The new bottlenecks
Honest accounting of an AI-augmented week shows "writing code" rarely tops the list anymore. Three things take its place.
Bottleneck 1 — Deciding what to build
A baseline classifier, a multi-modal pipeline, or a fine-tuning loop — any of them, in an afternoon. So which one?
Common failure: the agent is so capable you skip the planning step and ask it to "build something to predict X." Two days later you have working code that solves the wrong problem, evaluated on the wrong split.
Antidote. Before any agent session over ~30 minutes, write a one-page PLAN.md:
# What we are trying to learn
<one sentence — a question, not a feature>
# Smallest experiment that answers it
<the minimum viable run; data, model, metric, threshold>
# What "done" looks like
<a number, a chart, or a "kill the line of work" decision>
The agent reads this. So does future-you. Skip it and you rent compute to generate noise.
Bottleneck 2 — Verifying results are real
ML has always had a verification problem: train/test leakage, lookahead bias, metric scripts that silently divide by zero. Agents amplify it:
- They generate plausible-looking metrics quickly.
- They do not flag suspicious results unless told what suspicious looks like.
- They implement the metric the way you described it, even when the description is wrong.
A result that beats your baseline by 10 points is a bug until proven otherwise.
Bottleneck 3 — Distribution
The bottleneck nobody mentions: code nobody runs has zero impact. The agent will not write your blog post, present at a meetup, package your work for HuggingFace, or answer GitHub issues from confused users.
Three concrete distribution surfaces for ML work:
| Surface | What it requires | Frequency |
|---|---|---|
| Repo README + model card | Honest numbers, runnable example, license | Every release |
| Blog/notes (your site or HF) | Story, motivation, one chart | Every milestone |
| Talk / demo / tweet | A 30-second hook | Per major result |
The agent can draft any of these. You still have to ship them.
What leverage looks like
A good week with AI is not "I wrote 5,000 lines." It is:
- Tried six experiments instead of one.
- Caught two silent leakage bugs because verification was automated.
- Shipped the one that worked, with a real model card and a runnable benchmark, by Friday.
The five dead experiments are not waste. They are the system working.