Near the end of the night, a development engineer asked me whether he should buy a Ryzen AI Max+ box to run models locally. One of the little 128GB mini PCs. He'd been eyeing one for a while.
I told him it depends what he's optimizing for. "Run models locally" is a few different purchases wearing the same three words, and the marketing for them says the same thing.
One question is whether a large model will load at all. That's what the Halo chips are for. 128GB of unified memory, up to 96 of it addressable as VRAM, which means a 70B or a 120B that will not fit on any single consumer card (think GeForce RTX 3090 family) will load on something that sits on a desk. What that desk box costs depends entirely on whose badge is on it. Framework's 128GB desktop tops out around two grand. HP's Z2 Mini, same chip and same memory, runs about $3,700. If your problem is that the model won't load, one of those is your machine and the conversation is over.
The other question is how many tokens per second you get on a model that already fits. Different problem, different answer. That 128GB rides on a 256-bit LPDDR5X bus. Call it 256 GB/s on paper, closer to 215 once you actually measure it. A 3090 does 936. Put two of them in a box and you have 48GB of VRAM moving data about four times faster per card. On anything that fits in 48GB, the ugly box wins and it isn't close.
We got onto the DGX Spark too. Same shape of answer with a different tax. Also 128GB unified, slightly more bandwidth at 273, priced up with the expensive end of the Halo boxes, and CUDA instead of ROCm. Whether that last part is worth the difference depends on what you plan to run and how much of your weekend you want to spend making it run.
So figure out which purchase you're making before you make any of them. Capacity and throughput are not the same buy, and getting it wrong is the kind of "expensive" you don't find out about until the thing is sitting on your desk.
I had not walked in expecting to be the person answering that.
The room
I got to Gulp in Playa Vista early and talked to Gavin Johnson, who runs the meetup for Pinecone (LinkedIn). The event page had already laid out the trajectory. Seventy people in May, eighty in July, and "it'd be pretty cool to break 100 in August." No drinks, no food, nothing on the flyer that resembled a sponsored happy hour, which is usually what fills a room in LA on a weeknight.
Well over a hundred showed up. My count put it near 120.
I don't have a theory about that yet. Something is pulling people out of their houses on a Tuesday for a subject most of them can't monetize.
The build loop
The engineer asked what I actually run, so I walked him through it.
Hermes is the orchestrator. It sits above the agents, hands work between them, and watches the argument.
Claude Code drafts the spec. Codex tears it apart. They go back and forth until they converge, with Hermes doing the handoff each round, so neither one of them gets to decide when it's finished.
The cap is 60 iterations. If they hit it, Hermes tells both of them to stop and summarize what they're actually arguing about, and that summary comes to me. Then I break the tie.
It almost never gets there. A normal run is somewhere between 15 and 30 rounds, closer to 30 more often than not. That's a lot of tokens. I'm on Max plans for both, so I quit counting a while ago.
I read whatever survived the argument. Then they swap roles for the build, which Hermes also manages, so the one that wrote the spec isn't the one building against it. Claude Code watches for drift while the work happens.
The spec gets argued over by something with no investment in it before a line of code exists, and the review at the end isn't being done by the author. I'm in it at the review and at the deadlock, which are the two steps people try to automate away first.
He hadn't seen anyone run agents against each other that way. He was taking notes and looking things up on his phone while we talked.
An hour earlier
The best conversation I had that night was with a guy who'd run e-commerce automation at Alibaba, and it was about dropshipping.
I looked into dropshipping years ago and put it down. I understood the model fine. What lost me was where the real work lived. Advertising, media buying, split campaigns, tuning the thing after it's live. None of that is my background, so I filed the entire category under "a lot of marketing work" and stopped thinking about it.
So he walked me through how he sees it.
High level, not deep in the mechanics. It still held together better than the version I'd been carrying around in my head. There was real structure in there, and the part I'd dismissed as marketing turned out to be where the actual system was. I spent that conversation asking beginner questions about ad spend, at an AI meetup.
Nothing got resolved that night.
Friday
He messaged me today.
Really liked your idea and work. Love to get online again this Friday to show you things I build and hear from you too.
I have questions about ad spend.