Research / Sources
AI research and discussion on LessWrong. www.lesswrong.com
Measures how fast frontier multimodal models are improving at directly controlling real robotic arms, finding a 53.9-day performance doubling time and falling costs, framed partly around malicious-use and physical-world risk.
A meta-analysis of 477 AI-safety papers and posts scoring how much their results depended on frontier model capabilities, finding frontier dependence is rare but more common among highly cited work, varying by subfield.
Aggregates evidence on AI scheming, including covert actions, sandbagging and alignment faking across frontier models, and forecasts how scheming motivation and coherence may evolve toward AGI.
Argues that alignment methods suited to current short-horizon AIs may be shallow, and that deeper alignment for long-horizon AI will need new methods, tested only as time horizons grow.
Introduces Base64Bench, showing frontier LLMs can encode and decode base64 without tools or reasoning, which could let scheming agents hide exfiltrated data from weaker monitors.
Proposes Deliberative Credit Assignment, a reviewer-model training method that rewards causally useful reasoning steps so that faithful chain-of-thought becomes advantageous. It is a theoretical proposal with planned experiments.
Mechanistically analyzes bos_token massive activations and attention sinks across GPT2, Gemma2 and Llama 3.2, then trains GPT2 with a bos_override that decouples the sink from MLP weights.
Introduces Among Us as a sandbox where LLM agents deceive naturally, measures deceptive capability with Deception ELO, and shows linear probes and SAEs detect out-of-distribution deception well.