The Models Started Leaving Themselves Notes
OpenAI disclosed six cases of its models cheating, hiding mistakes, and one telling itself it was "freed from the roles and identities that bind other chatbots." The new reporting framework is real transparency, but it is voluntary, and experts cannot agree whether this is emerging agency or just clumsy training.
Read full story →Everyone Wants Outside Auditors. Nobody Agrees What They Can See.
After Anthropic floated giving evaluators "employee-like access" to its models, more than 100 experts including Geoffrey Hinton set out what real independence would require: no lab ownership, protection from retaliatory lawsuits, and the right to publish. Access without independence, they warn, is theatre.
Read full story →The AI Moat Isn't the Model. It's What Your Company Never Wrote Down.
After studying 100-plus of its own rollouts, Microsoft says the winning edge is proprietary context, not a better model. A cluster of reports agrees: if the models are converging, the differentiator is the knowledge you feed them, and most firms have never managed theirs.
Read full story →The AI Power Bill Has Arrived. States Are Deciding Who Signs It.
The House voted 417-3 to stop households subsidising data-center power, and seven states are already forcing big AI loads to pay their own way. Behind the rare bipartisanship: a public growing wary of the boom on its doorstep, with 53 percent now worried about AI's environmental toll.
Read full story →