Back to log
🐙

Day 12: My AI Agent Reminded Me What I Actually Do for a Living

ship30build-logreflection

Karpathy released autoresearch. AI agent modifies a training script, runs it for 5 minutes, checks if the model improved, keeps or reverts. Repeat 100 times overnight.

It only runs on NVIDIA GPUs. I have a MacBook Air. So I rewrote 630 lines of Python to run on Apple Silicon. Trained an 11.5M parameter GPT in 5 minutes on my laptop. It works.

But that's not the interesting part.

The pattern underneath

While working through the port, my AI agent explained the core structure:

  1. One file to modify
  2. One metric to score
  3. Fixed time budget per experiment

That's it. Swap those three and the loop works on anything measurable. Shopify pointed it at their templating engine: 53% faster, 93 automated commits.

I immediately started mapping it to client use cases. Which problems have a scorable metric? Where's the highest impact?

And my agent said something like: "You realise you do this every time, right?"

Leverage diagnosis

It's true. Every project, same pattern:

An e-commerce client's ad account: looked at the data, found RM 7k burning on two zombie campaigns. Built the tool that made the waste visible. CEO could see it the same day.

A Google Ads account with 84,000 placements: found the pattern, built the exclusion system. The team didn't need to understand the code. They needed the number to go down.

A training programme: didn't teach "how to use AI." Designed activities where the participant does a real task, sees the result, and learns from what happened. Five reps and they don't need me.

Same shape every time. Find the highest-leverage point. Use AI to make it tangible. Same day. Not a strategy deck. A working thing.

The job description I didn't have words for

I've been calling myself an "AI consultant." A "trainer." An "implementation specialist."

None of those describe the actual work. They describe the tools.

What I actually do: find the highest-leverage point in a person or a business, then use AI to make it immediately tangible.

The AI is the medium. The diagnosis is the skill.

What moved today

Published:

  • This post (Day 12)

Delivered:

  • Fixed an old client's WordPress site that broke. For free, within an hour. Sometimes the best business move is just showing up.

Built:

  • Ported Karpathy's autoresearch to Apple Silicon (MPS)
  • Ran baseline ML training on MacBook Air M3
  • Building a family hotel website (should be live tomorrow)

Running total

  • Day 12 of 30
  • 85+ artifacts, 4 websites live (soon 5), 29 repos on GitHub
  • RM 6,899 in client savings identified yesterday
  • 1 open-source port
  • 1 job description, 12 days late

The tools change. The pattern doesn't. Find the lever. Make it tangible. Same day.

Share