Vortex Gaming

· The TLDNE project (These Levels Do Not Exist) aimed to create an AI capable of automatically publishing generated levels in Geometry Dash every 20 minutes.
· The AI was trained on 80 GB of existing levels, learning patterns rather than memorizing complete levels.
· Its operation is compared to models like ChatGPT, predicting the next element of a level based on previous data.

· The AI processed millions of level strings from Geometry Dash versions 1.9 and 2.0.
· The resulting 1 GB model had to compress all the acquired knowledge.
· Instead of memorizing entire levels, the AI learned the general structure and patterns of levels, such as the sequence of blocks and obstacles.
· It generates levels object by object with a degree of randomness, without human supervision.

· The AI model learned to "speak" Geometry Dash's internal language, which consists of level string fragments, not conventional words or code.
· This intermediate language is what the game uses to save levels.
· The AI's source code includes a note stating that the pre-2.0 format works better, suggesting the AI was predominantly trained on older levels.

· Levels generated by TLDNE vary enormously, from strange copies of known levels to unplayable creations considered "Lost Media".
· Some levels featured unusual structures, such as long rows of blocks with UFOs, unexpected teleportation portals, or obstacle spam.
· Occasionally, the AI produced something vaguely resembling existing levels like Time Machine or Polargeist, or included elements like text written with objects or particles.
· Levels with 2.0 structures and complex triggers were found, often on white backgrounds that appeared unfinished.

· The creator shared the source code, written in Python, which reveals the simplicity of the generation process.
· The code uses the RWKV engine, an open-source project, fed with Geometry Dash level data.
· The model's "vocabulary" consists of 5,000 level string fragments, representing "chuns" or pieces of levels, not individual objects.
· The model generates 999 levels consecutively without interruption, explaining the high upload rate.
· Each generated level has a maximum length of 2048 tokens (string fragments), limiting the size of the levels.

· "Temperature" parameters control randomness: zero temperature produces predictable results, while higher values (like 0.9) introduce chaos and unpredictable outcomes.
· The generation process is summarized in four lines of code: the model predicts the next fragment based on the previous one and repeats it 2048 times.
· There is no explicit logic regarding obstacle or portal placement; generation is purely based on sequential fragment prediction.

· The analogy of monkeys endlessly typing to generate masterpieces like Romeo and Juliet is used.
· TLDNE is like monkeys who have "read" the entire library of existing levels before starting to "write," ensuring that what is generated has some sense, unlike purely random writing.
· While monkeys would need infinite time, TLDNE uploaded 5,000 levels in 2 months.
· RobTop banned TLDNE's account, an action criticized by the video's creator, especially considering the project's low cost.

· Training the AI cost $250, using 8 Nvidia A40 GPUs for approximately one day.
· The creator believes that with enough time and money, an AI could generate levels indistinguishable from human-created ones.
· However, TLDNE's model is not updated or learns from new levels; it's a "snapshot" of its training state.
· Despite the ban and the static nature of the model, the project inspired others, like "These Levels Do Not Exist Extra Electric Bualo," which continued level generation, albeit less frequently.

· Seven months after the original ban, a new account emerged, uploading fewer levels weekly, with some appearing more realistic.
· Currently, this account is inactive due to issues uploading old format levels in recent game versions.
· There is hope that the TLDNE project or similar initiatives may resurface in the future, leveraging new tools in the Geometry Dash editor.