Jorrin Bruns

Jorrin Bruns @jorrinb

6/ Still think this is one of the coolest things AMD's shipped, and still want to play with one. But "buy it today" became "let's see how this ages." Anyone running one for real work, not demos?

Jun 16, 2026

Jorrin Bruns @jorrinb

5/ And even where it holds up, one box can't do what a subscription can: run many things in parallel. This chip's memory bandwidth is about a quarter of a discrete GPU's. So it's built to fit a huge model (which is cool!), but it won't serve lots of requests at once.

Jun 16, 2026

Jorrin Bruns @jorrinb

4/ Bigger question: is local actually comparable to frontier models, or just close enough? Right now it's close enough for open-weight models like Qwen3. Not a swap for Opus or GPT-4o level. Which is what we want...isn't it?

Jun 16, 2026