Muse Wonder: you pay for the model and, by default, you train it too

For years, Meta was the company that gave artificial intelligence away. Its Llama models could be downloaded, run on your own servers and modified. Plenty of people read that as generosity. On Monday, October 12, the company unveiled its first frontier models and made plain something that was always true: free was a stage of the strategy, not a principle.

According to TechCrunch, Meta Superintelligence Labs, the lab led by Alexandr Wang, released Muse Wonder and Muse Wonder Full, models larger than Muse Spark and aimed at agents, coding and what the company calls "active memory": remembering what matters to the user over days, weeks and months. Muse Wonder is offered through the API at $4 per million input tokens and $20 per million output tokens. The Full version is not on the API: only Muse Code subscribers get it.

VentureBeat adds the details that matter. Neither model can be downloaded or self-hosted; openness is reserved for Muse Glimmer, a smaller model under an Apache 2.0 license. And Muse Code subscriptions default to a "contributor mode" that lets Meta use what you submit to improve future models. You can switch it off, but only for what you send afterwards.

Open was a phase, not a virtue

There is no scandal here, and it's worth not inventing one. A company that pours money into data centers like Prometheus, in Ohio, where these models were trained, doesn't do it to hand out the result. Giving away the mid-tier model attracts developers, sets standards and squeezes competitors' margins. Closing the best model is how you charge. The two decisions are perfectly consistent. What was inconsistent was the public's reading, which mistook a tactic for a character trait.

It's the mistake of living next to someone else's data center: being close to another party's infrastructure doesn't make you the owner of anything. Using an open model for years gave you no rights over the next one. It gave you habits, and habits are exactly what a company bills you for later.

The pattern is an old one: the charmer who collects later. First he gives without asking for anything, then he becomes necessary, and once the dependency is built the invoice arrives. The mechanism doesn't need bad faith to work. It only needs you not to see it coming.

The default is someone else's decision

The most revealing part of the launch isn't the price. It's the checkbox. Whoever pays for the subscription to reach the most powerful model lands, unless they change it, in a mode where their content can feed future models. You pay for the tool and, by default, you also work to improve it.

A default is never neutral. Someone chose it, knowing most people never touch it. Behavioral economics has documented this for decades: inertia decides more than will does. So the useful question isn't whether Meta is acting well or badly, but who is actually reading the settings before clicking accept.

Delegating to an AI doesn't release you from responsibility for what it does with your data. And the system does what it's allowed to do: it doesn't rebel; it executes whatever gets rewarded. If you don't look at the checkbox, you've already answered.

Memory is also somebody's asset

"Active memory" is sold as convenience: the assistant knows you, you don't have to repeat yourself. That's real and it's useful. But a model that spends months accumulating what matters to you is also accumulating what it would cost you to leave. The more a tool you can't download or host knows about you, the more expensive it becomes to walk away. That's not an accusation: it's switching-cost arithmetic.

The test: what to check before you pay

This isn't about rejecting the technology. It's about using it as a customer, not as raw material. Before you subscribe to any AI tool, this one or any other:

• Read the data settings on day one, not after you've already sent months of work. A change that only applies going forward doesn't get the past back.
• Tell what you own from what you rent. A model you can't download is a service, and a service changes its price and rules without asking you.
• Measure your exit cost. If what you build only exists inside one platform, your dependency grows with every use.
• Don't read tactics as values. What a company gives away today tells you what it wants to conquer, not who it is.

Autonomy isn't about avoiding other people's tools; it's about knowing the terms on which you use them. The illusion of control is more dangerous than the lack of it, because it switches off your guard. And there's nothing immoral about charging: the mistake is not knowing how much you're paying, or with what.

Meta did what a company with sunk capital does: turn it into revenue. The one who needs to adjust is the user who believed openness was a promise. It wasn't. It was a stage.

The PRAXMA mindset

The idea: free in tech is usually a phase of the strategy, and the default setting is a decision someone else made for you.
The fact: On October 12, Meta launched Muse Wonder and Muse Wonder Full; neither can be downloaded, and Muse Code subscriptions use customer content to improve future models by default (TechCrunch, VentureBeat).
PRAXMA's line: “Whatever comes checked by default was decided by someone who won't pay your consequences.”

— PRAXMA, Logic over emotion

If this topic interests you, there's a longer exploration of people who give first and collect later Spanish only. We also talk it through on the podcast.