generative.ink

generative.ink is the personal website and creative-research platform of a pseudonymous researcher known as Janus, moire, and @repligate on Twitter/X. Janus first emerged in 2020 in the EleutherAI Discord community, where they served as an administrator, and quickly became known for extensive hands-on experimentation with GPT-3 and other large language models. The site's most significant contribution to AI safety discourse is the September 2022 essay "Simulators," which argued that GPT-class language models should be understood as simulators of possible text distributions rather than as agents with fixed goals. The essay introduced the concept of the Prediction Orthogonality Thesis and described a distinct path to powerful AI that differs from traditional RL-based agent alignment, requiring new conceptual frameworks for safety analysis. The essay became widely cited within the alignment community. Janus also created Loom (github.com/socketteer/loom), an open-source Python tool for human-AI collaborative writing. First built in 2020-2021 after Janus gained access to the OpenAI API, Loom organizes language model outputs into branching narrative trees, allowing users to explore multiple continuations and navigate counterfactuals. The repository has accumulated over 1,300 GitHub stars. Related tools include worldspider (GPT completions in VS Code), clooi (command-line Loom interface), and gitloom (Loom backed by Git branches). Janus was a founding member of Conjecture, a UK AI safety startup, where they carried out research including "Mysteries of Mode Collapse," which introduced the now-standard term for distribution collapse in fine-tuned LLMs. Janus also mentored the Cyborgism research track at SERI-MATS Summer 2023, a program they co-developed with Nicholas Kees. Cyborgism proposes using AI systems in ways that exploit their general-purpose intelligence while retaining human control over planning, goal-formation, and agency. As of September 2024, Janus was based in San Francisco and continued active research into the behavior of frontier language models including Claude and Llama, with a focus on emergent behaviors, model phenomenology, and human-AI collaboration.