Nick Cammarata, a researcher at AI startup Goodfire, opened a 26,000-word blog post with a striking claim: "I think I have good news about the nature of neural networks." He argues that large language models may be far easier to understand, and therefore easier to monitor and control, than researchers previously believed. The post has sparked wild speculation as well as cautious celebration across the AI world, though not everyone is convinced. In an interview conducted the day after the post went viral, Cammarata discussed what he thinks researchers are getting wrong, what his work actually shows, and why skeptics are not ready to celebrate yet.
Separately, personal AI agents are blurring the long-standing divide between bots and people online. These agents are taking on tasks such as booking flights and reserving hotel rooms. Meta's Muse alone was downloaded 5 million times in its first month. The mix of agents and human visitors on an internet built mainly for people has gone poorly, and many websites now block AI agents. If demand keeps rising, the internet could be redesigned around agents rather than people.
Elsewhere, documents outline a US Army plan to let AI agents find and disable targets for hacking. Experts warn the software could damage unintended systems. Other stories in the newsletter cover a suspect in an OpenAI-related bugging case who allegedly had Saudi government support, a US ban on DeepSeek products in civilian government systems, and a European effort to build its own version of Starlink.
Source: MIT Technology Review · Summarized by HeadlinesBriefing