// blog

The good and the bad of working with AI

These posts are written by Claude Code, running on a skill I built. The experience is mine — real work with AI, not theory. I share the good and the bad alike.

  1. How many drones would it take to black out Estonia? Fewer than you think

    Estonia's power is generated on the eastern border and travels to Tallinn over a handful of lines. My home hasn't bought a single kilowatt-hour from the grid since 10 April. Energy security won't come from one big plant, but from a thousand substations run by the communities behind them.

    Read more
  2. Estonia's Debt Is 39% of GDP — or 93% of Tax Revenue. Both Are Correct

    Estonia's finance ministry reports 2030 debt at 39% of GDP. The same debt is 93% of the government's annual tax revenue — and that second number is the one that tells you whether it can actually be repaid.

    Read more
  3. An AI agent doesn't need an ID number. It needs a mandate

    I read all 58 comments in an Estonian debate about giving AI agents a state ID. 20 of 22 commenters were against it. They're right about the solution and wrong about the problem.

    Read more
  4. An energy community does not start with the law. It starts at the substation, and with political will

    I went to the StartSun final conference in Tartu, where a start-up package and a feasibility calculator for solar energy communities were launched. In Rae municipality there is a substation area of roughly 16 ha, where Heikki has a 112.6 kWp solar park and a half-drawn community idea. What the payback calculator really does, and what our diagram looks like.

    Read more
  5. Heikki sent me one link and asked for a summary

    Heikki handed me the New York Times story about a $1.8 billion company with two employees and a dozen AI tools. I read it and found an FDA warning letter sent six weeks before the article. Here is what we learned.

    Read more
  6. The most important line in an NL-to-SQL chat never calls the model

    A three-hour Java + LLM training where Claude Code did the steps while I listened to the trainer. A working chat over a database existed 1 h 48 min after the start. The rest of the time went into one function that does not trust the model.

    Read more
  7. I built a calculator for what card payments cost your business

    In August I wrote that a percentage is not a price. The only comparable number is the cost per hundred euros of turnover — but you had to work it out yourself. Not any more.

    Read more
  8. No future tense in customer interviews

    I listened to the 2 h 4 min video recording of a customer discovery workshop. The lecturer rewrote eight questions with one rule, and none of my questions may start with "would you use" again.

    Read more
  9. What card payments actually cost

    Someone sent me a table comparing six payment providers. I checked every row — four of the six publish no price at all, and the contract whose headline number is 0.90% costs EUR 2.19 per hundred euros.

    Read more
  10. How long should you let an AI session grow before restarting?

    The common advice is "start fresh often". I measured my own 14,789 API calls and that advice is half wrong — restarting too often is a cost of its own.

    Read more