SPOTLIGHTMeta's AI comeback runs through open models, again  Meta's latest model marks a return to the playbook that once made the company important to the AI space. Muse Glimmer is built for AI agents that can write code, use software tools, and recover when a task goes wrong. More importantly, developers can download and run it on a high-end computer instead of paying Meta each time they use it. The company is willing to go even further, as it plans to release the open-weight version of its flagship model Muse Spark 1.2 in the coming weeks. Call this a comeback, if you will, and Mark Zuckerberg's lengthy essay conveniently sets out the policy environment Meta wants for that to happen. Take distillation, which allows an AI model to learn from the output of a stronger one. Zuckerberg argues that restricting it would hinder open-source development. There is truth to that, and Meta itself used the technique to transfer capabilities from Muse Spark into Glimmer. But permissive rules would preserve Meta's ability to learn from any model it can observe. That matters for a company still trailing OpenAI and Anthropic. While there's no public evidence that Meta plans to distill OpenAI's or Anthropic's models, Zuckerberg's preferred policy would keep that route open. His call for faster construction of energy infrastructure and data centers has the same dual character. Zuckerberg frames it as a national race: the US must build faster to maintain its lead over China. Yet faster approvals and more energy would also benefit Meta's own infrastructure ambitions, which include plans to sell excess compute capacity to AI companies, similar to xAI's playbook. If adopted, Zuckerberg's proposals could benefit the open-source ecosystem and strengthen US competitiveness. They would also give Meta a leg up in the AI race. |
