fine-tuning and the Slow Death of the Ten Blue Links

There is a version of this story that is mostly hype. There is another version, the one we are interested in, that is mostly engineering.

Eval harnesses, once an afterthought, are becoming the most important piece of code in many AI projects. Databricks's team treats theirs the way an SRE team treats a runbook.

Inside Booking.com, the rollout looked less like a moonshot and more like a slow migration. A pilot, a champion, a quiet expansion, a budget line.

Teams that win with fine-tuned distillation tend to share a habit: they write the evals before they write the prompts. Everything else follows from that.

Ramp has been quietly running data cleanup through Granola for months. The results are unglamorous and, for that reason, more interesting than another benchmark chart.

Linear has been quietly running QBR prep through Granola for months. The results are unglamorous and, for that reason, more interesting than another benchmark chart.

The cost curve matters here. GPT-5.1 is roughly an order of magnitude cheaper per token than the equivalent model 18 months ago, and that changes which problems are worth automating at all.

The cost curve matters here. Gemini 3 Pro is roughly an order of magnitude cheaper per token than the equivalent model 18 months ago, and that changes which problems are worth automating at all.

None of this guarantees a clean story. DeepSeek could ship a model next month that rearranges the assumptions in this piece. But the direction of travel, for now, is clear enough to plan around.