i think this is the real story with local models. they don't need to be smarter, just cheap enough to run the boring repetitive stuff. fishing is the perfect test for that
I let DJev play Runescape at fishing for 30 min.
Every single action is decided by the 26B diffusion model itself on a single 5090. ZERO neurosymbolic setup.
+1.84M XP vs. Astra's +1.61M
~$0.32 of power + 5090 wear (roughly simulated -- assuming $0.45 kWh) vs. $13.62 in API.
42x cheaper.