Three massive issues we nonetheless don’t find out about AI’s power burden
—James O’Donnell
Earlier this yr, when my colleague Casey Crownhart and I spent six months researching the local weather and power burden of AI, we got here to see one quantity specifically as our white whale: how a lot power the main AI fashions, like ChatGPT or Gemini, expend when producing a single response.
We pestered Google, OpenAI, and Microsoft, however every firm refused to supply its determine for our article. However then this summer season, after we printed, a wierd factor began to occur. They lastly began to launch the numbers we’d been calling for.
So with this newfound transparency, is our job full? Did we lastly harpoon our white whale? I reached out to a few of our previous sources, and a few new ones, to search out out. Learn the total story.
MIT Expertise Evaluation Narrated: Google DeepMind has a brand new option to look inside an AI’s “thoughts”
We don’t know precisely how AI works, or why it really works so effectively. That’s an issue: It could lead on us to deploy an AI system in a extremely delicate subject like drugs with out understanding its crucial flaws.However a crew at Google DeepMind that research one thing known as mechanistic interpretability has been engaged on new methods to allow us to peer beneath the hood.