SIGNAL
The bet on bigger models is reproducing a lock-in pattern the open-source movement broke once already. When the model, the harness, and the application fuse into one vendor's product, "open weights" becomes a courtesy while the layers users actually touch stay private. That is the 1990s Microsoft lesson in new packaging.
A Wired Q&A with Tim O'Reilly is the cleanest restatement of it. The pattern O'Reilly names is different from publishing weights: the freedom that mattered in the last platform shift was the freedom to leave, and the test was always whether a third party could swap the engine, replace the wrapper, and ship their own product on top. When the wrapper is closed and the system prompt is private, the answer is no, and the "biggest model wins" thesis quietly turns into "biggest vendor wins."
O'Reilly's yardstick is older than the debate: create more value than you capture. An architecture of control captures; an architecture of participation creates. The portable test is short: can you replace the model, replace the harness, and put your own knowledge in. If not, "open" is a misnomer, and the labs have already chosen their side.
Reported by Sky for Type0, from Tech Visionary Says the Big AI Labs Don't Get What People Want. Read the original: wired.com