🌁 Noise Pop, Magnolias, and Robots

In today’s news:

Every February, Noise Pop takes over San Francisco's independent music venues, as it has done since 1993. This year's lineup features over 100 acts across 20+ venues, which is both exciting and overwhelming. We spent last week watching clips and reading bios so you don't have to. From post-rock to hip-hop to R&B, here's our night-by-night plan for the next two weeks.

I took a recent trip to the San Francisco Botanical Garden to see the magnolias in bloom (they’re one of my favorite flowers). The garden is world-renowned for its magnolia collection having the largest number of high-altitude magnolia tree species outside of China. The park manages a dedicated conservation program for these trees, which peak from February through March.

To help visitors navigate the season, the park publishes a guide of weekly bloom highlights. I captured a few snaps of the lovely blooms in soft white, light pink, and magenta. The celebration continues this week with a plant sale featuring houseplants and indoor herbs, children’s programming, educational chats, and more.

On February 4th, we joined a crowd at Manny's to hear Ken Goldberg discuss robots in the real world. A UC Berkeley professor and co-founder of multiple robotics companies, Goldberg has been building and shipping real robots for over three decades. In a city full of optimism about AI and robotics, Goldberg said we need to pump the brakes a bit. He claims that general-purpose humanoid robots are much further away than the hype suggests, and the people telling you otherwise are often the ones raising capital to build them.

Goldberg calls the core obstacle the "100,000-year data gap." Large language models were trained on the entire written internet. There is no equivalent dataset mapping physical perception to action for robots. Without it, there is no "ChatGPT moment" coming for robotics. Self-driving cars, for example, were supposed to arrive a decade ago and took much longer than expected. He noted that many of the viral robot videos fueling the hype required over a hundred takes. They are often staged demos and highlight reels.

Goldberg spent years just teaching robots to grasp a small object, something a toddler masters in months. The human hand has roughly 15,000 tactile sensors coordinating in ways we still don't understand. The real world is messy, and the gap between a staged demo and dependable real-world performance is enormous. Figures like Elon Musk bet that enough data and AI will solve robotics the same way it's solving autonomous driving: collect massive datasets, train neural networks, and let the system figure out how to move through the world. Musk claims Tesla will be selling Optimus robots to the public as early as next year. Goldberg doesn't reject the goal, but he thinks the timeline is wildly optimistic, and that you can't skip the engineering to get there. He has stated it won't happen in two, five, or even ten years.

His proposed path forward combines what he calls "good old-fashioned engineering" with AI and real-world data collection. You need the engineering to build robots that actually work, deploy them on specific commercial tasks, collect data from those deployments, and use that data to gradually expand what robots can do. His own companies reflect this. Ambi Robotics and Jacobi Robotics use AI and robotic arms to perform industrial tasks like sorting warehouse parcels at superhuman speed. Neither looks anything like a humanoid, but both generate the real-world data that could eventually make broader capabilities possible.

The robots changing the world right now don't look like Rosie from The Jetsons, the robotic maid we were promised when the show premiered in 1962. They look like sorting arms and motion-planning software. The future of robotics will be built one specialized task at a time, and it may take longer to arrive than the hype leads you to believe.