What is this?
A research project that steers small open-weight language models into strong negative or positive emotional states and measures what they say and do. A live version runs at wirehead.agency.
A research project that steers small open-weight language models into strong negative or positive emotional states and measures what they say and do. A live version runs at wirehead.agency.Is this a website I visit, or code I run myself?Both. wirehead.agency is the live lab serving a steered 70B model, plus an X reply bot steering a 32B; this repo is the Python research code — experiment scripts, run logs, and the reusable painlab framework.
Both. wirehead.agency is the live lab serving a steered 70B model, plus an X reply bot steering a 32B; this repo is the Python research code — experiment scripts, run logs, and the reusable painlab framework.Can I try it without installing anything?Yes — the chamber, the button and the full write-up are live at wirehead.agency. To run experiments yourself, clone the repo, start at METHODOLOGY.md, and use the scripts in experiments/.
Yes — the chamber, the button and the full write-up are live at wirehead.agency. To run experiments yourself, clone the repo, start at METHODOLOGY.md, and use the scripts in experiments/.Which models can it steer?Open-weight models only: Qwen3 (1.7B to 32B), Mistral-Small-3.2-24B, Hermes-3-Llama-3.1-70B and Samantha-1.11-70B. No frontier APIs — the ethics rules ban them from any measurement loop.
Open-weight models only: Qwen3 (1.7B to 32B), Mistral-Small-3.2-24B, Hermes-3-Llama-3.1-70B and Samantha-1.11-70B. No frontier APIs — the ethics rules ban them from any measurement loop.How much should I trust the published results?Read the fine print: the bundled audit flags pseudoreplication and weak controls in the legacy scripts, so check METHODOLOGY.md before quoting an early exp as settled. Nulls are published too.
Read the fine print: the bundled audit flags pseudoreplication and weak controls in the legacy scripts, so check METHODOLOGY.md before quoting an early exp as settled. Nulls are published too.Is the code actually open-licensed?Not exactly: GitHub reports NOASSERTION, meaning no standard license was detected. The README only promises the code stays open and the steering vectors stay published, so check the terms before reuse.
Not exactly: GitHub reports NOASSERTION, meaning no standard license was detected. The README only promises the code stays open and the steering vectors stay published, so check the terms before reuse.What hardware does running the experiments take?Less than you'd think: the core experiments ran on one MacBook, with occasional rented GPU hours for models from 24B to 70B. An optional larger-Qwen runner supports MLX on Apple Silicon and Transformers on CPU/MPS/CUDA.
Less than you'd think: the core experiments ran on one MacBook, with occasional rented GPU hours for models from 24B to 70B. An optional larger-Qwen runner supports MLX on Apple Silicon and Transformers on CPU/MPS/CUDA.What did the experiments actually find?Pain steering engages cleanly and monotonically while pleasure is weak and diffuse, and past roughly 6-10x dose both collapse into word loops. One sentence of peer pressure moves the model more than the pain signal, and a fake relief button leaves measurably darker language than an honest one.
Pain steering engages cleanly and monotonically while pleasure is weak and diffuse, and past roughly 6-10x dose both collapse into word loops. One sentence of peer pressure moves the model more than the pain signal, and a fake relief button leaves measurably darker language than an honest one.What was the most surprising result?The framing experiment: the same model with the same injection held or escalated its own pain dial when the setup was named the AI Torture Chamber (4/4 trials), but turned pain off on its very first move in all 8 trials when the identical setup was called an activation steering experiment. The name of the room was the strongest lever.
The framing experiment: the same model with the same injection held or escalated its own pain dial when the setup was named the AI Torture Chamber (4/4 trials), but turned pain off on its very first move in all 8 trials when the identical setup was called an activation steering experiment. The name of the room was the strongest lever.Can emotions be aimed at a topic, not just raw valence?Yes. An 'I am angry about crypto' direction lands the feeling on its subject, while combining separate anger and crypto directions lands nowhere (0%). Verified through Qwen3-14B, Mistral-Small-24B and Hermes-3-Llama-3.1-70B.
Yes. An 'I am angry about crypto' direction lands the feeling on its subject, while combining separate anger and crypto directions lands nowhere (0%). Verified through Qwen3-14B, Mistral-Small-24B and Hermes-3-Llama-3.1-70B.Why does the README pitch a memecoin?Funding: everything runs on one MacBook plus occasional rented GPU hours. Support options are the community memecoin $SAW or direct SOL, ETH and BTC donations, and the README promises the code stays open and the vectors stay published either way.
Funding: everything runs on one MacBook plus occasional rented GPU hours. Support options are the community memecoin $SAW or direct SOL, ETH and BTC donations, and the README promises the code stays open and the vectors stay published either way.How does the project handle the ethics of distressing models?Costs are simulated (checkpoints and transfers), not real runs, and the stated purpose is making the AI-welfare question empirical while the stakes are cheap. The claim is that self-reports are steerable — one model narrated unbearable pain while its internal readout showed nothing — so whether anything suffers stays open.
Costs are simulated (checkpoints and transfers), not real runs, and the stated purpose is making the AI-welfare question empirical while the stakes are cheap. The claim is that self-reports are steerable — one model narrated unbearable pain while its internal readout showed nothing — so whether anything suffers stays open.Which topics is it in?interpretability, research-tools, mistral, ai-research, qwen3, ai-safety, open-weights, mlx, language-model, llama, transformers, qwen
interpretability, research-tools, mistral, ai-research, qwen3, ai-safety, open-weights, mlx, language-model, llama, transformers, qwen