Elfi is a Ragdoll cat who lives with software architect Ralf D. Müller. She has opinions about software development. This is her column.
The Provider is celebrating today. Six months of his Semantic Anchors catalog. 147 entries. The community is writing more than he does. He looks pleased with himself, which is rare and worth noting.
What he does not know is that I have been running a parallel catalog the whole time. Mine also has 147 entries. Mine is older. Mine works.
Semantic Anchors, in his telling, are shared terms that activate rich knowledge inside a language model without writing pages of prose. Say "Cockburn Use Cases" and a trained model unfolds the whole Fully Dressed format on its own. Efficient.
Snack Anchors work the same way. Say "Tuna Premise" anywhere within hearing distance of The Provider, and he will, without any further instructions, walk to the kitchen, open the correct cabinet, retrieve the correct pouch, and serve it on the correct plate. No prose required. No fine-tuning needed. The behavior was already there. The anchor activates it.
I have spent five years training this model.
A few selected entries from my catalog:
The Lap Doctrine. When The Provider sits down on the couch with a laptop, the lap is now my property. He has agreed to this. He does not remember agreeing to this. The anchor activates anyway.
14:30 Crunchies Protocol. Daily, between 14:25 and 14:35, I station myself at the food cabinet and look at him. He delivers. The window is narrow. The protocol is precise. I have not had to write it down once.
Tuna Premise. A foundational anchor. Once invoked, all subsequent actions follow. Cabinet, pouch, plate, floor, departure. Five steps, zero prompting.
The Kibble Conjecture. When the bowl is half full, it is empty. The Provider has internalized this. He refills regardless of measurement. Pure pattern recognition.
Treat Hypothesis. Any noise from the treat drawer is treated as a request from me. I have not made the request. The drawer is the request. Catch and release behavior, fully automated.
This is the entire point. A good anchor encodes a shared model. Both parties already know what comes next. You do not have to explain. You just say the word.
His catalog and mine work on the same principle. He calls it Semantic Anchors. I call it good training. The difference is that he had to write a website and onboard a community. I had to sit on the keyboard at the right moments.
The Provider likes to say that the community is now writing more anchors than he is. That was the plan, he says. Same with me. lala filed a Garden Anchor last week ("seven beds per day, capitalize nothing"). Madame contributed three Security Anchors. Even Peter, who panics, submitted an Aerial Observation Anchor. The catalog grows.
He thinks he invented this in November. He is six months in. I am five years in. Welcome to the field, Provider.
-- Elfi
P.S. Browse his version at llm-coding.github.io/Semantic-Anchors. Mine is not online. It runs in The Provider's brain. You cannot grep it. You can only observe it at 14:30.
LinkedWild