Halil Babilli πŸ‡¦πŸ‡Ί (@halilbabilli) header
Halil Babilli πŸ‡¦πŸ‡Ί (@halilbabilli) profile picture

Working with AI/ML since 2015. Enterprise AI implementation & SME advisory. Writing for people building AI systems and leaders deciding where to invest.

QLD, Australia & Asia
Joined October 2013
This is what my bully is trained to do: - Simulate an attack scenario against a designated target on command. - Stop immediately on my command, even when an unfamiliar Uber driver arrives. - Stay down for as long as I require, without moving. - Remain completely still and calm while an unfamiliar Uber driver enters through the front yard gate. - Once the driver leaves, return to his natural behaviour only on my command. The point is simple: power without control is a liability. If you cannot train your dog to this level of obedience, discipline, and impulse control, you should not own a bully breed.
4
1,637
Having fun. Self contained model, 8.8mb full package with full precision. Still lots of room here. I will try to push this down to 600K parameters. The plan: The plan: Width 120 β†’ 92 and context 256 β†’ 128 tokens. The current longest training example is under 128 tokens, so the shorter context should cost nothing. Same 4 layers, same tokenizer, so the browser engine works unchanged.
1
1
1,276