When I first started running over a decade ago, I quickly built up to 2-3k. 5k+ was a little harder. I have flat feet, and it became a challenge to overcome the frequent foot pain I had. Then I invested in good shoes, and built up to the occasional 10k over a few years. Soon, I considered anything below 7k to be a “short run”.

I kept pushing into trail runs, 12-13k regular runs and finally I ran my first half marathon in 2018, about 3 years after I started running regularly. I ran 3 more after that, finishing in the 2:20-2:40 range. The first 3 HMs were extremely exhausting. The stretch after 18k was particularly brutal. Blisters on my feet, heart rate rising, my glutes complaining, I had to summon every bit of willpower to trot-walk-drag my way to the finish line.

An LLM has no intuition for when a contradiction means a human misunderstood something and when it means the spec is wrong and should change.

The Bug Bash

Last night I pointed Fable at my bugs list and asked it to review and fix everything. I mentioned some of them might already be fixed, so it had to review them before fixing.

It fixed four bugs, and I verified them. A good night’s work.

Getting value out of an AI is mostly context transfer: you hand it the constraints and context it can’t figure out on its own, add enough tokens and a few prayers, and value falls out. Or that’s the hope.

The problem is humans are not very good at sending context to AI. We’re impatient, so we send half-baked instructions and expect the model to fill in the gaps. Sometimes it does a good job; sometimes it totally misses the point.