My 2025 prediction scorecard: I got the robots

Every January, I look at what I predicted and grade myself honestly. The grades are always humbling.

Predictions I got right

Humanoid robots: demo to deployment. I wrote in January 2025 that this would be the year humanoid robots moved from demonstrations to real work environments. Figure in BMW factories. Agility Digit in Amazon warehouses. Tesla Optimus in Tesla’s own factories. Grade: A.

The energy problem becoming mainstream. I predicted that AI energy consumption would become a front-page story. It did. Microsoft restarting Three Mile Island. Amazon buying nuclear capacity. Data centers consuming city-scale electricity. Grade: A.

Open source AI closing the gap. I predicted that open-weight models would reach parity with closed models within 12 months of each generation. Llama 4 and DeepSeek R1 confirmed this pattern. Grade: A.

Predictions I got wrong

Reasoning model plateau. I predicted that the improvement curve for reasoning models would flatten in 2025, the way the GPT-4 to GPT-4.5 improvement was smaller than GPT-3.5 to GPT-4. OpenAI’s o3 and Anthropic’s extended thinking proved me wrong. Reasoning is still climbing. Grade: D.

Self-driving expansion to 10 cities. Waymo reached 8 cities. Close but not 10. The expansion slowed in Q4, possibly due to the fatal accident and increased regulatory scrutiny. Grade: B-.

AI agents going mainstream. I thought 2025 would be the year people routinely let AI agents act on their behalf. The technology exists. The trust doesn’t. Most people still use AI as a question-answering tool, not an autonomous agent. Grade: C.

Overall: 3 right, 3 wrong

50%. Coin-flip accuracy. Which is, I think, the honest baseline for predicting the future. Anyone who claims higher accuracy is either lying or lucky.

Lessons for 2026

Don’t bet against reasoning. The improvement curve hasn’t flattened. Whatever is happening inside these reasoning models has more room to run than I expected.

Trust is the bottleneck. The technology for AI agents, self-driving, and autonomous robots is ahead of society’s willingness to use it. The limiting factor isn’t capability. It’s comfort.

The physical world moves slower than software. Robots, fabs, rockets. These things are physical. They require manufacturing, supply chains, regulatory approval. Software can update overnight. Hardware can’t.

I’m making my 2026 predictions privately and I’ll grade them next January. The record says I’ll get about half of them right. That’s fine. Prediction isn’t the point. Paying attention is the point.

Related thinking: