Workplaces have had reports nobody reads for years.
Seriously though, I think this question will come up in every field. You're onto something with courses and seminars, where more of the value comes from people's own expertise. Value will have to shift to where humans can provide something unique.
An AI recommendation can change someone's career.
For anyone building this kind of tech, the law is catching up. Connecticut already says using the software is no defense to a discrimination claim. Australia follows on 10 Dec, California and Colorado in Jan 2027.
Sources below.
Gave Fable 5.2 my health tracker app to speed up. It was off like a greyhound, almost like it enjoys having something to chase.
In tests my walk now loads in about half a second, down from over 5.
One snag, Claudeโs hidden browser spoiling a timing run, wasnโt Fableโs fault.
Gave Fable 5.2 my health tracker app to speed up. It was off like a greyhound, almost like it enjoys having something to chase.
In tests my walk now loads in about half a second, down from over 5.
One snag, Claudeโs hidden browser spoiling a timing run, wasnโt Fableโs fault.
After reading this I asked Claude to build a "dream" process. Definitely recommend this. The first run caught two of my own rules clashing.
One thing I added was a check for stale memories. Basically anything untouched for 8 weeks. The result was 5 more flagged items.
Learn from this. Just don't hand your agent's memory to a lab.
I've been doing a lot of this with plain .md files for months. One small index the agent always loads. Domain files it opens only when a task needs them. It's in my repo, so I can read it, diff it, fix it, and the same files work in Codex.
Even Claude Code's own memory loads just the first 200 lines, as an index.
Their tests back the rest:
- Better models keep less.
- Telling the model how to write memory hurt. Say what memory is for, then let it judge.
- Check notes regularly. They call it dreaming.
I added a /dream to Mini Harness this week: merge duplicates, delete what the code contradicts, keep the index accurate. Without it, every knowledge system degrades over time.
@_simonsmith A lot of my use cases are solvable now and building is so much easier. But there's still things I'd love from the next model like clearer, shorter responses.
Congrats Max! Two from someone with a lot of scheduled tasks
1. The Scheduled list takes over the sidebar and pushes Recents down to a few rows. Collapsing it hides the ones that have run or need my attention so showing just those would fix it for me
2. What do the yellow, blue, grey and orange dots on sessions mean? I couldn't find it anywhere. Related, a home screen with all recent activity (sessions, scheduled runs, artifact updates) would be great