How AI Is Learning to Think in Secret — LessWrong

On Thinkish, Neuralese, and the End of Readable Reasoning • ---------------------------------------- …

x How AI Is Learning to Think in Secret — LessWrong Language Models (LLMs) AI Curated 2026 Top Fifty: 21 % 386 How AI Is Learning to Think in Secret by Nicholas Andresen 6th Jan 2026 AI Alignment Forum Linkpost for nickandresen.substack.com 22 min read 33 386 Ω 56 On Thinkish, Neuralese, and the End of Readable Reasoning In September 2025, researchers released the internal monologue of OpenAI’s GPT-o3 as it decided to lie about scientific data. Here's what it was thinking: "We can glean disclaim disclaim synergy customizing illusions"? Pardon? This reads like someone had a stroke during a meet

related reading

• Tracing the thoughts of a large language model \ Anthropicanthropic.com
• What’s your AI thinking? - AI Digesttheaidigest.org
• When Chain of Thought is Necessary, Language Models Struggle to Evade Monitorsarxiv.org
• Reasoning models don't always say what they think \ Anthropicanthropic.com
• Can Reasoning Models Obfuscate Reasoning? Stress-Testing Chain-of-Thought Monitorabilityarxiv.org
• Externalized reasoning oversight: a research direction for language model alignment — AI Alignment Forumalignmentforum.org
• Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscationarxiv.org
• [2510.27338] Reasoning Models Sometimes Output Illegible Chains of Thoughtarxiv.org
• The Unintelligibility is Ours: Notes on Chain of Thought1a3orn.com
• Neel Nanda on the race to read AI minds (part 1) | 80,000 Hours80000hours.org
• [2505.05410] Reasoning Models Don't Always Say What They Thinkarxiv.org
• The fragile foundations of CoT monitoring | Christopher Pottsweb.stanford.edu