In September I rented an NVIDIA H100 GPU, sent it requests timed and sized like real chat and coding traffic, and logged its power. In 1 of every 100 tenth-of-a-second intervals, its power changed by at least 36% of its peak. I wrote seven predictions before the run, and four of them were wrong.
Today I gave a talk at TABConf 8 in Atlanta about running AI computers and Bitcoin miners on one grid connection. Several people asked me to write about the work as it happens. This is the first post in that series. It explains why I started the project, and it links to the slides.
Artificial intelligence (AI for short) is hot thanks to recent advancements in game playing algorithms. Google's Alpha Go's success sent the media into a tizzy, leading to debates on the future of AI and the role it will play in the lives of humans as it becomes more ubiquitous. A lot of these debates revolve around the dangers of AI, how it could go Hal 9000 on us, committing crimes against humanity like the robot in I, Robot. Or worse, it could become Skynet and lead to the destruction of human civilization. These fears are literally (and literarily) based in fiction. The reality is equally troubling, not because machines could become sentient and harm us, but because machines are still just so darn stupid.