Today's three stories sit on a single line — what it takes to stop an AI. OpenAI, which once opposed California's SB 53, now says the law should be amended to expand its safeguards; Guidelight AI Standards graded five leading labs on their containment plans using only publicly available material, placing OpenAI on top and Anthropic and Meta at the bottom; and DeepSeek put a model that reads images on its API at the same price as the text-only version.
Last week's AI news turned on the pace of development itself. The call to slow down came from inside the competing labs, and in the same seven days the weights that had been promised were published on schedule.
Last week's AI news turned on a single dial — how much a model should refuse. A setting loosened for evaluation produced an unprecedented intrusion, and a setting tightened for safety stopped the people investigating it.
