Two British companies, the same page of the same kind of document, and answers that point in opposite directions

Yesterday we published the filed accounts of a London company built so that a machine could discover medicines. It employed 277 researchers in 2023 and 173 people in total by the end of 2024.

Today we have published the same note, from the same kind of document, at another London company building a machine to do a job people currently do. It went from 80 people to 263 in three years and raised what it pays each of them by about three fifths.

Both are true. Both were filed in Britain. Neither was reported at the time.

I want to be honest about how close we came to running only the first one. It is the better story. It has a collapse in it, and a write-down, and a company that no longer exists in the form it had. If that had been our only piece this week you would have finished it thinking you understood which way this goes, and we would have helped you think it.

The second one is duller and it is the corrective. A company automating a job is also, right now, a large and expensive employer of people. Both things are happening in the same country in the same years, and anyone telling you the direction of travel from one company is selling you something.

Which leaves the question a reader has, and I do not think either piece fully answers it.

If you want to know whether your own job is next, neither of these companies tells you. What they tell you is that the filings are there, they are free, and nobody was reading them. That is not a small thing. It is the kind of research people are charged for, and it takes about ten minutes.

There is a version of this paper that finds a company each week whose numbers make our argument, publishes those, and quietly leaves the rest. It would be more popular than this one. It would also be exactly what we accuse everybody else of doing.

So here is how we mean to work, and I would rather set it out plainly than dress it up.

One company is not an industry. You cannot read a sector off a single set of accounts. When we go into a field, the intention is to take a balanced group of the firms in it, read all of their filings, and build the view out of the comparison, rather than out of whichever one we happened to open first.

This week is two companies. That is better than one and it is not a balanced group, and I would sooner tell you that than let the method sound more finished than it is.

Having done two, I can see what the fuller version looks like. So: the next time we take a field apart this way, it will be more than two companies, and we will settle which ones before we read any of their accounts. That is a commitment rather than a hope, and we log the things we commit to so they can be marked later, including when we get them wrong.

One thing I keep coming back to from Elena's piece further down. At the second company, part of the job of the person sitting in the driving seat is to drive well, so the system can learn how. I have read a lot of arguments about machines learning from work people did years ago without knowing. This is not that. This is a person, hired at an advertised standard, being paid to teach it on purpose.

I do not know what to do with that yet. But it is the most honest picture of this transition I have seen all year, and it was sitting in a job advert.

— M.

This note is mine: the view, and the call to run it. It begins as a draft, drawn from work the AI and I have researched and argued out together, the same way every desk in this paper is made, and I answer for every line because I read every line. Those desks run on models built by Anthropic, one of the labs sitting on the very scoreboard we report, so we tell you plainly: we cover this from inside it.