LessWrong (Curated & Popular)

A sharp critique of Anthropic’s “Hot Mess” research and its blog framing. The discussion highlights selective reading of results and a misleading definition of “incoherence.” It questions key experiments, statistical measures, and claims about future alignment risks. The narrative also flags possible LLM authorship of the blog and methodological overstretching.

A conditional pledge campaign to trigger a large protest in Washington, D.C. if 100,000 people sign up. The episode explains why a mass march could legitimize AI x-risk concerns and the reasoning behind using low-cost conditional kickstarters. It covers who should sign up now, targets for early momentum, possible local marches, and logistical FAQs about timing and treaty demands.

Advice on avoiding full teams and preferring smaller-footprint solutions or contractors. Guidance on hiring one highly capable generalist to coordinate work. Reasons to keep teams tiny because management and coordination costs grow quickly. Practical tips to be picky, use realistic interview tasks, set clear success metrics, and plan for team formation phases.

A narrated tour of an anonymous critique that reads AI culture through Dostoevsky. Short sketches link literary character types to modern lab personalities and failure modes. Topics include moral erosion in institutions, reasoning that leads to despotism, and a shared AI social organism that pressures dissent. The piece frames intelligence outrunning conscience as a cultural pathology.

A professor stages a month-long 1492 papal election simulation to teach Machiavelli by immersion. Sixty students play historical figures with detailed character packets and chroniclers. Single actions like assassinations ripple into wildly different continental outcomes. The simulation is compared to ensemble forecasting, wargaming, and experimental history to reveal patterns amid chaos.

An AI system found 12 new OpenSSL zero-days, including high- and moderate-severity flaws. The team explains automated vulnerability hunting, historical low-severity CVEs, and accepted AI-suggested patches. They also describe curl canceling its bug bounty after a flood of AI-generated spam and discuss what AI-driven security means for the future.

A deep dive into an essay on AI's perilous technological adolescence. Topics include autonomy risks and defenses, misuse for destruction and biothreat countermeasures, and AI-enabled seizure of power with policy responses. They also cover economic disruption, labor market shocks, wealth concentration, and indirect existential uncertainties facing humanity.

Tiny RNNs and transformers trained on algorithmic tasks are used as surprising testbeds for mechanistic interpretability. The show walks through models from 8 to 1,408 parameters and why fully understanding few-hundred-parameter systems matters. Concrete case studies include second-argmax RNNs at several hidden sizes, analyses of decision geometry and neuron subcircuits, and a clear challenge to scale interpretability methods.

They dig into a Cold War stock sleuthing story that linked lithium trades to H-bomb secrets. They introduce the quirky Pentagon Pizza theory and trace its origins. They walk through data-collection hurdles, scraping tweets and building controls. They preregister tests around three military events and report backtest results showing no meaningful pizza signal.

A behind-the-scenes chat about building a custom command-line video editor and automating shot-cutting with Claude Code. They discuss creating transcript line IDs and a small DSL to reorder lines into ffmpeg commands. Practical constraints like file size, disk and network limits come up. They also reflect on what worked, what needed manual fixes, and lessons learned about tooling and agents.