🤖 Another vector for recursive improvement — using agentic AI to design better chips than any human. Keysight disclosed Sphere Semi's results today: "5-10x productivity gain, 6 dB better isolation, and 30% less area versus traditional design processes." Sphere uses a code-defined design loop running thousands of candidates through circuit and electromagnetic simulation. Keysight has separately documented first-pass fab success from Sphere's autonomous workflow "employing agents, generative models, and reinforcement learning". "Human designers often waste space by thinking, consciously or subconsciously, in terms of symmetry and reduced time to optimize manually. We focused on constraining the designs such that they would be DRC-clean by construction, leaving room for AI to innovate component placements, spacing, and structures to get to the best design, and it's almost always not what humans would choose to do." "Sphere Semi has already taped out working silicon for an AI-generated tunable bandpass filter design, which they believe is an industry first, validating that AI-designed layouts can translate into functional hardware." "We are now at the point where humans put in the spec, AI generates designs and test benches, and evaluations happen – with minimal human guidance, all in a few days." • The Keysight slides below show how the Sphere AI goes alien in the layout, but it works: • News: • Sphere Semi: