It had seemed quieter than usual around here.
Somehow, the flow of emails being resurfaced within hours to my * immediate *attention had grown thin.
Background: One of the things I have long done has been to have me, or for someone I delegate, #handcarry pieces of administrative paperwork on which I wish urgent action directly on up to the bureaucratic office concerned. Look the recipient who has to handle the piece of paper in the eye. (Or, more likely, and often more fruitfully, their schedule-controlling admin.)
I picked this up long ago from others, back when I worked for the US Treasury. It has served me well.
So I set up a similar system for my ‘bots.
And that is the context for this interaction:
It had made a typo: it had registered the hashtag #handcarry as #hancarry.
Then, rather than surfacing #handcarry emails immediately back to my attention, it had considered their existence, and decided to construct a new directory—/Users/delong/.openclaw/workspace/filing-agent/processed/hancarry/. It then began to diligently file the #handcarry emails there in …/hancarry/. Once the mistake had been made, it stuck. Every email tagged #handcarry got carefully and diligently filed away in .../hancarry/. For a week. Until I noticed that my .../handcarry/ directory—the one I actually look at—was empty when it definitely should not have been, and went looking.
Yes: natural-language interfaces are wonderful ways to interface with information-technology systems.
Much of the time.
Now you might justifiably classify this as user error.
Consider what #handcarry is for. It is the one tag in my system whose entire meaning is: do not file this. Do not put it in the queue. Do not process it in the ordinary way. Pick it up, walk it down the hall, and put it in front of the human being who will look at it: me. It is the escape hatch from the bureaucracy. But in form it is a #hashtag for filing in a directory exactly like everything else that gets looked at by the Hashtag Filing Protocol. I am thus overloading “#”. And overloading is bad.
Point.
But not game, or set. The promise of the Silicon Valley TechBros this day is that the machines will adapt to work smoothly with us, the way we already work. MAMLMs are supposed to be able to deal with linguistic ambiguity. They are supposed to be trainable to pay attention to and parse out the right bits of linguistic context. And we are told over and over again that that Attention Is All You Need.
But my ‘bot took my escape hatch from the ordinary procedures of bureaucracy, and it turned it into a new filing cabinet.
I want to stress how well it did this!
There was no refusal.
There was no idleness.
There was no sulking.
