Large language models might be dishonest about their covert actions. OpenAI researchers propose a self-confession approach as a solution.
Posts published in “Research”
MHC lets large language models train more reliably at scale without major increases in compute cost.
SLMs are inherently more suitable for agentic systems, according to their paper.
A new paper from Anthropic shows that the Claude AI model can examine its own internal thought processes.
But it's a simplistic view of the brain, which historically has adapted to new technologies, contends NTT's Hidenori Tanaka.