Own the Master and the Board

Last Friday the Government issued an export-control directive against Claude Fable 5. Anthropic complied by revoking access for everyone. The conversation I started with Fable on the subway home reverted to Opus after dinner.

The reaction, as you can imagine, wasn't surprising: a renewed call to stop "renting" intelligence (from OpenAI and Anthropic) and start owning. Own the weights! Own the loop! Your landlord can evict you, so hold the deed yourself.

The rental framing treats intelligence like real estate: deeds, evictions, who can take what from whom. Where I think this analogy betrays us is that it conflates ownership with availability. Again, not surprising, since we were collectively reacting to someone taking away our new model. If you own your weights, it can't be taken away by a vendor's roadmap, a huge price hike, a deprecation notice or a letter from the Commerce Department.

But there's more to owning your intelligence than owning your weights. Availability is the floor. The word I want is fidelity. In music it's how much survives the trip from the master to your ears. A flawless recording can still come out of the speakers hissing and clipped.

The weights are a master tape: mixed once, pristine, sitting in a vault. You have never once heard a master tape. You hear it played back, through a mixing board someone else owns, in a room you don't control, mixed down live to fit the bandwidth of the moment. Fidelity is how close that playback comes to what's on the tape. It's a property of the board and the room, and both are mostly hidden from you.

What the hell does this have to do with AI?

Spend any time in engineering channels and you'll see the same questions every few weeks:

"Is Claude dumber today, or am I?"

"Is Codex totally shitting the bed today?"

"We didn't change the prompt and yet our eval suite is flaky as hell today."

You flock to Twitter, check your provider's status page, and scan every post in r/{{whatever}}. The consensus from us users is clear: Flowers for Algernon, but AGI. The consensus from providers is steadfast: the weights are unchanged. So what moved?

You find out months later, and the cause lives somewhere other than the weights, every time. A routing error that sent short requests to servers provisioned for million-token contexts. A new optimization that broke generation after seeing Thai characters, or another that made the system disagree with itself about its own argmax.

In all these cases we had the same model weights, and yet the code and infrastructure that served them warped our experience. These are all knobs on the board: routing, precision, caching, the effort knob. The master tape sat unchanged! The hiss and the clipping were the speakers'.

Owning the board means at least owning the code that runs the weights.

The whole time Anthropic, one of the most transparent, most careful, most capable operators in the business, spent six weeks unintentionally shipping you a duller model while saying, truthfully, that it left the weights alone.

It's clear to me, then, that intelligence here is a runtime phenomenon. Two deployments of byte-identical weights can give you two wholly different experiences: the precision and quantization they're served at, how the cache gets managed, the pool your request lands in, the speculative decoder racing ahead, the reasoning effort, the noisy neighbors of multitenancy.

Having your model taken away sucks, but IME being benevolently gaslit is infinitely more frustrating. And it's your future as long as they run the board, even if you own the master. They downsample the instant the room fills, and tell you, truthfully, that it sounded great.

The calls this week to "stop renting" and ownnnn skip this part. Again, owning the master solves for availability; nobody can take your weights. But we've seen for years the importance of fidelity, which is decided by the board. You need to own the master and the board.

If you don't own the serving engine, the precision, whether the KV cache is paged or quantized, how requests get batched and scheduled then every one of those choices is a vendor's right to change without telling you. If you do, then when the output changes you have recourse: you can git blame or bisect instead of flaming your vendor on r/{{whatever}}.

There's no free lunch. Anthropic made honest infra mistakes and they're no chumps. The ground can shift under you as it did for them. But the bugs can be instrumented and caught. You can watch the regression happen, measure it against your own evals, and chase it down instead of refreshing status.provider.tld and wondering whether you or your model is the dummy.

On Friday morning we had Fable. On Friday evening we didn't. We felt a sharp and acute loss of availability. The discourse this week papers over the fact that the board had been swapped under us for weeks, months, years before that, and that quieter loss, of fidelity, is the one we need to take more seriously. You can own the master, but you don't truly ownnn unless you can open the board when it starts to hiss.