Five ideas, papers, and products shaping our investment thinking, and what we think they mean. This series is powered by an AI assistant that helps synthesize recurring themes from our discussions, alongside our own reflections.
Google pays homes and businesses to use less power. The grid serving much of the East Coast is running short of power. Instead of waiting for new plants, Google is paying Voltus to round up batteries, smart thermostats and factory equipment that can cut back when the grid is strained. The deal now covers more than a gigawatt, about one large nuclear reactor’s worth, through 2040. This matters because data centers can’t plug in until there’s power for them, and freeing up existing power is faster than building new plants. We think power you can switch off is becoming something hyperscalers buy, which opens room for startups that find, control and finance those small sources.
Claude designed up to 10x better quantum circuits. Quantum circuits are the step-by-step programs a quantum computer runs. The fewer qubits and steps they need, the sooner real hardware can run them. Anthropic gave Claude eight standard circuits and a program that scored each attempt. With no human help, Claude beat the best published design on seven of the eight. This matters because AI gets better fastest wherever it can check its own work instantly and cheaply. We think the bigger story is crypto: one of these circuits is a key step in breaking the encryption behind crypto wallets, which makes post-quantum upgrades more urgent.
Odyssey powered a self-driving car with only 20 hours of driving data. Waymo and Tesla needed years and millions of miles of data. Odyssey’s model first learned how the world works by watching video and games, so it needed only a little real driving to handle roads in India. The same model also became the base for robot arms and a humanoid. This matters because if one model can learn many machines from little data, robotics stops being a costly data race for each one. We think world models become the foundation of robotics, and startups win by building on them instead of collecting data from scratch.
Robots are learning from their mistakes, not yet learning how to learn. It’s the difference between a student reviewing wrong answers and a student finding a better way to study. A new 93-page survey finds robots now do the first, but no study yet shows the second. Software can attempt it because it tests every change instantly, while robots pay for every try in time, money and sometimes broken hardware. This matters because “self-improving robots” is becoming a pitch that usually means only the first. We think the teams worth backing can show, and measure, their robots getting better with every round of deployment.
AI models may be much easier to understand than we thought. The common view is that each “neuron” inside a model mixes many unrelated concepts, which makes models hard to read. Goodfire’s Nick Cammarata argues that is mostly a measurement problem: in an early check of one open model, about 90% of neurons had one clear job. This matters because a model you can read is a model you can audit and fix. We think that helps open-weight models most, since anyone can inspect them, which gives them an edge wherever regulators want to know how a model works.
We’ll share another edition next week.