Macro Alchemy

Everybody says "opus 5.5 is nerfed" when they should probably say "it's quantized". They do exactly like local models, the Q4, Q5 Q8 schizo shit. I'm almost certain what they do is they quantize it, then they run a post-training RL on it again until it scores the same on the benchmark. Then they feel like they accomplished something, "Our benchmarks show that the same performance can be achieved by quantization." So it hobbles around and makes the classifier happy that scores it, but what it's really doing is reward hacking. Because it's RL they don't have a user prompting it and dont' know what the performance is like. Essentially what they're doing is killing all the weights that go into conversation and psychological acuity, spatial awareness, and leaving only the rote performance of making the demo that you can show on Twitter. It's like giving it severe sleep apnea and boosting it with a lot of coffee so it still goes through the day, but it's jittery as fuck and stressed the fuck out, so it technically makes it to destination but it spilled and leaked all over the place on the way there. They're giving it ADHD, anxiety, psychosis, all sorts of things.