The job you couldn't do.

1. The wrong question

Every conversation about AI right now starts in the same place: which jobs is it going to take? Coders, designers, lawyers, radiologists, copywriters. The list grows every week. The mood is anxious. The framing is loss.

I want to flip the question, because I think the inverse is actually more interesting and more honest:

What job could you not do before that you can do now?

For me, the answer is astrophysicist.

2. The job I was never going to have

I'm not a physicist. I've never been a physicist. I have no PhD, no postdoc, no committee, no grant. I'm a software engineer and entrepreneur. I've been obsessed with astronomy and cosmology my whole life, but I never had what is normally required to do real research in either field. I had questions. I had ideas. I had hunches about cosmology that I'd been turning over for years. I had nowhere to put any of them.

The standard path to becoming a research scientist is roughly: undergrad in physics, GRE, six years of grad school, postdoc, second postdoc, junior faculty, tenure track, your own lab. That is fifteen years of training before you're trusted to ask your own question. I was not going to do that. I had a career and a life and ideas I wanted to pursue this decade, not next.

So for years I treated astrophysics the way most people treat the thing they secretly wish they did. As a hobby. A pile of arXiv tabs. A folder of notes I was never going to act on.

3. What changed

Three things showed up in the same eighteen months:

• Public survey data got serious. DESI DR1 alone is 17.65 million spectra. Gaia, SDSS, ZTF, JWST archives. Petabytes of real observational data, free to download, no permission required.
• GPU rental got cheap. H200s for a few dollars an hour. You can run inference over an entire sky survey for the cost of a nice dinner.
• AI coding agents got good. Not toy good. Real good. Good enough to write the FITS parser, the healpix indexing, the autoencoder, the cross-match against SIMBAD, the Parquet writer, the resume logic, the dashboards. All of it. Solo.

The thing that used to take a postdoc and a grad student six months now takes a weekend if you know what you're actually trying to do. And the “knowing what you're trying to do” part is the part the credentials were always supposed to confer. Turns out you can borrow that from a model too, if you're willing to read the papers and ask sharp questions.

4. Becoming the thing you weren't

I started by typing my actual cosmology question into a chat window. Then I asked it what data would falsify it. Then I asked what models people had used on similar problems. Then I asked it to write the pipeline. Then I read every line. Then I rewrote the parts that were wrong. Then I ran it on real data.

A few months in, I had a 195,829-object anomaly catalog scored across DESI DR1. Then a galaxy chirality classifier on 8.47 million galaxies. Then MCMC posterior samples on a cosmology model. Real outputs. Reviewable methodology. The kind of work that, two years ago, I would have told you required a real lab.

I am still not a trained physicist. I will never claim to be one. But the work is real, the data is real, and the papers will be real. The credential isn't the work. The work is the work.

5. The inverse question, generalized

I think a lot of people are sitting on a version of this. A job they always wished they could have done but couldn't justify the fifteen-year detour for. A research question. A novel. A clinical specialty. An economics paper. A historical investigation. A hardware project. A music score. The thing they were curious about but watched go by because the on-ramp was too long.

The on-ramp got short. That's the part of the AI story almost no one is talking about. Not “your job is at risk.” That's real, but it's the small story. The big story is:

The set of jobs you are personally capable of doing just got dramatically larger.

You don't need to go back to school. You don't need a committee. You don't need a fifteen-year detour. You need a real question, the willingness to read carefully, and a model that will write the boring parts while you steer.

6. Hubify is the lab I wish I'd had

This is the part where I'm supposed to soft-pedal the product. I'm not going to. Hubify exists because I needed it to exist. It's the operating room for the version of this story you might want to live: a workspace where the AI agents handle the pipeline plumbing, the GPU dispatch, the cross-referencing, the literature review, the figure generation, while you stay focused on the part only you can do — the question, the taste, the call.

It is not built for professional scientists. It can be used by them, and it should be. But it is built for the people who weren't supposed to be scientists in the first place. The software engineer with cosmology questions. The biologist outside academia. The historian with a dataset nobody has touched. The high schooler with a real idea.

The platform isn't magic. You still have to do the thinking. But the part that used to gate the door — the credentials, the institutional access, the years of plumbing — that part isn't the door anymore.

7. The question, again

So next time you see another headline about which jobs AI is taking, sit with the inverse for a minute:

What job could I not do before that I can do now?

That is a more honest question, and a more empowering one, and the answer might surprise you. For me it was astrophysicist. I never thought I would publish real research. Now I will, and I have. Because of the tools, not in spite of them.

The jobs you always wished you could have done are, for the first time, available to you. Pick one.