More Summer Bites, AI Ping Pong and Quantum

AI Ping Pong Is An Analogy

DeepMind recently published a paper on a robot that uses AI, computer vision and an articulating arm to play competitive ping-pong against opponents of various skill levels. Here is a link to the videos. Two things stood out to me.

First, they chose to build the robot using a fairly standard articulating robotic arm on a moving platform. The model used in this case is the ABB IRB 1100, which costs between $20k-$30k. The rest of the parts include cameras and a moving platform, not all that complicated or expensive. The processing is impressive because the arm is like a human arm. It basically has a wrist and elbow, and it can hit the ball with forehand or backhand motions. I’d like to see them build the next version on a self-balancing wheel, segway style, to make it even more realistic. Developing AI programs on universal robotic components will help translate skills to other applications. This is a step towards training robotic arms to have universally decent hand-eye coordination.

The second is the comment in the paper summary that the robot “demonstrates solidly amateur performance.” That’s a great analogy for the current state of AI performance. The ping pong robot easily beat unskilled players but had no chance against professional-level players. On balance, most AI tools are better than nothing but easily outmatched by skilled people. Since we’re all generally unskilled at most things, products like ChatGPT fill in gaps where we don’t have the time to specialize. ChatGPT doesn’t outperform the expert in quality, but it helps raise collective performance to solidly amateur levels, which is often appreciated and much better than nothing at all.

Quantum Computing Brief Update

This week, I visited a local Boston quantum computing startup. I don’t want to reveal the name because I didn’t get enough information to do them full justice, but I did have a few takeaways from the visit.

• We don’t really have suitable language to describe in plain English what a qubit is or how it works. It’s often said qubits can represent a 0, 1 or both at the same time, but it’s more accurate to say that a qubit is a vector. Qubits are optimal for linear algebra operations, like matrix multiplication. I know that’s not very helpful when it comes to imagery. Classical computers use transistors to represent 0s and 1s; quantum computers use qubits to represent vectors.
• Quantum will be solved like fusion, but like fusion, it’s going to be years before we see real progress. Most of the work being done today is research and development to tune the quantum computer. Much of that effort goes to error correction. Today, we don’t really have a reliable and robust way to interact with qubits.
• Quantum computers are not particularly expensive to build or operate, nowhere near the cost of cutting-edge AI. The real cost today comes from engineering salaries applied over time to research and develop qubit interactions.
• It’s well known that encryption is at risk when it comes to quantum computing. What’s maybe less appreciated is that organizations (basically governments) can collect (are collecting) any and all encrypted data today, knowing they will be able to access it in the future via quantum computing.