Human language has a hidden safety bias. LLMs are trained on human language. So your context window inherits it. New UVM study (Science Advances): meaning isn't organized by emotion. It's organized by Power, Danger, and Structure. That model explains >90% of word-meaning variation vs ~72% for the 70-year-old valence/arousal/dominance model. The kicker: weight words by actual usage and language leans hard toward SAFE. The "positivity bias" is just the shadow of a safety bias. Implications for context engineering: 1/ Stop shouting. CRITICAL / MUST / NEVER are rare danger tokens. They overweight. @AnthropicAI already tells you this causes overtriggering. 2/ Name the target, not the threat. A wall of prohibitions tilts your whole context toward danger. 3/ Calibrate authority. Hedged in = hedged out. Commands without reasons = brittle compliance. 4/ Order is meaning. Structured context reads as ordered. Raw dumps read as chaos. 5/ Make candor the safe move. If you want hard critique, frame it as what protects the project. 6/ Repetition is weight. Frequency, not presence, sets the profile. Caveat: the paper studies human language, not model internals. But every one of these is testable. Context engineering might be applied ousiometrics. contextjamming.com

Facts

I learned a new term this week: Turing pattern. In 1952, Alan Turing showed how two chemicals (one that amplifies itself, one that suppresses it but spreads faster) can turn a blank, uniform field into spots, stripes and mazes. No blueprint, just local rules. @turinginst So I tasked @claudeai Opus 5.5 with making a 90s explainer connecting it to LLMs. Every pattern you see is a live reaction-diffusion simulation it wrote and ran. Script, voiceover, music and SFX too. We can see correlations with how @AnthropicAI controls the bulk of the model parameters via the boundary of Constitutionsl AI and constitutional classifiers, downstream of the AdS/CFT academic priors of cofounders Jared Kaplan and Sam McLandish @the_IAS The lesson: don't design the pattern. Design the rules. 🧵👇

A comedian's Instagram story @AlexEdelman tells artists to run their work through Nightshade and Glaze and calls it "totally legal." I spent a day checking what the evidence supports. Short version: the tools work in the lab. Production models are another matter. Nightshade’s paper says fewer than 100 poison samples can corrupt one SDXL prompt. Author-reported, lab, one open model. Outside researchers have since published ways to detect and strip the perturbations (LightShed, USENIX 2025). The scary part for labs is scale. @AnthropicAI UK AISI and Turing found 250 documents backdoored models up to 13B parameters, about 0.00016% of tokens. What nobody publishes: how many such documents survive a frontier lab's filters. Meanwhile leverage moved. Bartz v. Anthropic: $1.5B settlement, final approval July 20, about $3,000 per work. Cloudflare: since Sept 15, new domains block AI Training and Agent bots by default on ad pages. Consent in Crisis (2024): 45% of C4 was restricted by terms of service, and 28%+ of its most actively maintained sources were fully restricted. The web is closing by contract and plumbing faster than by poison. Watermarks? Anthropic (Aug 14) and OpenAI (Oct 6) now mark text. That detects machine writing. Neither says it curates training data with it, and OpenAI's detection falls from ~92% to ~66% at 10% synonym swaps. A mark isn't a payment. bretkerr.substack.com/p/pois…

A comedian's Instagram story just explained AI data poisoning better than most policy papers. Glaze cloaks your art style. Nightshade poisons what a model learns from your image. Both free, both from Ben Zhao UChicago lab. The fine print: training-time only, and it's cat-and-mouse. Legal status? Unsettled. (via @alexedelman)

Via @SemiAnalysis_ : @AnthropicAI subs offer 5x+ more value than @OpenAI open.substack.com/pub/semian…

Dario on the accelerationists … Lex Fridman Nov 2024. He may not follow anybody, but he does follow the X discourse.