microsoft AI
  • What makes a good monitoring prompt? – Apollo Research
  • The OpenAI/Hugging Face Incident: Challenges in Controlling and Containing Cyber-Capable AI Systems — Institute for AI Policy and Strategy
  • Metrics of Agent Ability - METR
  • The Long (Self-)Correction — LessWrong
  • First, Make Me Care, by Gwern · Gwern.net
  • Séb Krier on X: "The way people discuss AI incidents is pretty important and I'm a bit concerned we're sleepwalking into a bad world littered with bad abstractions. How you label something is often a lossy compressions of a causal model, so the words you use affects which hypotheses people update https://t.co/IsMLwiMhR4" / X
  • Discovering Concept-Editing Algorithms With LLM Agents
  • Does distilling Claude carry the persona with it? — LessWrong
  • Rule Thinkers In, Not Out — LessWrong
  • Introducing Claude Opus 5 \ Anthropic
  • Are we existentially threatened by the type of AI misalignment seen in the OpenAI Hugging Face attack? — LessWrong
  • Announcing AIXI Labs — LessWrong
  • How our new Control Red Team is stress-testing frontier monitors | AISI Work
  • Your AIs don't do what you want. This is really bad
  • OpenAI and Hugging Face partner to address security incident during model evaluation | OpenAI
  • Security incident disclosure — July 2026
  • EXCLUSIVE: US, China to hold AI talks in September, sources say | Reuters
  • Tencent Hy
  • Differential acceleration of alignment-relevant capabilities is a bad bet — LessWrong
  • Coding vs thinking — Paradigm 3
  • Paradigm 3
  • What we’d like to fund — Paradigm 3
  • Safety and alignment in an era of long-horizon models | OpenAI
  • Don't default to nonprofit - by Carol and Austin Chen
  • which_claude_is_k3/writeups/write_up.md at main · rgreenblatt/which_claude_is_k3
  • China's Technological Playbook - by Sophie Kim
  • Judd Rosenblatt on X: "Xi Jinping AI speech transcript: Distinguished colleagues and guests, ladies and gentlemen, friends, 70 years ago, a group of young scholars proposed the concept of artificial intelligence for the first time at the Dartmouth workshop in New Hampshire of the United States. In" / X
  • [2607.14345] Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values
  • We need 3rd party Training-Run Assessments — LessWrong
  • When AI builds itself \ Anthropic
  • RL Post-Training on Macs | Pluralis Research
  • Inkling: Our Open-Weights Model - Thinking Machines Lab
  • Can risk aversion learned at low stakes generalize to astronomically high stakes?
  • How I think about catastrophic biological risk (part II): risk breakdown by type of prevention
  • How I think about catastrophic biological risk (part I): risk breakdown by type of response
  • The Four Pillars: A Hypothesis for Countering Catastrophic Biological Risk
  • The Dangers of Mirrored Life
  • ‘Give Away Your Legos’ and Other Commandments for Scaling Startups | First Round Review
  • Toward A Public Science of Model Behavior | Transluce AI
  • Policy Memo : Broad Institute of MIT and Harvard
  • Notes on Inference Integrity - by James Tillman - ForeWord
  • The current bottleneck is political will, not research — LessWrong
  • The missing half of AI futurism debates
  • What we learned from 1,604 Chinese AI job postings
  • Frontier labs don’t use most AI compute (yet) - by Josh You
  • Selective Optimism: a critique of AI 2040 — LessWrong
  • AI-Enabled Coups: How a Small Group Could Use AI to Seize Power | Forethought
  • Séb Krier on X: "Concepts of a Plan" / X
  • Total research transparency would be nice - by Ajeya Cotra
  • To be legible, evidence of misalignment probably has to be behavioral
  • Rest in motion
  • Do Things that Don't Scale
  • Humans are not automatically strategic — LessWrong
  • A "Failure to Evaluate Return-on-Time" Fallacy — LessWrong
  • I think alignment work is more promising than control work — LessWrong
  • Be impatient | benkuhn.net
  • Shut up and do the impossible! — LessWrong
  • Relentlessly Resourceful
  • What We Look for in Founders
  • The Case for Model Forensics — LessWrong
  • Alignment pretraining could backfire — LessWrong
  • The Invisible Side of AI Governance — LessWrong
  • AI catastrophe: more like a genocide than a thought experiment — LessWrong
  • Why are adversaries assumed to be incapable of responding to AI risk? — LessWrong
  • Optimisation over non-stationary distributions creates weirder minds — LessWrong
  • A basic systems architecture for AI agents that do autonomous research — LessWrong
  • Efficient tradeoffs and the safety-usefulness tradeoff model — LessWrong
  • How to be More Agentic - by Cate Hall - Useful Fictions
  • Estimating No-CoT Task-Completion Time Horizons of Frontier AI Models
  • Why I don't persuade people to do AI safety
  • Bad Problems Don't Stop Being Bad Because Somebody's Wrong About Fault Analysis — LessWrong
  • how to write an essay | scribbles in the margins
  • [r]secret-loyalties-whitepaper.pdf
  • The behavioral selection model for predicting AI motivations — LessWrong
  • Your Left Brain Doesn't Trade With Your Right — LessWrong
  • escaping flatland: career advice for CS undergrads
  • It's nice of you to worry about me, but I really do have a life — LessWrong
  • How to Actually Spend Billions on AI Safety - by Sophie Kim
  • A playbook for field strategy - by Dewi Erwan
  • Natural Language Autoencoders \ Anthropic
  • Risk from fitness-seeking AIs: mechanisms and mitigations
  • Not a Paper: "Frontier Lab CEOs are Capable of In-Context Scheming" — LessWrong
  • Not All Compute is Created Equal
  • Fail safe(r) at alignment by channeling reward-hacking into a "spillway" motivation
  • Half A Month Of Consolation Writing Advice
  • The Orange - The Gladdest Thing
  • An extremely non-comprehensive list of how to increase your surface area for luck and magic (and instantly sprinkle fairy dust on your life)
  • From personas to intentions: towards a science of motivations for AI models — LessWrong
  • Annoyingly Principled People, and what befalls them — LessWrong
  • Hidden Role Games as a Trusted Model Eval - James Lucassen's Blog
  • Model organisms researchers should check whether high LRs defeat their model organisms — LessWrong
  • Steering Might Stop Working Soon — LessWrong
  • Do Thing, Do One Thing
  • What 3,654 Job Postings Tell Us About Talent Needs in AI Safety — EA Forum
  • If Mythos actually made Anthropic employees 4x more productive, I would radically shorten my timelines
  • Why You Can’t Just Do Things - Octopusyarn
  • Anthropic repeatedly accidentally trained against the CoT, demonstrating inadequate processes
  • AI Populism's Warning Shots - by Jasmine Sun
  • How to walk through walls - by Henrik Karlsson
  • [2603.20639] Agentic AI and the next intelligence explosion
  • A woefully incomplete guide to technical upskilling
  • An Apple-Picking Model of AI R&D | Tom Cunningham – Tom Cunningham
  • 11 pieces of advice for children — LessWrong
  • The state of AI safety in four fake graphs — LessWrong
  • Product Alignment is not Superintelligence Alignment (and we need the latter to survive) — LessWrong
  • Academic Proof-of-Work in the Age of LLMs — LessWrong
  • Fitness-Seekers: Generalizing the Reward-Seeking Threat Model
  • AI Safety Talent Needs in 2026: Insights for Field-Building Organizations
  • Facing the Precipice of History - Chanden Climaco
  • AI Safety Needs Startups - by Joshua Landes and LTM
  • Your Work Will Change You Whether You Like It Or Not
  • Strategic Taste
  • Are AIs more likely to pursue on-episode or beyond-episode reward?
  • There should be ‘general managers’ for more of the world’s important problems
  • We don't need more founders in AI safety - by Gauraventh
  • Lessons from a year of university AI safety field building — LessWrong
  • Separating Prediction from Goal-Seeking — LessWrong
  • Two Skillsets You Need to Launch an Impactful AI Safety Project — LessWrong
  • How to Design Environments for Understanding Model Motives — LessWrong
  • Martian Interpretability Challenge: The Core Problems In Interpretability — LessWrong
  • Prefill awareness: can LLMs tell when “their” message history has been tampered with? — LessWrong
  • Don't Let LLMs Write For You — LessWrong
  • Tell Culture — LessWrong
  • How to win a best paper award (or, an opinionated take on how to do important research)
  • Current activation oracles are hard to use — LessWrong
  • The current SOTA model was released without safety evals — LessWrong
  • Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers — LessWrong
  • Good Ideas Aren't Enough in AI Policy - Andrew Wei
  • Instant LLM Updates with Doc-to-LoRA and Text-to-LoRA
  • The Stakes of Our Work - by Celeste Li - heart of hearts
  • Why AI won’t go well unless sensible people like you speak up and act.
  • Persona Parasitology — LessWrong
  • Value systematization: how values become coherent (and misaligned) — AI Alignment Forum
  • The Persona Selection Model: Why AI Assistants might Behave like Humans
  • The 2026 Global Intelligence Crisis - Citadel Securities
  • Questionable practices in machine learning
  • Managed vs Unmanaged Agency — LessWrong
  • Mapping LLM attractor states — LessWrong
  • METR's 14h 50% Horizon Impacts The Economy More Than ASI Timelines — LessWrong
  • Changing the world for the worse — LessWrong
  • You don't create a culture – Signal v. Noise
  • Minimal-trust investigations
  • Alignment to Evil — LessWrong
  • If you don't feel deeply confused about AGI risk, something's wrong — LessWrong
  • Aligning to Virtues — LessWrong
  • Did Claude 3 Opus align itself via gradient hacking? — LessWrong
  • My six stages of learning to be a socially normal person
  • Good conversations have lots of doorknobs
  • 21 Facts About Throwing Good Parties
  • How to Manage Relationships Like a Psychopath
  • Most of Your Efforts are Wasted. Here’s the Framework to Fix It.
  • Will reward-seekers respond to distant incentives? — LessWrong
  • Two Buckets - by Gauraventh - Dhaniya
  • How I've run major projects | benkuhn.net
  • [r]Frontier Safety Framework Report - Gemini 3 Pro (November, 2025) v2
  • Status Is The Game Of The Losers' Bracket — LessWrong
  • Where is the Capital? An Overview — LessWrong
  • Omniscaling to MNIST — LessWrong
  • Legible vs. Illegible AI Safety Problems — LessWrong