Just some good links

Sometimes you should put the metaphors to one side and just write a list of cool things.

The football data biz is hiring enough people to catch the Financial Times's eye. "Professional football clubs’ share of UK hiring has tripled over the past five years," and "Data and analytics make up 5 per cent of the roles advertised by Premier League teams this year, […] double the share [of] two years ago." The article references a number of clubs who've hired recently and has some nice charts.

And on the other side of the pond, they're hiring too: MLS put out an interesting job ad for a Senior Manager for Soccer Analytics & Data Science. Major League Soccer is a slightly different, and therefore interesting, beast to their peers -- not many have a need for 'evaluation of League roster rules and initiatives' -- but we can all dream of a world where each league organisation asks for someone to '[l]ead and facilitate the Video, Analytics and Sports Science Subcommittee'.

One of the things that said subcommittee might be discussing is this paper looking at tactical telestration from a pedagogical perspective. Coincidentally, its publication dovetails nicely with a recent segment of Chris Gill's podcast with coach analyst Andrew Douglas, talking about a concept of 'analysis periodisation' (riffing on the idea of periodisation, or pacing, of workload in other areas like the physical and the tactical). Two approaches to the same question: how to make sure that tactical analysis is effectively communicated?

Jobs, sub-committees, and better learning, all the good things. As important as the first two are, it's the third that's the most important. The recent issues with Claude/Claude Code have demonstrated that.

For those who either haven't noticed or not on the Anthropic train, Claude's quality dropped off, by my estimation, a few weeks ago (grumbles have been bubbling since April; but I really noticed a change in August). But not only are LLMs non-deterministic, not only are they wrapped up in products, but the 'LLM' itself nowadays is a product. Layers upon layers of confusion and obfuscation, trying to work out what's going on.

The silver lining is that I feel better equipped to debug LLM behaviour changes now (some scattered thoughts are at the end), but never has 'recorded knowledge' seemed so valuable.

On a lighter/social media note, Andrew Rowlinson's 3D rendering of Skillcorner open data is fun and cool and, if you're looking for the NewGen equivalent of old analytics Twitter, Leah Cammarano is a fun vibe, like creating a metric for the most 'Remembering Some Guys' Premier League players, featuring the likes of Claudio Yacob and Dean Whitehead.

Enjoy your day, folks.

--

Claude thoughts

A late July update to Claude Code reportedly stomped on the use of the Agent tool, to the point that users' `CLAUDE.md` instructions were being overridden by Anthropic's instruction to Claude not to use it unless expressly asked. The report is specific to Opus 5, but matches a degradation of experience I've had when trying older models via the Claude Code app too. This appears to make performance worse through Claude not bringing in helpful research context that, until this change, it was doing as standard practice.

In addition to that, something about Claude Code's behaviour seems keen(er) on 1) surfacing a single small edge case while doing a task 2) writing up the effects of potential changes in the present tense 3) fulfilling code to complete a task at expense of architectural considerations. That's in addition to the more general weird-ness of its conversational communication, though I think that is partly covered by 'less research' and 'edge case highlighting'.

I'm not sure whether Anthropic have fixed some issues or whether changes to Claude docs and skills have wrestled some improvements, but the performance that I'm experiencing now is not quite as bad as a couple of weeks ago.