I love the phrase “intelligence too cheap to meter”. For us, that plays out as every random piece of our lab equipment soon will have a 140+ IQ. An anecdote was that in the early days, as we were quickly scaling up materials synthesis, we had technicians and scientists recording data on our characterization machines. This is the usual way. But when the scientists couldn’t keep up with our throughput, we made a script to programmatically capture the data. That scaled nicely, but we found the data wasn’t particularly useful. It wasn’t a complete or intelligent record of what happened experimentally. It also didn’t have the context for what we intended to do. We weren’t only interested in the mean behavior but were looking for anomalies or details that programmatic capture often missed. Now, with improving AI that is (almost) too cheap to meter, our data capture is both scalable and intelligent. The system knows the intent behind what we’re doing, can review our full records, and can direct and find the interesting evidence using the machine. And that improved data capture yields ever more useful future AI systems. It’s our own real lab version of Rick and Morty’s butter robot: “what is my purpose?”, “you watch the characterization machine”, “oh my god”.