TL;DR. Eleven days ago we argued that the hard part of a personal assistant is not the model, it is reaching your life at all. Last week Nvidia published research pointing at a second bottleneck that is also not the model: the software wrapped around it. Same conclusion, different mechanism, and between them they cover most of what actually decides whether an assistant is any good. We think this matters more for personal AI than for anything else, and we say below why that part is our claim rather than theirs.
On 14 August we published a piece arguing that the bottleneck was never the model. The short version: “perfect context of your whole life” is a claim about data, not about intelligence. Your life is in ten years of Mail, a Messages database going back to your first iPhone, and a WhatsApp export that officially does not exist. On a Mac every one of those sits behind a permission gate, and those gates do not open for a cleverer model.
On 21 August Nvidia published research that lands on the same conclusion from a completely different direction.
What they actually found
They took Claude Opus 5 and ran it against ARC-AGI-3, a reasoning benchmark. Unmodified, it scored 30%. They then left the model entirely alone and rebuilt the software around it: the tool set, how memory and context are managed, how feedback returns to the agent, plus a supervising component that steps in when the agent wanders. Same model. The score went to 100%.
Nvidia calls that wrapper the harness. Their VP Adel El Hallak put it about as bluntly as it can be put: “It is the model. It is the scaffolding around the model, which we call the harness.”
They are not the only data point in the piece. OpenAI reportedly tripled its own ARC-AGI-3 score by changing two harness settings. Databricks is cited finding that harness choice can swing operating cost by 2x, which is more than the choice of model does.
Two different bottlenecks, and it is worth being precise about that
It would be easy, and slightly dishonest, for us to present this as vindication of the exact thing we wrote. It is not quite that, and the difference is the interesting part.
Our bottleneck is getting the data in: permissions, exports, formats, the decade of material that no model has ever seen. Nvidia's bottleneck is what happens around the model at run time: tools, memory, feedback, supervision. Those are not the same claim. You could fix either one and still be stuck on the other.
What they share is the shape. Both are plumbing. Both are unglamorous, invisible when they work, and absent from every launch keynote. And both are now, on the evidence, more decisive than the number of parameters everyone argues about.
Why we think this bites harder for personal AI
This next part is our argument, not Nvidia's. They tested agents on puzzles and said nothing whatsoever about personal assistants. But the logic transfers, and it transfers with more force rather than less.
A general model answering a general question can lean on what it absorbed from the internet. It has read a great deal about most things. But nothing on the internet says what you promised your brother in February, or that the person emailing you today is the one you met at a conference four years ago. There is no amount of parameters that recovers a fact the model was never given. For a personal assistant, the scaffolding is not an optimisation on top of the intelligence. It is the part that supplies the subject matter.
Which means the two bottlenecks compound. The harness can only work with the context it can reach, and the context can only be reached if something did the unglamorous work of getting it out of Mail and Messages and WhatsApp first. Fix the second without the first and you have an excellent reasoning system with nothing personal to reason about.
What we are not going to claim
ARC-AGI-3 is a puzzle benchmark. Going from 30% to 100% on abstract reasoning grids is a real result about agent scaffolding, and it is not a result about summarising what happened with your sister last year. Nobody has demonstrated that transfer, including us, and we are not going to quote the number as though it were ours.
Nvidia also has an obvious commercial interest in the conclusion that you should invest more in scaffolding and in the compute to run it. That does not make the finding wrong. It does mean it should be read the way you would read any research published by a party who benefits from believing it.
And a harness that supervises itself with a second agent costs tokens and wall-clock. On a Mac running one model, that budget is real in a way it is not in a data centre. Some of what Nvidia can afford to do, we cannot, and we would rather say so than imply our scaffolding is measured to be better than anyone's. It is not measured at all. That is a fair thing to hold us to later.
The part that does not depend on who is right
There is a strategic reading here that survives whichever bottleneck turns out to dominate.
If the model is not the thing that decides quality, then the model is the most swappable part of the system. That is worth sitting with, because it is the same conclusion we reached from the supply-chain side last week, by an entirely different route. If the weights are a file on your disk and the file is replaceable, then the layer everyone worries about losing is the layer that matters least.
The parts that are genuinely hard to replace are the ones nobody photographs: the permission work, the parsers, the identity resolution, the graph, the retrieval, the thing that notices. That is most of what we have been building, and for a long time it was difficult to explain why it took so long, because you cannot demo plumbing.
It is a little easier to explain this week.