Building in the Unknown

We are at an incredible point in technology, walking a fine line between what these new models can do and where a human still needs to hold the wheel and make the call. Inside organizations there is enormous resistance to letting go of control over things we have done the same way for twenty years. After months of conversations with executives and architects about systems that have never been built before, the same pattern keeps showing up: the initial idea sounds impossible, unfounded, and not worth the time, and the pushback steers everything toward immediately solvable, deterministic outcomes for the business. From inside two years of deep work with AI, that instinct looks shortsighted; plenty of what is now standard engineering practice seemed impossible even six months ago.

The Bottom of the Mountain

I’ve climbed half a dozen of the tallest peaks in the lower 48, and building in the unknown produces a feeling I recognize from every one of them. At the bottom, the mountain looks insurmountable. Driving up, watching the summit get bigger through the windshield, there is no visible path to the top, and you think repeatedly about turning around and doing something else with your day.

Part of the flow involves dropping off a ledge knowing there is ground below you without knowing exactly how you’ll land. Or leaping from one boulder to the next with hundreds of feet of exposed rock underneath, where one slip means you’re either dead or leaving in a helicopter. A friend of mine lived the second version last year on the Siphon Draw trail at Lost Dutchman State Park; one slip snapped his humerus in half and we had him airlifted out.

At the bottom, the mountain looks insurmountable.

Is This Even Real?

As we move fast and start replacing major pieces of how engineering work gets done, that exposure feeling shows up constantly. You oscillate between absolute confidence that this is the way and extreme uncertainty about everything ahead. Is this even real? Are the models convincing me this is good when it’s actually bad? What if it’s all a fad and ends tomorrow? How do I even begin to quantify the business impact of something that changes how an entire organization works?

Quantifying a new feature, a more efficient architecture, a streamlined process, or a computationally more correct algorithm is a known quantity with established patterns. Reasoning about changes that are still extremely early, and that rewrite the mental model behind every item on that list, is not.

Resistance as the Feature

After watching an interview with Dario Amodei, the pattern seems clear to me: keep pushing the vision forward and use the resistance as a feature. I’ve written about this before at the learning layer, deliberate friction between me and the models; this time it’s happening at the organizational layer, where whole industries are being asked to change their mental models at once. Resistance at that layer is extreme, and it’s also valid. Many things remain unknown, and many scenarios will not work easily, or ever, with nondeterministic next-word predictors. Yet.

For the first time in my life, the machines have caught up to how my brain works.

The Machines Caught Up

This has been the hardest part for me personally. I swing through what feel like manic episodes of disbelief and elation. New ideas arrive suddenly, out of nowhere, that would change how entire industries operate, and that is frightening at the same time that I have never felt so certain of anything in my life. My ideas sound crazy at first in most cases, and almost always, one or two months later, the people who resisted hardest come back to tell me how much the thing changed their lives, inside and outside of work. That loop is satisfying and intellectually stimulating in equal measure. I’m probably on some sort of spectrum, but for the first time in my life the machines have caught up to how my brain works.

This is probably the most incredible stretch of my career so far, a major shift in capability and reach arriving all at once. I’m dictating this article by voice from a mountain just outside Phoenix, the same way I shipped a post from a trail in March. By the time I’m back at the car, an agent running on a computer somewhere in the cloud will have cleaned this up for clarity, generated the audio version, and staged it for review, all autonomously.

Charge Forward

Keep pushing against the force that wants you to stop climbing, the one that says this is impossible and the models are not as good as you think they are. On this particular Friday, the models are the least capable they will ever be. That thought is inspiring and scary in the same breath. Lean into the fear and uncertainty, the way Pema Chödrön teaches in Comfortable with Uncertainty, and let the resistance push you to think deeply and critically about what is not currently possible. Dario bet on scaling compute while serious people in the field were in violent disagreement, and a few years later the result is so obvious we forget it was ever contested.

On this particular Friday, the models are the least capable they will ever be.

Fable 5 from Anthropic is a significant shift in model capability. People are realizing it now, and it’s another moment like last November with Opus. Headlines will shout absurd claims and fear in both directions. Underneath the noise, this is another slingshot for the human race, for our ability to work alongside machines as a superpower. Charge forward, optimistically skeptical.