Apple is reportedly about to spend a billion dollars a year to put one model under Siri. Andrej Karpathy already made the same pick for the model that runs his AI council. In both cases the model is Gemini — and in both cases it sits at the top, deciding. The synthesizer-on-top pattern is winning. That's also, quietly, how a council turns back into a single model.
The same model, handed two different jobs
The reporting ahead of Monday's WWDC keynote is unusually specific. Apple will reportedly pay Google around $1 billion a year for a custom ~1.2-trillion-parameter Gemini model — roughly eight times larger than the biggest cloud model Apple has built on its own — to power a rebuilt Siri running inside Apple's Private Cloud Compute. None of it is confirmed. Apple has announced nothing, and every line is a pre-keynote preview. But the direction is hard to miss: the most-used assistant on the largest installed base of phones on earth is about to think with Gemini.
Now look at the other end of the stack. The open-source project that put the phrase "AI council" into developers' mouths — Karpathy's llm-council — ships with Gemini in the chairman seat. The chairman is the model that reads what every other model said and writes the final answer. Apple is putting Gemini under the assistant. Karpathy put Gemini over the council. Same model, opposite positions in the stack, the same week.
That's a coincidence of news cycles, not a conspiracy. But it points at something real about where this is all heading.
A chairman is only as good as the disagreement it resolves
Slow down on why a council is worth running at all. The entire value of asking Claude, GPT-5, and Gemini the same hard question is that they fail differently. Different training data, different blind spots, different ways of being confidently wrong. When one of them is off, the others are usually off somewhere else. The disagreement between them is not noise to be smoothed away — it is the signal. It's the part a single model can never give you, because a single model can't see its own blind spot.
The synthesizer's job is to sit above that disagreement and decide what to do with it. Andy Hall's research on chairman-mediated synthesis — the basis for Shingikai's Chairperson Synthesis strategy — makes the point sharply: the value isn't in averaging the votes, it's in the chairman deciding when the model that disagreed was the one that was right.
So here's the two-part line that should make anyone pause before they cheer for the synthesizer pattern: a council is only as good as the diversity beneath it, and a chairman is only as good as the disagreement it has to resolve. Take away the diversity and the chairman has nothing left to do.
One vendor on both ends is a monoculture, not a council
Stack the layers and the problem shows up. If Gemini is generating the candidate answer and Gemini is also adjudicating the answers, you don't have deliberation. You have one model talking to itself with extra steps and a higher bill. The structure looks like a council — there's a panel, there's a chair, there's a final synthesis — but the errors are correlated all the way down. When the substrate is wrong in a particular way, the chairman built from the same lineage tends to be wrong in the same way, and it nods the mistake straight through.
This is the failure mode nobody puts on a slide, because it doesn't look like a failure. Nothing crashes. The answer comes back fluent and assured. You just quietly lost the one thing the council was supposed to buy you: a second pair of eyes that doesn't share the first pair's mistakes.
Watch it on a genuinely contestable question — a hiring trade-off, a contract clause, a "should we kill this product" call. Ask one model and you get a clean, confident paragraph. Ask Claude, GPT-5, Gemini, and an open model the same thing, and the first useful surprise is where they split. One flags a downside the others sailed past. A Traditional Council surfaces that fault line; a single model paves over it. The gap between those two outputs is the whole pitch, and you can only see it when the voices come from different places.
Diversity is the product, not the panel
The lesson Apple's reported deal and Karpathy's default both teach, from opposite directions, is that the synthesizer layer is becoming the prize — and that's exactly why the layer beneath it matters more, not less. A picker that lets you choose Gemini, or Claude, or ChatGPT is a real step up from having the choice made for you. But choosing your one model is the easy half. The harder half is deciding when one model is enough and when you need several that don't share a brain, and who gets to adjudicate when they disagree.
Sometimes one model genuinely is enough — that's what a Quick Take is for, and pretending otherwise just burns tokens. But for the questions that are expensive to get wrong, the answer and the judge should not come from the same place. The moment they do, the deliberation is theater.
Gemini under the assistant and Gemini over the council is a fine story for Gemini. It's a worse one for anyone who thought they were getting a second opinion. The next time an AI hands you a confident answer, the question worth asking isn't which model wrote it — it's whether anything in the room was built to disagree with it.
Try it free, no signup. shingik.ai