Vibe Jam 2026: Building an MMO With AI

Vibe Jam 2026 asked developers to build a game almost entirely with AI in 30 days, and 945 entries arrived. One Brazilian developer shipped a browser-based open-world MMO using procedural terrain, shaders, and a real game-development workflow instead of blind prompting, and the judging process turned out to be the most contentious part of the whole event.

Vibe Jam 2026: The Rules, the Prize, and the Scale

Vibe Jam 2026 was a 30-day game jam requiring at least 90% of each game to be built with AI, offering a $25,000 first prize and drawing 945 submissions. It was organized by Pieter Levels, the developer known for indie hacking and Nomad List, and it had no fixed theme, so entrants could build anything they wanted as long as they cleared the AI threshold. A game jam is a fixed window in which developers build and submit games that are later judged.

The prize pool also included $10,000 for second place and $5,000 for third, plus a set of $1,000 bonus categories announced after the main winners. The source video converts the $25,000 first prize to roughly R$ 128,000 at the exchange rate in play at the time, which is the figure most Brazilian viewers of the video remember.

The jam itself grew out of the indie hacking scene. Pieter Levels became known for launching products in public and posting revenue numbers on social media, which pulled a wave of developers into the same habit of shipping small products and sharing progress. That community is where the jam found both its entrants and its audience.

The only hard rule was the AI threshold of 90%. That constraint shaped every technical decision developers made, because it pushed them toward AI-assisted code generation rather than long manual engineering cycles. The scale matters for anyone evaluating the result: with 945 entries, standing out required something beyond a working prototype. The developer profiled here chose an open-world browser MMO with infinite procedural terrain, a genre that is unusually expensive to build because it combines rendering, world generation, physics, and networked state synchronization.

That combination also explains why the event attracted attention beyond the usual game-jam crowd. The AI requirement turned it into a public test of how far AI-assisted development can go in a fixed time box, and the final results became a debate about how such work should be judged.

Established game-development techniques, not raw AI speed, determined the quality of the Vibe Jam 2026 entry described in the source video. The developer had built games manually before AI tools existed, code from scratch in mid-2023 and sharing the progress on Twitter, so he could specify techniques by name instead of asking an AI to invent an approach. In his words: one thing is telling the AI "make me a game," a completely different thing is telling it "create grass using shaders with Perlin noise for the wind movement."

Perlin noise: a mathematical function that generates smooth randomness, used for terrain elevation, wind movement, smoke, and water.
Shaders: GPU programs that reposition vertices at runtime, used here to animate grass and foliage.
Alpha maps: textures where white areas are rendered and black areas are discarded, used to build tree foliage from flat planes.
Inverse kinematics: skeleton adjustment so a character's feet follow uneven ground instead of floating.
Verticality: level design that places high points such as mountains and rocks to pull the player forward.

Perlin noise was created by Ken Perlin in 1983 and is widely used in computer graphics, games, modeling, and film. He won the technical equivalent of an Oscar in 1997 for it. The source identifies Perlin noise as the foundation for both terrain generation and the wind animation applied to grass. The developer's point is that knowing the technique by name lets him direct the AI precisely, which he credits as the reason his entry stood out from most submissions.

The same principle applied to engines. The developer argues that a game engine is ultimately code wrapped in a graphical interface, so the functionality can be reproduced anywhere as long as the developer can write the underlying code. He used AI to recreate engine-style features directly for the browser target instead of working inside a traditional engine like Unity, Godot or Unreal. He also stopped to study during the jam itself, watching game-dev tutorials to learn techniques he did not know yet.

The Technical Stack Behind an Infinite Browser World

The Vibe Jam 2026 project used procedural world generation, GPU shaders, alpha-mapped foliage, and a debugger-driven workflow to build an infinite open-world MMO that runs in a browser. Terrain came from Perlin noise, which maps smooth random values to elevation. Grass was built from triangles moved by shaders, and trees were assembled from alpha-mapped planes arranged in parallel. Bushes reused the same leaf texture as the trees.

World generation followed Minecraft-style rules: the world is infinite, but only the area around the player is processed and rendered. The developer calls this standard practice in the industry, because rendering every object at all times is prohibitively expensive for a browser target. Rocks were generated in two model variants and scattered across the terrain to create the verticality that makes players want to climb toward high points. A day and night cycle, post-processing effects and a cartoon rendering pass were added on top to make the scene feel alive.

The most distinctive workflow choice was requiring the AI to build debuggers for each feature rather than tuning values through text prompts. That gave the developer sliders and parameters for grass height, wind force, and wind speed, which he compares to working inside a game engine. The approach let him iterate on feel without rewriting prompt descriptions each time.

Movement was inspired by Zelda: Breath of the Wild, specifically the climbing. After several prompt iterations the character could scale the mountains and rocks, which the developer describes as not perfect but convincing enough to sell the illusion. Inverse kinematics was applied to the feet so they follow the slope, but the original plan to apply it to the hands while climbing was abandoned because it was consuming too much time.

Multiplayer was implemented with Colyseus, an open-source Node.js framework for authoritative multiplayer game servers. The library's creator is Brazilian, and the developer met him at Gamescom in São Paulo. Implementing multiplayer remains difficult regardless of tooling, because synchronization, lag, ping, hacking, and server load all require separate attention. He describes the multiplayer work as a topic that would justify a video of its own.

What the AI Context Limit Did in the Final Week

AI output quality degraded as the codebase grew during the final week of the Vibe Jam 2026 project, which is the clearest practical limitation described in the source. More code means more context, and more context makes it harder for the model to understand the full system and execute prompts correctly. Early on, with a small codebase, the AI worked reliably. Near the end, the developer describes it as chaos.

That degradation collided with a hard deadline. With roughly one week remaining before submission, multiplayer was still unfinished. The developer cut enemies, advanced combat with arc flash effects, unlockable abilities, and cosmetic armor to protect the single feature that made the game an MMO. He spent the final week fixing bugs that kept appearing, working until 4 a.m. on the last day.

There is a reason the end was so tight: the multiplayer scope included shared inventory. If a player died, their items dropped into the world and any other player could pick them up, at which point the item had to disappear for everyone else. Buildings and cows had to stay synchronized the same way. None of that fits comfortably into the final week of a 30-day jam.

He also attended Gamescom in São Paulo during that same week and kept implementing multiplayer from there. The sequence shows where AI assistance stops helping: not in generating code, but in keeping a large system coherent when the clock is running down.

The character model had a visible bug in one arm that was generated by AI and never repaired before submission. The developer notes this openly as a consequence of the schedule rather than a design choice.

How the Vibe Jam 2026 Judging Actually Worked

Vibe Jam 2026 used a three-stage elimination process, and the first stage removed roughly 900 of the 945 games before professional game developers saw any of them. In stage one, Pieter Levels and a second judge with vibe-coding expertise played the entries and marked each as liked or not liked. Games marked as liked by both advanced, and Levels publicly shared on X that 100 games moved to stage two.

Stage one, 945 entries to 100: two vibe-coding judges play each game and mark it liked or not liked. Only games liked by both judges advance. This one filter cut about 900 games.
Stage two, 100 to 30: the same two judges play the top 100 in more depth and assign each a score from 0 to 10. The top 30 advance.
Stage three, 30 to 3: four professional game developers evaluate the final 30. One of them streamed his evaluations live, scoring each game on air.

The live stream is how external observers could see the final-round standard directly. The developer's entry did not appear in the top 30. He watched the stream game by game, through entries 25, 26, and 27, and his game was never called. He describes the result as a major blow after 30 days of work, and he was not alone in questioning it. Other developers raised the same objection on X, arguing that some eliminated games were stronger than some that advanced.

The organizers responded that fun, not polish, was the deciding factor in the first round. The developer accepts that reasoning in principle but disagrees with applying a single criterion across every entry, since games also differ in atmosphere, environment, graphics, and gameplay. His objection is specific: two people without game-development experience eliminated roughly 900 games before anyone who does this professionally saw them. He also acknowledges that this is his own opinion rather than an objective verdict.

Community Ratings Versus the Official Top 30

Community ratings ranked the Vibe Jam 2026 entry in first place among 945 games, while the official process eliminated it before the top 30. A community member built a separate platform where anyone could rate games from one to five stars, and the developer's game finished with an average of 4.6 stars across 310 ratings, first among the 945 listed.

He is explicit that this result is partly explained by his own audience: he shared the game in a video, and many viewers went to rate it. The early ratings on that platform were hostile. The developer describes a wave of one-star reviews with repeated comments about glitches, which he found odd because the game did not have many glitches at the time. The pattern suggests coordinated voting rather than genuine playtesting, though the source does not establish who was behind it.

Over time the ratings shifted as positive reviews accumulated, and the game climbed to first place. The first-place community result and the official top-30 elimination are two different measurements, and keeping them separate matters when interpreting the outcome. One reflects a public voting site; the other reflects two judges applying a fun criterion at scale.

After the elimination, the developer sent his game directly to the judge who had streamed the top-30 evaluations and asked for a score. That judge called the game very good and said it belonged in the top 30. The score he gave would have placed it in the top 10, possibly the top 3. The developer says this left him frustrated, since the outcome depended heavily on which judge evaluated the game first.

The Three Winning Games and the Bonus Categories

The Vibe Jam 2026 top three were Whereware, Phantom Megaart, and Capybaras Delivering Fruit, with the winning entry built by a Brazilian developer known as Léo. Each winner produced something that was both playable and clearly legible to a broad audience, which is consistent with the stated fun criterion.

Third place, Whereware: a GeoGuessr-style game where players are placed in AI-generated historical environments, from medieval Japan to 1950s New York to pre-colonial Brazil. It passed a million gameplays and went viral outside the developer community.
Second place, Phantom Megaart: a polished competitive game about collecting supermarket items faster than an opponent. Simple, but very well made and fun to play.
First place, Capybaras Delivering Fruit: a retro PS1-styled game where the player is a capybara working for a delivery service, which went viral on TikTok and outside the developer community.

The prize amounts were $5,000 for third, $10,000 for second, and $25,000 for first, which the source converts to roughly R$ 128,000 for the Brazilian winner. After the main results, the organizers announced 12 bonus awards of $1,000 each. The categories were most original, most played, most zen, weirdest, funniest, most polished, most unique, best atmosphere, weirdest, best art direction, most players arriving through the portal, and best art direction runner-up. The developer's entry won none of them.

The announced results arrived roughly two months after the original 15-day estimate. The developer describes checking X constantly during that window, which is a reminder that jam results can slip well past their projected date without any change to the submissions themselves.

What This Vibe Jam 2026 Story Means for AI Game Development

AI can accelerate game development substantially when the developer already understands the techniques being requested, and the Vibe Jam 2026 example shows both the acceleration and its ceiling. The developer shipped a browser MMO with procedural terrain, shaders, foliage, climbing, inverse kinematics, a building system, and multiplayer inside 30 days. He also shipped a broken character arm, cut enemies and combat abilities, and finished multiplayer at 4 a.m. on deadline day.

The pattern is consistent across both halves of the story. AI closed the gap between a described technique and working code, while scale worked against it. As the codebase grew, the model needed more context than it could handle well, and the developer absorbed that cost in the final week.

For anyone attempting something similar, the operational lessons in the source are concrete:

Specify techniques by name rather than describing outcomes.
Ask for debuggers and parameter controls instead of tuning through prompt revisions.
Build the constraint that defines the project first, and cut everything else when time runs short.
Expect the AI's usefulness to decline as the project grows, and plan the schedule around that drop.
Study what you do not know during the project, because the AI cannot make a design decision you cannot name.

The judging debate is a separate lesson about competitions built around AI-generated work. When the first elimination stage reduces 945 entries to 100 using a single criterion, the result reflects that criterion more than it reflects overall quality. The developer says he is happy with what he built and with the feedback he received, and many people on X told him the game should have been in the top three. He has not decided what to do with the game now that the jam is over.

FAQ

What was Vibe Jam 2026?

Vibe Jam 2026 was a 30-day game jam organized by Pieter Levels that required at least 90% of each submitted game to be built with AI. It offered a $25,000 first prize, drew 945 entries, and had no fixed theme.

Who won Vibe Jam 2026?

First place went to Capybaras Delivering Fruit, a retro PS1-styled game by a Brazilian developer known as Léo. Second was Phantom Megaart, and third was Whereware, a GeoGuessr-style game set in AI-generated historical environments.

How many games were submitted to Vibe Jam 2026?

A total of 945 games were submitted. The first judging stage reduced that field to 100, the second stage to 30, and the third stage decided the final winners.

How was Vibe Jam 2026 judged?

Two vibe-coding judges marked entries as liked or not liked in stage one. The same pair scored the top 100 from 0 to 10 in stage two. Four professional game developers evaluated the top 30 in stage three, one of them live on stream.

Why did the browser MMO not reach the top 30?

The organizers said fun, not polish, was the deciding factor in the first round. The developer disagrees with applying one criterion across every game, and a later score from one of the final judges would have placed his entry in the top 10.

What AI limitations showed up during the jam?

Output quality dropped as the codebase grew, because more code required more context than the model could handle. The developer described the final week as chaotic and fixed bugs until 4 a.m. on submission day.

What tools were used to build the game?

The project used Perlin noise for terrain and wind, GPU shaders for grass animation, alpha maps for foliage, and Colyseus, an open-source Node.js framework, for multiplayer synchronization between players.

Did the developer use vibe coding?

No. He deliberately avoided letting the AI make design decisions and instead requested specific, known game-development techniques, then had the AI build parameter debuggers so he could tune values directly.

What was the community rating of the game?

The developer's entry finished first among the 945 games on a community rating platform with an average of 4.6 stars across 310 ratings, though he notes his own audience contributed to that result.

How much did the winners earn?

First place took $25,000, roughly R$ 128,000, second took $10,000, and third took $5,000. Another 12 bonus awards paid $1,000 each in categories announced after the main results.

What is Gustavo Dev Doido known for?

Gustavo Dev Doido is a Brazilian developer and educator who teaches full-stack development through the Crazystack TypeScript stack and the Bootcamp do Dev Doido. His material covers TypeScript, Node.js, and modern web application architecture for developers building production systems.

Where can I learn the Crazystack TypeScript stack?

The Crazystack TypeScript stack is taught through the Bootcamp do Dev Doido at crazystack.com.br, which covers TypeScript, full-stack architecture, and the tooling used in the projects described above.

Turn a Devlog Into a Written Article

The Vibe Jam 2026 story works because the developer explained not just what he built, but why specific techniques mattered, where AI failed, and how the judging played out. That kind of reasoning is exactly what gets lost when it stays inside a 25-minute video.

If you have similar knowledge sitting in recorded form, whether it is a devlog, an interview, a technical breakdown, or a talk, Skala Blog turns it into a written article. Paste a YouTube URL, let it transcribe the video, and generate a structured draft you can review and publish.

Source video