Two things I switched off this week.
One was a clever idea that couldn't do the job. The other was a safety feature that had been slowing down the person it was meant to protect.
Not every update is something new. This week two of my own changes went back off.
The phone answering on its own
Since late August, Neeos could decide per message whether to answer from a small AI model that runs on your phone, or to send the question to the server. Small talk and questions about notes it had already found went to the phone. Writes, web lookups, attachments, and long requests went to the server. A tiny symbol showed which one had answered. I turned it on by default on August 27th.
On September 4th I turned it off. The reason is simple: the phone's model can only reach what it's handed in the moment, and it couldn't reach most of what a conversation is actually about. It answered quickly and said very little. A fast, empty answer is worse than a slower, useful one. The code stays, and the switch is there if it earns its way back. I'm also being careful about a related claim: Neeos does not “work offline”. The phone model has real limits, and I won't pretend otherwise.
The brake that slowed me down
The spending brake I described at the end of August worked, including on me. For about five days it was quietly turning down my own account, which is the one I use to test everything. I'd been judging quality on a deliberately weakened version of the product without knowing. My account is now measured and never turned down, so what I test is what I built.
Both of these are the same lesson. A feature that sounds good and passes its tests can still be wrong in practice, and the only way to know is to use it.
Written October 21, 2026 from the build log, and dated to when this happened.