Brier scores, language rants, and data-intensive applications
We learn about how we can compare forecasting accuracies, and a little about protocol design for avoiding forward compatibility.
I hope your week is going well!
New articles
An Opinion is not a Take
Very brief language rant. Don't call subjective writing a "take" – use many of the other more descriptive words we have.
Full article (0–1 minute read): An Opinion is not a Take
Brier Score
The most standard measure we have of forecasting accuracy is the Brier score, which effectively goes from 0.5 (very bad) to 0 (perfect accuracy) with people off the street sitting around 0.4–0.3 depending on question difficulty. This article lists a few more reference points in that space, and additional considerations.
Full article (2–8 minute read): Brier Score
Flashcard of the week
I see a lot of people gushing over Designing Data-Intensive Applications and I just don't get it. It's a good book, but it's not that good. Maybe I'm missing something. Anyway, one of the things I took away from it when I read it ages ago was the answer to
What guarantees about forward and backward compatibility do you get from upgrading servers before clients?
The neat thing is not the answer itself, which is
Doing it in that order guarantees you only need backward compatibility, not forward compatibility.
The neat thing is the simple reason why: clients initiate interactions. The client cannot know whether to use the old or the new protocol, because it's starting from a vacuum. The server, which sees the client's initiation, can decide based on it whether to switch to a compatibility mode or keep talking the new version of the protocol.
This also implies in which cases it is okay to upgrade the client first without needing forward compatibility: when the client, as part of its handshake, can figure out whether to use its backwards compatibility or not.
Premium newsletter
The third premium newsletter went out this weekend. It had three links, two book recommendations (covering investing and polar expeditions), a link to a future article that's not yet published, and the ACX 2025 forecast rationales.
If any of this sounds interesting, you should upgrade to a premium subscription! If you upgrade now, you will pay only $2/month – as interest increases and I learn to hit my stride with publishing these, the price will be set higher for future subscribers. As always, you can cancel at any time.
To upgrade, click the subscription link at the top of this newsletter and fill in your email again.
Your opinions
I cannot improve without feedback. Reply to this email to share your thoughts on any of the topics above, or anything else!