Summary I report evidence that conversational “operators” (e.g. summarize, critique, reframe) correspond to stable, decodable internal states in transformer mid-layers, distinct from surface instruction wording and generalizing across content. These states are weak but persistent, geometrically separable via simple centroid methods, and survive instruction masking. I’m sharing this as a...
This post extends on the importance of studying subjective experience by considering the quality and relevance of the evidence and its acceptability in scientific discourse. I understand that first person reports are typically not considered objective or reproducible and therefore quite often fail to be accepted as scientifically rigorous data...
Current AI alignment relies almost entirely on reinforcement learning, which is operant conditioning. We successfully use operant conditioning to train rats. The problem is that humans don't develop sophisticated moral reasoning through operant conditioning alone. We develop it through higher-order social learning: forming relationships, building trust, evaluating principles autonomously, and...
The Allure of the Dark Side: A Critical AI Safety Vulnerability I Stumbled Into TL;DR: A casual thought experiment led me to discover that philosophers and psychologists possess exactly the skills needed for sophisticated AI jailbreaking. Worse, they're being systematically undervalued by AI safety researchers, creating a recruitment pipeline for...
Epistemic Status: Conceptual framework with initial validation protocols designed. I am highly confident that the problem statement (the "measurement crisis" in AI evaluation) is accurate. I am cautiously optimistic that this psychometric framework is a robust and necessary direction for the solution. The most speculative parts relate to autonomous evolution,...