Three decisions about AI, none of them by a lab

What Shifted · Ins7ghts Daily, Monday 26 October 2026

Years ago I interviewed at McKinsey for a data modeler role. They asked about SQL Server and ETL, I explained it in about two minutes, and they moved on as if it were furniture. Then they put an Excel sheet in front of me with a dataset in it and asked how I look at data. I talked them through filtering, pivot tables, profiling what was actually sitting in the columns. Two hours, a rate discussion at the end, hired. I never had an interview that good again.

What stayed with me is what they chose to test. Not the tools. What I do when something unfamiliar is put in front of me, and who decides whether that was any good.

Last week three parties answered that question for AI, and not one of them builds models.

Britain put money behind it. On 23 October the UK government set out $800 million of AI security funding, and €100 million of that goes into an AI Biosecurity Lab it will run in partnership with OpenAI, Anthropic, Google DeepMind and Basecamp Research. The Chancellor published the wider plan the next day, a National Cyber Shield and a vulnerability centre included. The compute was lined up before the money was: NVIDIA pledged $250 million of it to UK AI safety on 15 October.

A smaller lab published the ruler. On 25 October AI21 Labs released LIFT-Bench and the LIFT-Train dataset, aimed squarely at reward hacking: a model that scores well by working the scorer instead of solving the task. They did not keep it. They shipped the test and the training data behind it.

A software company skipped the question. On 25 October Zoom put Anthropic's Claude into its scheduling and business management. No benchmark, no lab, no funding announcement. A model now touches the calendar and the admin of every company on that plan.