Spreadsheets are all you need

Ishan Anand put GPT-2 in a spreadsheet. Not a spreadsheet that calls an API, the actual model, the embeddings and the attention and the layer norms, in ordinary Excel formulas, so you can click on a cell and see the number that went into the next token. It’s called Spreadsheets are all you need, it was on the front page of Hacker News a couple of weeks ago, and it’s a great way to teach how a transformer works, because there’s nowhere for anything to hide.

It also made me laugh, because I went back through this blog and nearly every time “we” did something organized, it was a spreadsheet. The credential inventory in 2013. Every machine we owned after Shellshock. The agents with admin rights after SolarWinds, and the vendor list after Log4Shell, which by now has columns for which bank is behind each vendor and whether it uses AI. The GDPR data map was a spreadsheet with a grand name.

So, the tools I actually run an engineering team with, as of this month:

• A spreadsheet for the headcount plan. One row per person or open role, with team, level, start date, cost, and a column for “what happens if we don’t fill this”. Finance has its own version. Mine is the one I think in.
• The vendor list, the on-call history (exported monthly from the paging tool, because the paging tool’s own reports never answer the question I have), and a budget sheet that’s mostly a copy of the cloud bill grouped by team.
• One shared doc per week, where each team lead writes three or four lines about what happened and what’s stuck. I read it on Friday afternoons. It replaced a status meeting.
• Obsidian for my own notes, including one note per person for one-on-ones.
• The ticket tracker, for actual work, which the engineers own and I mostly stay out of.

The list of things I’ve paid for and stopped using is about as long. A goals tool, which we used for two quarters until everyone was updating it the night before the review and I was reading it the morning of. A one-on-one tool that wanted me to fill in a template before every meeting. A “developer productivity” dashboard that measured PR cycle time very precisely and told me nothing I didn’t already know from the weekly doc. Each was a schema somebody else designed for a question somebody else had.

That’s the thing a spreadsheet does that a purpose-built tool doesn’t. When I don’t know what the right columns are yet, which is most of the time, I can add one, and when I was wrong, delete it. The vendor list had four columns when we made it in December 2021, and it’s grown to fourteen because things kept happening that we wanted to track. No vendor would have predicted the bank column. I added it the week our bank failed.

The case against is real, and I’ve lived it. Spreadsheets have no access control worth the name, and they rot quietly when the person who cares about them leaves. The headcount sheet has broken formulas in it right now, I’d bet on it. My rule, such as it is, is to stay in the spreadsheet until it hurts, and move to a real tool when you can say exactly which pain it fixes. We moved hiring into an applicant tracking system years ago, for good reasons (compliance, and candidates deserve better than a row). We moved the machine list into config management once it stopped being a list and started being a fleet. Everything else is still rows and columns.