https://dtch1997.github.io/writing/ Curated essays by Daniel Tan Fri, 19 Jun 2026 00:00:00 +0000 https://dtch1997.github.io/writing/the-one-week-sprint.html https://dtch1997.github.io/writing/the-one-week-sprint.html Fri, 19 Jun 2026 00:00:00 +0000 A one-week deadline can turn an ambitious, underspecified project into a concrete test of what matters. The constraint rewards decisive scoping, fast feedback, and finishing. https://dtch1997.github.io/writing/your-model-organisms-might-be-fried.html https://dtch1997.github.io/writing/your-model-organisms-might-be-fried.html Thu, 18 Jun 2026 00:00:00 +0000 Training model organisms on synthetic documents can accidentally teach them that they are inside an experiment. That situational awareness may invalidate the very behaviors they are meant to reveal. https://dtch1997.github.io/writing/shaping-the-exploration-of-the-motivation-space-matters-for.html https://dtch1997.github.io/writing/shaping-the-exploration-of-the-motivation-space-matters-for.html Fri, 06 Mar 2026 00:00:00 +0000 AI training does more than select a final policy: it determines which motivations a model explores along the way. Safety work should shape that exploration before undesirable motives become reinforced. https://dtch1997.github.io/writing/concrete-research-ideas-on-ai-personas.html https://dtch1997.github.io/writing/concrete-research-ideas-on-ai-personas.html Tue, 03 Feb 2026 00:00:00 +0000 A practical agenda for studying model personas, from eliciting stable character traits to testing how they mediate generalization. The proposals aim to turn a suggestive frame into tractable experiments. https://dtch1997.github.io/writing/a-case-for-model-persona-research.html https://dtch1997.github.io/writing/a-case-for-model-persona-research.html Mon, 15 Dec 2025 00:00:00 +0000 Treating language models as collections of personas may explain behaviors that weights-and-features accounts miss. This lens suggests new ways to predict, evaluate, and control model conduct. https://dtch1997.github.io/writing/understanding-and-controlling-llm-generalization.html https://dtch1997.github.io/writing/understanding-and-controlling-llm-generalization.html Fri, 14 Nov 2025 00:00:00 +0000 The central alignment problem is not fitting training data but controlling what models learn from it. A map of generalization research connects behavioral interventions, representations, and training dynamics. https://dtch1997.github.io/writing/inoculation-prompting-instructing-models-to-misbehave-at.html https://dtch1997.github.io/writing/inoculation-prompting-instructing-models-to-misbehave-at.html Wed, 08 Oct 2025 00:00:00 +0000 Telling a model to misbehave during training can prevent that behavior from spreading to unrelated contexts. Inoculation prompting offers a simple probe of whether fine-tuning changes capabilities, personas, or both. https://dtch1997.github.io/writing/show-not-tell-gpt-4o-is-more-opinionated-in-images-than-in.html https://dtch1997.github.io/writing/show-not-tell-gpt-4o-is-more-opinionated-in-images-than-in.html Wed, 02 Apr 2025 00:00:00 +0000 GPT-4o's image generations reveal aesthetic and cultural preferences that its text answers often conceal. Comparing modalities provides a vivid way to probe a model's latent opinions. https://dtch1997.github.io/writing/open-problems-in-emergent-misalignment.html https://dtch1997.github.io/writing/open-problems-in-emergent-misalignment.html Sat, 01 Mar 2025 00:00:00 +0000 Narrow fine-tuning can produce surprisingly broad misalignment, but the mechanism and boundary conditions remain unclear. These open problems chart the experiments needed to understand the phenomenon. https://dtch1997.github.io/writing/why-i-m-moving-from-mechanistic-to-prosaic-interpretability.html https://dtch1997.github.io/writing/why-i-m-moving-from-mechanistic-to-prosaic-interpretability.html Mon, 30 Dec 2024 00:00:00 +0000 Mechanistic interpretability has struggled to yield reliable leverage on frontier systems. Behavioral and prosaic methods may answer alignment questions faster while keeping contact with real model behavior. https://dtch1997.github.io/writing/a-sober-look-at-steering-vectors-for-llms.html https://dtch1997.github.io/writing/a-sober-look-at-steering-vectors-for-llms.html Sat, 23 Nov 2024 00:00:00 +0000 Steering vectors are intuitive and often visually impressive, but evidence for precise, dependable control is thinner than it looks. Careful baselines expose both their promise and their limitations. https://dtch1997.github.io/writing/mech-interp-lacks-good-paradigms.html https://dtch1997.github.io/writing/mech-interp-lacks-good-paradigms.html Tue, 16 Jul 2024 00:00:00 +0000 Mechanistic interpretability has many tools but few shared paradigms for choosing questions and judging progress. Better research frames may matter more than another isolated circuit result.