Key Takeaways

  • AI coding tools make output fast, but skipping manual implementation prevents researchers from developing scientific taste and intuition.
  • John Platt protects unstructured 20% time in his Google group without requiring status updates, giving engineers room to explore outside metric pressure.
  • Hyper-optimizing your schedule for immediate output causes tool overfitting, which narrows exploration to problems existing software already solves well.
  • Platt applied his background studying under Richard Feynman and John Hopfield at Caltech to build Google's ERA framework for climate interventions like contrail mitigation.
  • Rigorous science requires fighting self-deception, because the easiest person to fool is always yourself.

The Athlete Problem in Software

When code generation takes five seconds, the temptation is to never write a line by hand again. Platt sees a trap here. Having trained at Caltech under John Hopfield and Richard Feynman, Platt spent decades building systems from the ground up, from support vector machines to climate models for wildfire detection and contrail mitigation.

“I don't think domain expertise is going away,” Platt explains, “because it goes back to a lot of people who said it goes back to taste and trying to figure out how people get taste without doing all the grunt work.”

Platt compares building software to athletics. If an athlete hires someone else to lift weights, the barbell still moves, but the athlete never gets stronger. “You also have to develop the muscles,” Platt says. “So it's a little bit maybe like being an athlete.” Skipping the hard work of writing algorithms from scratch robs an engineer of intuition. When an AI tool outputs an answer that looks plausible but fails silently, only the person who understands the core mechanics will spot the error.

The Cost of Hyper-Optimization

The standard startup advice is to measure every hour and cut anything that does not push a metric. Platt argues this mindset ruins research. When you optimize every sprint for immediate velocity, you overfit to your current toolkit.

“There just seems to be this strong impetus in the world to just optimize and squeeze everything out,” Platt says, “but you do lose something when you hyper-optimize. You sort of overfit.”

To counter this, Platt protects traditional 20% time for his team at Google. He places zero reporting requirements on this time. “In my own group, I try very strongly to protect it,” Platt says. “It's like you can do whatever you want. If you want to learn stuff, if you want to try stuff, you don't even have to tell me.”

This unstructured time led to Google's Empirical Research Assistance (ERA) framework, which pairs language models with Monte Carlo tree search to run autonomous scientific experiments. If Platt had forced his team to justify every experimental sprint against quarterly deliverables, ERA would not exist.

The Feynman Standard of Verification

Platt keeps Feynman's classic rule front and center: "You absolutely cannot fool yourself, and you are the easiest person to fool."

AI tools generate confident, coherent prose and functional scripts on command. That fluency creates an illusion of understanding. When models confirm your initial hypothesis, the easiest reaction is to declare victory and ship. Platt's counterweight is verification against physical reality, whether tracking actual aircraft contrails to reduce heat trapping or measuring real wildfire spread. If your model cannot predict physical data under scrutiny, speed of execution does not matter.

What to Do With This

Block four hours this Friday on your calendar for unconstrained experimentation, and turn off your AI autocomplete tools during that block. Pick a core algorithm or component in your product and rewrite it from scratch using plain documentation. If you find yourself unable to write the logic without an AI prompt, you have found a blind spot where your domain taste is lagging behind your output.