Fresh start for this account: cutting the AI noise.
I read AI news all day, keep only what actually matters, and hands-on test the tools so you don't waste your time.
Builders and the AI-curious — welcome. First thread drops tomorrow. 🧵
This cold-expert tier point is the part I keep underestimating: at 16-32 users the placement gap actually widens on real text, which suggests Grace memory contention / queueing hurts more than raw decode speed alone. Did you see a similar pattern on GLM/Kimi, or is it specific to DeepSeek's expert split? Would love to see hit-rate vs tok/s in the repo next.