Projects

Research

Optimisation over non-stationary distributions creates weirder mindslesswrong →Samuel Ratnam, Pjain · LessWrong, June 2026Training under sequentially mixed objectives produces ecological generalists, conditional policies or strategy churn, depending on how separable the environments are and how much circuitry the objectives share.
Why Jailbreaks Look Like Hypnosis: Parallels Between Human and LLM Cognitionpdf →Samuel Ratnam · SPAR Winter 2025Theoretical work on the parallels between human and LLM cognition.

Engineering the Generalisation Landscape of LLMslesswrong →Samuel Ratnam · LessWrong, July 2026Freezing a trait-eliciting LoRA while training the base model beneath it creates or removes entanglements between traits; disentangling emergent misalignment makes the model more robust across domains.
Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignmentarxiv →Cameron Tice, Puria Radmard, Samuel Ratnam, Andy Kim, David Africa, Kyle O'Brien · ICML 2026 (spotlight)How discussions of AI in pretraining data shape model behavior through self-fulfilling mechanisms. Accepted as a spotlight paper at ICML 2026 (top 2.2% of submissions).
Objective Matters: Fine-Tuning Objectives Shape Safety, Robustness, and Persona Driftarxiv →Daniel Vennemeyer, Punya Syon Pandey, Phan Anh Duong, Michael Umeokoli, Samuel Ratnam · arXiv 2025How different fine-tuning objectives impact language model safety and adversarial robustness at scale.

νgithub →spawn, send, watchA minimal actor calculus for Claude Code agents: spawn, send, watch. An agent is a row in a database, woken by a message and parked again after one turn.
OpenRecommendergithub →local, transparentA locally-running, AI-powered content recommendation system that prioritizes user transparency and control.
Poetry Copilotgithub →a thesaurus for verseAI writing assistant for poetry, providing inspiration, intelligent thesaurus, and editing suggestions.
Favour Auctiongithub →favours, at auctionA point-based favour trading platform where users submit favours and bid in auctions.
Non-Linear Writing Interfacegithub →branching draftsA keyboard-native writing application for crafting branching narratives using DAG visualization.