tldr * Frontier models can be prompted into a weaker model's capability tier, but not its identity: they adopt a generic weaker-model error pattern, not a specific predecessor's per-question fingerprint. * Targeted sandbagging capabilities: where a stronger model throttles down to a weaker one without reasoning showed as a largely...
tldr * We explore stylized but concerning out-of-context reasoning where the same prompt can contain computation-dependent hidden information that influences a stronger model, but does not a weaker one e.g. a cost-effective monitor. * This is done through simple JSON “pointer-chains” in prompts. Chain-depth tunes difficulty (deeper-chains influence Opus but...
In Brief: I've been developing an interactive tool that I believe is helpful in accelerating transformer mechanistic analysis and that has the potential to reduce the barrier of entry. Motivations For a while now, my focus has been shifting towards alignment research, but getting involved and building intuition in this...