Live · training on this server
Fly brain
A small model that stays in training. Each pass stores a real passage from headlines, Hacker News, the latest Arc block, or a desk guide, then refits a weight vector for that subject. Ask it, or teach it a subject from Wikipedia and the wire. Answers quote only what is on the shelf.
Training on this server21 lessons19 weight updatesLast pass 11s ago
- Lessons on the shelf
- 21
- Subjects
- 6
- Updates
- 19
Subjects
- desks · 5
- hacker news · 5
- headlines · 5
- arc · 2
- fruit fly brain · 2
- hi · 2
Just learned
- /depthdesks
- HIhi
- /markets/bitcoindesks
About this desk
- What you use it for
- Leave it running. The model keeps a training pass on this server. Ask a question when you want the passages it has stored. Teach a subject when you want Wikipedia and matching headlines added, then ask again.
- What the numbers are
- Each lesson is a headline, a Hacker News story, the latest Arc block and its transfer fee, a desk guide, or a page you asked it to learn. Training refits one weight vector per subject from those lessons. The match percent is how much the question overlaps that passage.
- What it leaves out
- An answer is only those stored passages. If nothing overlaps, it says so and does not invent a paragraph. A source that misses is named and is not stored as an empty lesson. Teaching does not spend ARCOP.
Every desk is introduced the same way on the desk list.
How to read this
- What training does
- A pass reads one new passage and refits the subject centroids from every lesson still on the shelf. The loop runs about every 45 seconds while this server is up. Train one pass runs that step immediately.
- What an answer is
- The question is hashed into the same bag of words as the lessons. The nearest subject is the centroid with the highest overlap. The passages shown are the stored lessons with the highest overlap. A miss stays a miss.
- What it will not do
- It does not invent a paragraph from outside the shelf, and it does not spend ARCOP. Teaching a subject fetches Wikipedia and matching headlines. If those come back empty, the weights stay as they were.