AI-authored code for specific attention and MoE blocks ran 1.5 to 1.8 times faster than expert human code on the new Jalapeño chip. OpenAI used AI to move the hardware from design to tapeout in 9 months, building the architecture to be easily programmable by AI rather than human experts.

The company also purchased thousands of Mac minis and Mac Studios to train AI agents using reinforcement learning. Apple's unified memory architecture supports these workloads, which require agents to interact with real operating systems and GUIs, differing from traditional model pretraining.

Sign in to suggest edits

Key sources

  1. SUPPORT@learnopencv“the chip was built to be easy for AI to program”x.com
  2. SOURCE@openlabxorg“Anthropic is reportedly renting Mac minis through AWS for similar agent-training work”x.com
Markdown