Garry's Mod's Successor Walks Into a Slop Tide on Day One

s&box, Facepunch Studios' Source 2 sandbox and game creation platform, released on Steam on April 28, 2026, after years in public development. The open-source follow-up to Garry's Mod lets users build and run their own game modes on community servers, with the option to ship finished projects as standalone Steam titles. Garry Newman, who started Facepunch on the back of the original mod two decades ago, leads the project.

Steam's overall rating settled into "mixed" within hours of release, with negative coverage clustering around the volume of AI-generated game modes already filling the discovery hub. Facepunch flagged the problem in its own launch-day post and acknowledged the team needed a plan for the obvious AI submissions. Newman has promised a weekly update cadence and active moderation steps aimed at burying the worst entries.

A discovery page full of Terrys

Open s&box on launch day and the first thing it shows is the discovery feed: a grid of user-made modes ranked by popularity, with names where you can usually tell within two seconds whether a person was thinking when they wrote them. PC Gamer singled out "Terry's Granny (Now with Co-op!)" as the kind of name that's intriguing for both flattering and unflattering reasons. It's a fair pick. I clicked it. What was loaded was the joke the title implies and not much else, which is the joke and also the problem.

The Source 2 base shows in the right places. Lighting is sharper than Garry's Mod ever managed, physics interactions feel less rubbery, and the cartoon character models slot into scenes without the prop-physics jankiness of the old engine. Those models have been described by people who don't mean it as a compliment as Roblox-shaped blob blokes. They aren't entirely wrong about the resemblance. Where they're more wrong is that the resemblance is doing real work: the simplified silhouettes make user-built scenes legible from a distance in a way Garry's contorted Half-Life models never quite were. Newman has gestured at the platform becoming a kind of Roblox for dads, and the visual language is earning that pitch even when the early library isn't.

None of this translates into curation. The front page treats six-month projects and six-minute projects as the same kind of object, which is what an open submission feed plus a popularity sort always does in its first weeks. Five or six modes in, one was genuinely funny, a couple folded after a minute or two, and the rest sat in undercooked rather than actively broken territory. The amount of obviously bot-built content is real. It is also not the only thing on the page, and that distinction tends to vanish in the louder takes.

Two complaints, one bundle

Looking at the negative Steam responses in aggregate, two complaints keep arriving together. One is that AI-generated modes are smothering the discovery feed. The other is that the underlying platform itself is unfinished, unoptimized, and missing features that the comparable competition already has. Both complaints are true. They aren't the same complaint.

Kotaku caught the bundling cleanly, calling s&box Garry's Mod spun into Roblox and quoting one player's verdict that the platform is "an unfinished mess that doesn't hold up against modern engines or actual game platforms." Two charges in one sentence. The Roblox structural comparison is honest. An open submission feed plus a popularity sort plus generative tools that can produce a passable-looking mode in an afternoon equals the slop tide every UGC platform eventually weathers. Steam Workshop has weeks like this. So does itch.io's free section. The unfinished-mess part of the verdict is on Facepunch specifically, and it's the part the studio actually has working levers to fix.

I think the bundling makes the conversation around s&box cheaper than it has to be. Treating the AI flood as evidence of platform failure conflates a curation problem any host of this kind eventually faces with the real technical and feature gaps that are the studio's responsibility. The first one happens on a rolling basis to anyone running an open feed of user submissions, and Newman's stated optimistic case, that the slop will eventually fall to the bottom, depends entirely on how the ranking algorithm gets tuned. The second is what the weekly patches need to actually move on, and what the loud responses are correctly pointing at when they aren't smearing it together with the slop talk.

Newman's weirdest analogy and what's actually inside it

Ahead of launch, Newman did the usual press circuit. Most of what he said was sober. He framed AI as a fact of the new development cycle and said the team used it for code scaffolding in pre-production. Facepunch's artists were more skittish about it, he acknowledged, particularly given how the models have been trained. The studio's storefront moderation approach was, he said, to wait for actual misuse first and build the systems around the real problems rather than imagined ones.

Then he reached for an analogy. He compared over-reliance on AI in coding to over-reliance on porn, suggesting that if you let the tool do the imaginative work for you, you eventually lose the muscle, and that the right calibration sits closer to fifty-fifty between AI-assisted and unassisted work.

That bit got the laugh in coverage. Headlines treated it as a quotable swerve. I read it differently. Newman is describing what he thinks happens to a user when calibration drifts, and the description maps neatly onto the discovery page he's about to inherit, where the imagination has been outsourced, and the muscle is visibly atrophied. Whether the diagnosis is useful or self-implicating, given that the platform he runs is hosting all that atrophy, is a separate question, and one the press cycle hasn't actually gotten to yet.

What "push obvious slop off the main page" actually requires

In the days after launch, Newman told Rock Paper Shotgun that low-quality AI-generated content was going to be a growing problem across creative outlets, and that Facepunch intended to surface human-made work and bury the obvious AI builds on the main page. That's the specific commitment to watch.

The options for actually doing it are limited and well-known. Manual review works at small scale and breaks at large. Community flagging works in the long tail but creates pile-on dynamics in the short. Algorithmic detection of AI-generated content is a moving target, and it gets harder when the content in question is a Source 2 project stitched together from generated assets and human-written code rather than a single image with telltale artifacts. Newman's preferred framing in his GDC interview was to let people break the rules first so that whatever moderation system Facepunch eventually builds is shaped around the real misuse rather than guessing at it. That's defensible. It is also a multi-month timeline at minimum, and the discovery hub in its current state is already setting the conversation about the launch.

The other gap is the line between tools used in development and content presented as a finished mode. Facepunch's own Steam page concedes that while the studio doesn't generate anything with AI itself, user-submitted content on s&box "may contain AI-generated stuff." That language is honest. It is also load-bearing in a way the studio hasn't yet shown a plan around. There is a meaningful difference between a creator using an LLM to scaffold the boring code in a real project and a creator using an LLM to generate the entire experience and ship it as the same kind of object. The platform currently treats both the same way at the discovery level. Whether and how it stops doing that is the work of the next several months, and it's the bit that determines whether the launch reception ages into a footnote or sticks.

Garry's Mod is still alive and that matters

The original Garry's Mod isn't quite finished, somehow. An update arriving the day after s&box's launch is grafting the fan-made Half-Life remake Black Mesa onto it as a playable map base. That's the kind of community moment Facepunch built its first sandbox to host, and it's the kind of moment s&box has to start producing before its critical narrative hardens around the slop framing.

Newman has been candid about how the road here looked from inside the studio. He told players in his launch post that it had become embarrassing how long the project had taken him, and that if he were starting over now he could probably ship the whole thing in under a year. Someone made the obvious counter, which he relayed in the post: that's like a songwriter saying they could have written the song in five minutes if they'd already known what the song was. The honest version of why he didn't take Garry's Mod and reskin it for Source 2 is that he wanted a platform that could publish standalone games on Steam, with a real distribution pipeline behind it. That ambition bought s&box its longer development arc, and now it has bought the studio a discovery hub that looks disproportionate to the actual scope of what the project is trying to do.

I keep thinking about the Workshop comparison. The thing that made Garry's Mod the project people still talk about wasn't that the original release was clean. It wasn't. It was the years of community work that landed on top of it, much of which was sharper than anything the studio could have planned for. s&box at launch is a worse object than the version of Garry's Mod people remember, and it's possibly a better object than the version of Garry's Mod that actually shipped a decade and a half ago, before any of that community work existed yet. Whether the gap closes the way the previous one did is the question every weekly patch is now answering, and the answer takes a while to come in. The discovery page is loud right now. Most of what determines whether s&box becomes a thing won't happen there.