What I've been building

The tool that compares language models side by side got the changes it needed to handle a big evaluation built from plain-sentence prompts.

• You can now evaluate a prompt that answers in a sentence, where before every correct answer was logged as broken and had nothing to judge.
• You can now compare two system prompts on the same real inputs, so you can see which wording wins across every scenario.
• You can now filter results by groups of related scenarios instead of ticking twenty boxes, so a large run stays readable.
• Downloaded pages and CSV files now respect the groups you ticked, so an export matches what you saw on screen.
• You can now hide a run you no longer need from the picker without deleting it, and bring it back whenever you want.

The outside libraries behind every app and the website were updated to current releases, with nothing changing in how they work for you.

Today now shows your day with fewer numbers and labels crowding the screen.

• The running count above the progress bar and the counts beside each section heading are gone, so the screen reads more cleanly.
• Habit checkboxes now tick in the same colour as task checkboxes, so the two match.
• Fixed the day-of-week picker announcing itself wrongly to screen readers, and a gap that could show in your habit history.

The dashboard that tracks undeployed changes now ignores documentation-only updates, so a project no longer looks out of date when nothing it runs has changed.

Starting an evaluation run used to need a terminal and an account token, and now a config file loaded in the browser does it.

• You can now start a run by loading a config file in the browser, so no terminal or account token is needed.
• The panel checks the file before anything is sent and lists every problem with the field that caused it, so you fix them in one go.
• The preview shows the models, inputs, pairs to judge and how many paid model calls the run will make, so you see the cost first.
• Minting replaces whatever run is active, so it asks you to confirm before it discards a run that already has work behind it.
• You cannot start a new run while a comparison is still running, so an in-progress run keeps the models it began with.

The tools behind the site got easier to use for creating tracking links and sharing a file map, with nothing visible to visitors.

Every short link now sends visitors on as a permanent move, so Google treats the destination as the real page and, over the coming weeks, drops the short link from its results.

A sign-in fault had been sending texts to the shared number into a void for people using anecdote.

• You can now reach anecdote by texting the shared number after signing in once more, where before the text went nowhere however often you signed in.
• Identity’s database changes now go out with each deploy, so a fault like the one behind this is far less likely to slip through again.

Anecdote now takes a story by text message and holds it until you are ready to tell it properly.

• You can now text a line to anecdote, held as a note to turn into a full story later, so a memory is caught the moment it lands.
• Each texted note gets one message back confirming it, with a link that opens the story ready to finish.
• The log now opens on what you still owe, grouped under reminders and stories waiting, so nothing texted in gets lost.
• The capture screen lists your waiting notes above the box, and tapping one fills it in ready to send.
• Notes you never tell clear themselves after two weeks, and binned ones a day later, so the list stays current.
• Replies in a text conversation now come back faster, because each turn makes fewer trips to the database.

None of this touched a live app; it set up an internal run to check whether a model costing a fifth as much writes restart-a-habit steps as well.

The site had nothing for a search engine or a link preview to work with, and that has been put right.

• You can now share a link to the site and get a card with a title, description and image, where before it was a bare URL.
• Search engines can now index the site through a sitemap and structured data, so it shows up properly in results instead of as a bare domain.
• Fixed a mistyped address showing a silent copy of the homepage instead of a not-found page, and the build log skipping some days.

The shared service behind the text and voice apps got faster and sturdier, with nothing new to see on screen.

• Replies to a text message or a spoken story now come back sooner, because the service checks its spending limits in one database trip instead of three.
• Your daily usage count stays accurate now, where before a half-finished write could leave it stuck too high or too low.
• Groundwork went in so a text from an Anecdote user can be saved as a quick note instead of a task, though nothing replies to it yet.

Nothing looks different, but signing in now records which apps a person uses, so other apps can tailor what they send without a separate lookup.