AI Weekly: The Model Is the Swappable Part

The Model Is the Swappable Part

Mike Cannon-Brookes gets two kinds of questions from Atlassian's enterprise customers. One kind he can satisfy. "there is certainly some customer pushback in some places to quote unquote Chinese models. There is not customer pushback to open models." So Atlassian runs open models, and draws one line: "And so we don't use any Chinese models in production."

He does not expect the line to hold everywhere. "But for an enterprise to say, I'm not gonna use a cheaper alternative that'll give me better results, when I know what's in it, I think it's gonna be a very hard, long term choice". Plenty of companies already cross it quietly. The model a startup calls its own, he says, is often "a Chinese model that they've trained or added their own fine tuning specialization on top of, and then they call it their own model." His job, as he describes it: "It's not to pick a model. It's to use the best model for the best job".

If the model is the part you swap, the value moves to what sits around it. Harrison Chase, who runs LangChain and sells that layer, warns about lock-in one level up. "And so if you build this memory in a way that is linked to the harness and the harness is linked to the model, then you're kind of stuck." The smart routing everyone talks about is rarer than advertised. "I think that's actually being done like way less than people think." His episode was sponsored by a model-routing vendor.

Uber built for this before most. Uday Kiran Medisetty, a distinguished engineer there, describes one gateway for every model, open or closed, and wrappers that make the coding tools interchangeable. The goal: "we can easily port across different models, different harnesses." Nothing gets around it. "So there's no model request that doesn't go through this, so that gives us like one single entry point for across all the models."

Misha Laskin, whose Reflection AI sells open-weight models and the stack to run them, calls this the end of renting. "the initial market for AI was entirely a rental market, right?" Companies with production workloads now want "control of your data, lowering your costs, being able to customize the model against your data, being able to optimize your margins." He concedes who leads. "And right now, the Western frontier has not yet caught up with the Chinese models, but it is close."

Nathan Lambert, who studies open models, defines what owning buys: "real ownership, which is model weights, no terms and conditions that they can run on their hardware". The closed labs helped, without meaning to. "the model companies have been pressurized to make it so that their model can work in all sorts of different harnesses."

Last week the question was whose card the agent holds. This week it is whose model, and who can swap it.

Sources: Interviews from 20VC (Oct 12, Mike Cannon-Brookes of Atlassian; and Oct 17, Harrison Chase of LangChain), The Pragmatic Engineer (Oct 14, Uday Kiran Medisetty of Uber), This Week in Startups (Oct 14, Misha Laskin of Reflection AI), and a16z (Oct 14, Nathan Lambert of Trillium).