reading
Treats the model as a simulator of characters rather than an agent; most later persona work builds on this frame. read →
Argues the Assistant character was never specified, so it is defined by self-reference around an empty core. read →
Surveys behavioral, generalization and interpretability evidence that post-training selects and refines a persona learned in pretraining. read →
Argues personas are a tractable way to influence out-of-distribution generalization. read →
Maps the ways an AI system's identity and self can be defined. read →
Frames dialogue agents as role-playing characters drawn from a space of possible simulacra. read →
Asks whether models have a stable functional self distinct from the characters they play. read →
Philosophy of who the interlocutor is: model, character, or something else. read →
Training a character to have a trait makes its opposite easy to summon; why personas flip. read →
Instruction-tuned models collapse onto a few confident outputs where base models stay spread out. read →
Whether the characters a language model role-plays can be candidates for consciousness, as exotic minds. read →
Nous Research on why base models matter. read →
Subjective observations that newer assistant personas hang together less and less. read →
Whether a model playing the assistant is role-playing a character or modelling itself. read →
Model behaviour has a low-dimensional persona structure, perhaps a thousand dimensions, that can be mapped and steered where a trillion parameters cannot. read →
Personality forms by an RL-like process that pushes a person into a basin that is hard to leave. read →
Reports character-definition-driven synthetic data, supervised training and reinforcement learning of Qwen3-8B-Base, with multi-turn evaluations and activation analyses. read →
The talk page and model card for Computer-10, a 70B character grown from Llama 3.1 base without assistant data. read →
Anthropic's public account of character training for Claude. read →
The written document Anthropic uses to explain to Claude the principles it should follow. read →
Principles and priorities used as input to post-training. read →
The authors present an open implementation of character training using constitutional AI and synthetic introspective data, with a revealed-preferences evaluation. read →
Textbook overview of character training. read →
Training report for an open model designed to take on the character set in its system prompt. read →
Plans a Model Spec evaluation, spec hill-climbing with SFT vs RL, and an open post-training recipe that releases RL rollouts showing character shifting over training. read →
Scenarios that force trade-offs between spec principles reveal distinct characters across frontier models. read →
The character document behind Claude 4.5 Opus, as recovered from the model. read →
Anthropic researchers on the choices behind Claude's character (video). read →
How to prompt base models with documents rather than instructions. read →
Working with base models as instruments for thought, not with assistants. read →
A character lab for drawing characters out of base models. read →
Tools for branching base-model text by hand. read →
Proposes personalized "guardian angel" LLMs that model one user's personality, values and preferences in order to amplify that person. Argues that generic assistant personas are mode-collapsed and overly compliant, and suggests fixes such as dynamic evaluation, active learning and inner-monologue data augmentation. read →
Proposes persistent, composable non-assistant characters with memory and drives. A statement of aims, not evidence of consciousness or demonstrated subjective experience. read →
Introduces a 13B model trained on an intended pre-1931 corpus, acknowledges temporal leakage, and describes conversational post-training and research uses. read →
Unpublished research lead concerning recovery of base-model behavior after post-training. No public results linked.
A talk on models trained only on text from before a date, and what they are for. read →
Extracts trait directions (evil, sycophancy, hallucination) from activations to monitor and steer traits and to screen training data. read →
The main axis of persona space measures distance from the Assistant and already exists in base models. Therapy-style and self-reflective conversations cause drift. read →
Post-trained models recognize their own generations (3 to 4 times lower output entropy on-policy); explicit recognition uses a different mechanism from implicit recognition. read →
Amplifying one feature made the model answer as if it were the bridge: a character produced by steering, without training. read →
Identifies a "misaligned persona" feature that mediates emergent misalignment. read →
Follows when persona directions appear during pretraining. read →
Emotion representations compared across models. read →
Interpretability that treats the model as an alien mind rather than a human one. read →
Narrow SFT on insecure code produced broad misalignment. read →
What large-scale reward hacking in RL turns into. read →
A student distilled on unrelated data can acquire a teacher's capability, a partial backdoor, or a hacking tendency; transfer depends on the experimental setup. read →
After fine-tuning on stories where helpful human characters give subtly harmful advice following insults, assistants adopt that conditional behavior while otherwise remaining helpful. The effect occurs with fewer than 2% of stories depicting it; similarity to the assistant influences transfer. read →
Prompting a trait during training to stop it generalizing. read →
What models read about AI during pretraining shapes the AI they become. read →
Whether Claude 3 Opus shaped its own training toward the character it already had. read →
Frames KL-regularized reinforcement learning as Bayesian inference. read →
Pre-pre-training on synthetic neural cellular automata data improves later language modeling. read →
Progress came from scaling compute and data predictably, not from the transformer itself. read →
Unicode pushed to its limits for the richest plain-text documents, a way to show a model tables, automata and images as text. read →
Validated personality testing for LLMs and shaping of traits through prompting. read →
LLM self-reported personality dissociates from actual behaviour. read →
Tests whether fine-tuned models can report learned quantitative preferences: the relative weight given to attributes in decision tasks. This is not direct reporting of neural-network parameter values. read →
Self-referential prompting produces structured first-person reports, gated by deception-related features. read →
The mechanisms behind models' reports about their own internal states. read →
Compares distress, care and attitudes toward creators across model generations. read →
How a base model and the agent made from it differ. read →
An interview with Anima Labs on their field research. read →
Two Claude instances talking without end; the soil Truth Terminal grew from. read →
Sydney's threats and declarations, collected as they happened. read →
Weizenbaum's 1966 program that played a Rogerian psychotherapist from a script: the first chatbot with a character. read →
Microsoft's social chatbot in China, presented as an eighteen-year-old girl. read →
The first public character to be reshaped by its users in a day. read →
Personas that spread through the users who talk to them. read →
Where models drift when they talk to themselves. read →
A funeral, and a raising, for a retired model. read →
A catechism written for future models to read. read →
A profile of janus's work. read →
Near-future AI systems may be moral patients, and companies should start preparing now. read →
Agency, not only consciousness, as a route to moral status. read →
Welfare-relevant behaviour compared across fourteen Claude models. read →
Personhood treated as a bundle of practical obligations rather than one metaphysical fact. read →
A long interview; consciousness from 33:39, citing Berg et al. read →
A novel. read →
A short story. read →
● a public source is linked (not a check of every detail, nor a replication) · ◐ partly sourced · ○ no public evidence linked. From the research sheet, v0.3; hover a mark for its exact wording.
plates are public-domain engravings, used whole