The parts of AI I could no longer delegate

I was on the subway recently, watching a lesson about neural networks.

A familiar set of symbols came up on the screen.

Weights, gradients, derivatives, loss functions.

I had met versions of them before. I studied computer science at Waterloo, and I've spent most of my adult life building technology companies.

But the formulas felt different this time.

They weren't attached to an assignment, an interview question, or something I was learning because I thought I should know it.

They were attached to GPUs we pay for by the month, a model we were trying to train, and decisions that could cost us weeks.

The math had become real.

Building more while understanding less

AI has made me much more technically capable.

With coding agents I can go from an idea to a working system faster than I ever could before.

I can read unfamiliar codebases, test product ideas, and direct work across infrastructure, research and design.

Things that used to need a much larger team can now be attempted by a few people working closely with AI.

But there's a strange contradiction inside that progress.

I can build more than ever while personally touching less of the machinery.

An agent writes the code. Another agent reviews it. A third explains the concept underneath it.

When something fails, they can produce five plausible diagnoses in a few minutes.

That's extraordinarily useful.

It's also dangerous.

A good explanation can create the feeling of understanding before the understanding is there.

A technically fluent postmortem can sound convincing even when its conclusion is wrong.

If you keep moving fast enough, it becomes easy to confuse access to intelligence with having judgment yourself.

For a while that distinction was manageable.

We were mostly building products on top of models other people had trained.

I understood the user, the product and the outcome we wanted.

AI and the people around me could go deeper into the implementation than I could, and that was fine.

Then Vinci started taking model training seriously.

When the training didn't work

Our most ambitious training run so far did not produce what we wanted.

This wasn't our first attempt at training anything. We'd trained smaller models before, learned from them, and shipped some of that work.

This one was supposed to be different.

We wanted a model with better judgment. One that could understand what a user actually intended, recognize when work was incomplete, gather the right evidence, and avoid claiming success when it hadn't succeeded.

The first attempts fell short.

Afterwards there was no shortage of explanations.

Maybe the data mixture was wrong.

Maybe the learning rate was wrong.

Maybe the training stages were sequenced badly.

Maybe the evaluations were misleading.

Maybe the thesis was sound and the implementation was weak.

Maybe the implementation was fine and the thesis was wrong.

Several of those could be true at the same time. Most of them sounded reasonable.

That was the problem.

I could direct the discussion. I could ask questions and compare results. I could follow the broad tradeoffs.

But I couldn't yet independently adjudicate every important technical claim in the room.

I was a strong model-product founder and an emerging evaluation owner.

I wasn't yet an independent model-training supervisor.

It's uncomfortable to write that sentence while running an AI lab.

It's also true.

A founder can't delegate the final judgment

I don't think a founder needs to be the best engineer or researcher in the company.

If I build the kind of institution I want Vinci to become, there will be people who understand individual parts of model training far more deeply than I ever will.

That's how serious organizations work.

But delegation only works when someone still owns the objective, the evidence and the tradeoffs.

Someone has to decide what the model is actually being trained to do.

Someone has to decide whether an improvement is real or a benchmark artifact.

Someone has to ask whether another run is justified, whether the data supports the conclusion, and whether the company is learning or just producing more sophisticated stories about its own failures.

AI can't own that.

Neither can an outside specialist.

The founder doesn't need to personally run every training job.

But if model training is central to the company, the founder has to be capable of challenging the people and the systems that do.

That's why I'm picking up machine learning again.

This is not a return to school

I'm not doing this to collect credentials, finish every textbook, or become a conventional academic researcher.

I also don't have six months to disappear and study first.

The business has to move now.

The training work can't wait for me to feel ready.

Our experiments, data work and evaluations have to proceed while I learn alongside them.

That makes it different from school.

One morning I might be learning why gradients vanish through a network.

That afternoon I may have to decide whether an experimental result justifies another run.

I can study post-training on the train and apply it to a real pipeline, with real compute and real consequences, the same day.

The feedback loop is much shorter.

It's also less forgiving.

My goal isn't to memorize formulas.

It's to understand what they represent well enough to know when something is behaving incorrectly.

I want to understand training dynamics, data quality, supervised fine-tuning, preference optimization, reinforcement learning and evaluation design well enough to separate evidence from technical theatre.

I want to hear a convincing explanation and still ask the question it's avoiding.

Learning in the age of AI

There has probably never been a better time to learn this.

I have courses, books, papers, working implementations and infinitely patient tutors.

If an explanation doesn't land, I ask for another one.

I can ask for a visual intuition, a numerical example, or an explanation tied directly to a problem we're actually seeing inside the company.

That would have felt miraculous when I was younger.

But AI also introduces a new failure mode.

It can remove the struggle that makes the learning yours.

If an agent does the exercise, summarizes the chapter and tells me what to conclude, I move faster without becoming more capable.

I end up renting the appearance of understanding from a machine.

So I'm trying to use AI as a tutor without letting it become a substitute.

It can explain. It can challenge me. It can debug something after I've attempted it myself.

It can't quietly do the hard thinking and then let me pretend the result transferred into my head.

This is harder than it sounds.

The same tools that make learning easier make avoiding learning easier.

The company changed, so the job has to change

A year ago my most important technical responsibility was understanding how AI products should work.

Now I need to understand how the models underneath those products are trained, measured and improved.

Later, if Vinci becomes what I think it could become, the responsibility changes again.

The experiments get larger. The organization gets more complex.

The cost of a plausible but wrong conclusion gets much higher.

My job isn't to preserve the version of myself who was qualified to run yesterday's company.

It's to become qualified for the company we're trying to build next.

Right now that means neural networks, transformers, post-training and evaluation.

It means sitting with formulas until they stop looking like symbols and start describing systems I can reason about.

It means admitting what I don't know without treating that admission as a reason to slow the company down.

What I want to change

I don't know if I'll ever call myself an ML researcher.

That isn't the outcome I'm after.

The outcome I want is this.

When a training run fails and five explanations arrive, I can tell which ones the evidence actually supports.

When a result looks unusually good, I can recognize what might be hiding underneath it.

When we commit serious compute to an experiment, I know exactly what belief we're testing and what result would prove us wrong.

I want to be able to supervise increasingly ambitious training programs without handing my judgment to either humans or AI agents.

That takes longer than finishing a few courses.

But I've started.

For now it means sitting on the train with neural network formulas on a screen, occasionally feeling like a student again.

The difference is that there's no exam this time.

There's only the company we become because I learned it, or the company we don't because I didn't.