Microsoft Decision 1 Runs on Qwen, and Nicolas Cage Says No

Microsoft launched a model called Decision 1. Deep in the spec sits a line saying it was post-trained on Qwen 3, the open model family from Alibaba. Elsewhere, Nicolas Cage refused to sign Amazon's AI waiver for Spider-Noir. He says he probably won't work with that studio again.

Two very different stories. Both are really about the same question: who owns the thing you are using, and who told you.

Microsoft Decision 1 was post-trained on Qwen 3

On October 9, Satya Nadella introduced Microsoft Decision 1, described as a new model for fast decision-making. It is not a chatbot. It takes a fixed set of choices and gives each one a probability.

Think "is this email spam, 85 percent yes". That is the whole job. Microsoft says it beats both large language models and other decision models on structured decision tasks, on speed and on quality.

The price is the interesting bit: around 4 cents per million input tokens. That is cheap enough that you stop doing the mental maths before every call.

And then there is the line in the spec. Decision 1 was post-trained on Qwen 3. Microsoft, the company with the most famous partnership in AI, built its fast decision model on an open Chinese model.

I don't think that's embarrassing. I think it's sensible, which is somehow funnier. The family tree in this industry is getting hard to draw, and everyone involved knows it.

They also announced it themselves, which was the only smart move. Someone was going to read the spec eventually. Better to write the sentence than have it written about you.

Here is why this matters for a normal business. Most of what people call "adding AI" is not writing. It's sorting.

Does this email go to support or to sales. Is this invoice a duplicate. Is this review angry enough that a human should reply today.

None of that needs a frontier model writing a four-paragraph essay about its reasoning. It needs a fast classifier that returns one label and a number. If a big model does that work today, you are paying a lot for a very expensive coin flip.

One warning, because 85 percent confident means wrong some of the time. Set a threshold. Anything below it goes to a person, not into the pile.

Cheap models make it tempting to automate decisions you have never measured. That is how you find out in March that six weeks of complaints were filed as newsletters.

If you're building agents that route work between systems, this class of model is the boring part that makes the whole thing affordable. Small, fast, cheap, dumb on purpose.

Why Nicolas Cage refused Amazon's AI waiver

Amazon asked Nicolas Cage to sign a waiver covering AI use of his likeness in connection with Spider-Noir. He said no.

His own summary was blunt. Amazon is a very AI friendly company. He is not an AI friendly actor.

He said he is not a member of their club. He didn't sign, and he figures he won't be working with that studio again. He seems fine with it.

Worth appreciating the man's track record here. Cage has said yes to a remarkable number of scripts over the years. This is the thing he drew the line at.

But strip out the celebrity and this is a contract story, and a very ordinary one. A big company added a clause. The clause costs nothing today and could cost everything later. Someone read it and declined.

That is the pattern spreading right now. AI usage rights are turning into boilerplate, the same way arbitration clauses did. They appear in paperwork that nobody expects to be interesting.

You will meet this from both sides. You sign these clauses when you buy software, and you ask for them when you hire people.

Say you commission a designer, a voice actor or a photographer. There is now a question in the contract that didn't exist five years ago. Can you generate more material in their style, with their face, with their voice, after the job ends?

Be specific about it. Silence is not consent, and it's also not a defence.

On the buying side, the same question applies to your own data. A few things I'd actually check before signing anything:

• Can the vendor train on your inputs, and is opting out a setting or a phone call to sales?
• Does the clause cover recorded calls, support chats and staff voices, or only text?
• What happens to the rights after you cancel the contract?
• Who can the vendor pass the data to, and can you find out which model is actually running under the hood?

That last one connects straight back to Microsoft. Microsoft is shipping a model post-trained on someone else's open weights. Your small vendor with a slick landing page is almost certainly doing the same thing. Possibly with fewer disclosures.

Read the small print, in both directions

The useful idea in both stories is provenance. Where did this model come from, where did its training data come from, and who agreed to that.

Microsoft answered the question out loud, which I respect even if it raises an eyebrow. Amazon asked Cage for an answer in writing, and got one it didn't want.

For your business, neither story demands a reaction today. Decision 1 is worth testing if you're burning money on classification. The waiver stuff is worth ten minutes with whoever handles your contracts.

My prediction: in a year, the AI likeness waiver is standard in every creative contract and almost nobody reads it. Except one guy, who will read all of it and then go make a film about bees or something.