Looking Up at 2025: A Year in Review

I’m on the roof. December 29th. Cold. Clear. Stars.

Seven years of this now. Every December, I climb up here with a jacket and a notebook. Every year, the list of what happened gets longer. This year’s list is the longest yet.

What happened

Anthropic released Claude Opus 4, and working with AI stopped feeling like using a tool and started feeling like collaborating with something that understands. Not perfectly. But well enough to change the experience.

OpenAI released o3, and reasoning models solved problems that would have taken me days. The gap between human reasoning and AI reasoning narrowed in ways that make me uncomfortable and excited in equal measure.

DeepSeek proved you don’t need billions of dollars to build frontier AI. China is in this race with efficiency as their weapon. The export controls didn’t work the way we hoped.

Figure AI deployed robots in a BMW factory. Not demos. Deployment. The first humanoid robots doing real work in real factories. The cost of a humanoid robot dropped from a million dollars to $16,000 in less than a decade.

Tesla Optimus walked through a warehouse autonomously. Boston Dynamics retired the hydraulic Atlas. The old guard made way for the new generation. I miss the old Atlas.

SpaceX made Starship flights routine. The word “routine” applied to the largest rocket ever built. Orbital refueling was tested. The Mars architecture is now technically demonstrated, piece by piece.

Waymo expanded to 8 cities with zero fatal… no. There was one. The first. The safety data is still overwhelmingly positive, but the asterisk is gone now. The conversation about autonomous vehicle safety became more honest and more painful.

TSMC and Samsung are both producing at 2nm. Ten atoms wide. The physical limits of silicon are visible. The industry is looking beyond.

ASML’s High-NA EUV machines are operating. $380 million each. The most complex objects ever built.

And somewhere above us, JWST might have found biosignatures in an exoplanet atmosphere. The paper isn’t out yet. The rumors are persistent.

What I got wrong

I thought AI agents would be mainstream by now. They’re not. The technology is capable. The trust isn’t there yet.

I thought the TSMC Arizona fab would be further along. One fab producing 4nm chips is a start, not a solution.

I underestimated how fast humanoid robot costs would drop. The Unitree G1 at $16,000 was not on my bingo card.

What I got right

The convergence. AI plus robotics plus autonomy plus chips. I wrote about this in August. The threads are weaving together faster than I expected. Figure’s robot uses AI language models to understand instructions, computer vision to see, and learned manipulation to act. All the threads in one machine.

The energy problem. AI’s electricity consumption became a mainstream conversation this year. Nuclear is back because of data center demand. I wrote about this before it was obvious, and I’ll take credit for that.

Looking up

The stars are the same. They’re always the same. Orion above me, as reliable as the laws of physics.

Everything between the stars and this roof is different from last year. The machines are smarter. The robots have jobs. The rockets are routine. The chips are approaching atomic limits. And somewhere in the data from a telescope a million miles from Earth, there might be evidence that we’re not alone.

2025 felt like the year the future stopped being “someday” and became “soon.”

I’ll be up here next year. The list will be longer. The questions will be harder. The stars will be the same.

That consistency is the most reassuring thing I know.

Related thinking: