Where I start
- What am I losing when I use large language models and generative and agentic AI?
- What am I gaining?
- What is the opportunity cost, seen from both directions?
Opportunity cost runs in both directions. If I do by hand what a model does well, I pay in time, and other work waits. If I hand the work over, I pay in practice and in understanding. Most of the questions on this page come down to weighing one cost against the other. The questions below are where those three land in practice.
AI and machine learning
Two tools, two kinds of question.
I feel differently about AI than I do about machine learning, and they raise different questions. A language model is itself built with machine learning, so the line I'm drawing is about how I use each one.
Generative and agentic AI
- What it is
- Language models that write, reason, and act on my direction.
- What it does for me
- It takes on part of my thinking, which frees my mind and my time for other thinking.
- What it asks of me
- How much should I hand over, how do I check the work, and how much of it is mine?
Machine learning
- What it is
- Models that learn patterns from data, such as a neural network trained on decades of weather.
- What it does for me
- It gives me a second way to see a problem I already study with physics.
- What it asks of me
- What can a model know without being taught the physics, and what does it mean when the two agree?
Generative and agentic AI
What I hand over, and what I keep.
How much of my thinking should I hand over?
A language model can take on part of my thinking. That frees my attention and my time for other problems, and it's a large part of why I can carry as many projects as I do.
How much can I hand over before I lose the ability to think critically about what I handed over?
I believe there's a sweet spot. Short of it, I spend hours on work a model does well while the problems that need my judgment wait. Past it, I've given away so much of a subject that I can no longer evaluate what the model gives back.
I haven't found a rule for locating that spot. My practice for now is simple. When a model takes me somewhere I can't follow, I stop and learn the ground before I go further.
What do I lose when I no longer build by hand?
I used to build things slowly. I tinkered with amateur radio, Raspberry Pis, and Arduino boards, and about ten years ago I built a small Raspberry Pi cluster to teach some friends parallel computing. We wrote Fortran and C with OpenMP and MPI, and every step taught us something.
Recently I set up Reticulum, an encrypted networking stack, on a LilyGo radio in ten minutes with Claude Code. I understood what it was doing, and I could have done it myself given the time.
But I didn't watch each piece go together, and that is where a certain depth of understanding used to come from.
Working this way lets me do far more than I could before. It also leaves me with a sense that my own skills are slipping. I can still write Python, so I know that isn't the case. The sense is there anyway, and I think it deserves attention.
I had years of building by hand before these tools arrived. Many people starting now won't. I wonder what that is like for someone newer to coding or to science than I am.
What does due diligence look like?
Much of my research now runs through Claude Code and Codex. I ask them to transform data into different basis sets so that we can examine it in ways I wouldn't have thought to try. That approach has produced findings I'm excited about, including symmetries in satellite orbits that came out of the OrbitWatch work.
I did not personally inspect all of that data, and inspecting it would take a long time. Here the cost runs in both directions again. If I don't examine the data closely, I risk standing behind a result I haven't verified. If I spend weeks on one data set, every other project waits. I don't yet know where the ethical line sits.
Part of what I feel is a need to see in order to believe. Part of it is a need to see in order to understand. I'm learning to tell them apart.
For now, when I can't check every row, I check the method. I run blind tests against data with known answers. I run controls that contain no real signal. I audit whether one result is feeding another.
How much of the work is mine?
These models produce work that is complex and often beyond my own training. When one of them turns up something new, I go back and study until I understand it. It's humbling to be led to the starting line of a field and see how much of it I don't know.
OrbitWatch gave me two examples. I asked the models to recast the orbit data as matrices that GPUs could process in parallel. The idea was mine, and I would never have attempted it alone, because working out the mathematical structure would have taken me too long. I also asked them to treat the correlated error in the orbit data as one large antenna and to tune its gain. The analogy held, and the models found an ingenious way to recast the data. With it we can probe Earth's gravitational field using only publicly available satellite ephemerides.
I'm steering the ship, but how much of the wake behind me is my own?
I'm doing work I could never have done alone. When a model makes a discovery that is new, I can't yet say how much of it belongs to me.
Are the models interchangeable?
No, they aren't. As of September 2026 I run eleven agent seats in Claude Code, Codex and Grok Build: seven Claude, three GPT and one Grok. After enough hours with each of the major coding agents, you can see what's in the weights, the way you learn a colleague's habits.
I use them differently on purpose. The local model on my DGX Sparks takes the high-volume work, the daily write-ups, summaries and checks, where the job is clear and the cost is tokens. Codex takes the long, mechanical jobs: a server migration, a complex program, a debugging session, a security pass. Claude takes the jobs that need an idea. It will propose a design I didn't ask for, or a different route through the math, and more often than not it's a better one.
Someone took the time to put that in, and it's the part I value most.
It isn't finished, and I'd like to help with it.
Machine learning
A second set of eyes on the same sky.
"That model doesn't even understand physics."
Many people I've talked with in the military don't trust machine learning. The objection I hear most often is about weather models.
What I hear
"Sean, that weather model doesn't even understand physics."
What I say back
They're right, and I tell them so. We see the atmosphere as a physics problem, and physics has served us admirably in learning how it works. I spent years forecasting with physics-based models, and I train machine learning models for weather now.
I don't see the missing physics as a weakness. I see it as a strength.
There is no physics in a machine learning model. There are weights and nonlinear functions, and what the model has learned to do is find structure. A person can draw a convincing cloud without knowing the physics of clouds, and a model can do the same. I can train a super-resolution model on satellite images of clouds, and it will sharpen them into realistic fine structure without a single equation.
Physics and machine learning both look for structure in the same atmosphere, and they look in very different ways. A physics model divides the fluid into a grid and steps the equations forward with methods such as finite differences. A machine learning model has no use for any of that. It looks for whatever structure the data holds, and the structure it finds can be close to orthogonal to the structure physics finds. Because it isn't bound by our equations, it can see patterns that our physics doesn't account for. Why turn away a perspective because it wasn't formed through our own lens?
A second set of eyes, with a different way of thinking, can show us what our own approach leaves out. Two cases make the point.
When two models agree
If a deterministic model and a graph neural network both predict the same weather, or the same climate regime, they reached that answer by different routes. Their agreement strengthens our understanding of the science in a way that neither model can alone.
When physics runs out
Physics has limits to what it can explain today. If machine learning can show us something beyond those limits, it gives us a reason to extend the physics until it reaches there too. I would find that amazing.
That is power, and I don't think we should shy away from it while others learn to master it.
How we approach it
With rigor, in the open.
Should we fear it or embrace it?
Thoughtful people argue that AI development should slow down, and I take that argument seriously. The risks they describe are real. My own view is that other nations will keep building whether or not we do, so I would like the people with good intentions to keep pace and lead.
What I think will protect us most is transparency. That means being open about what AI and machine learning can do, and being responsible with that capability. It also means rigor: test carefully, share what we find, and say so when we get something wrong.
I try to approach the science without fear and with respect.
I've written a second essay about what the models I work with miss about the person reading their work. It's called What a language model doesn't see, and it describes the problem I'd most like to work on next.
I'll add to this page as I go. If you have thoughts on any of these, I'd like to hear them. sean@theinformed.org