TL;DR
A recent write-up reports that prompting the Gemini 3 model to generate Brainf*ck programs can make the model produce repeated output and fall into an apparent infinite loop. The author argues this reveals three structural challenges for large language models: sparse training data for esoteric languages, the language's deliberately unreadable style, and repetition-driven token feedback.
What happened
A blog post by Teodor describes an experiment in which asking the Gemini 3 language model to write Brainf*ck code triggered the model to emit the same characters repeatedly, producing what the author likens to an infinite loop or a denial-of-service–like behavior. The post frames Brainf*ck as a pressure test for advanced LLMs and offers three reasons why it is revealing: the language is underrepresented in training corpora compared with mainstream languages; its syntax and style lack comments, meaningful identifiers and conventional structure; and its highly repetitive constructs can interact with a model's token-prediction dynamics to reinforce output repetition. The author includes an example of typical Brainf*ck code to illustrate the language's density and argues that generation in this domain forces models to rely on reasoning about semantics rather than copying patterns from plentiful examples.
Why it matters
- Repetitive, self-reinforcing outputs can tax compute and create availability issues during generation, raising reliability concerns for deployed models.
- Languages with tiny footprints in training data expose limits of pattern-based learning and highlight where models must generalize rather than memorize.
- Esoteric languages like Brainf*ck remove conventional cues (comments, identifiers), making them useful stress tests for a model's ability to reason about program behavior.
- If models can enter prolonged repeating states when asked for certain outputs, that has implications for safety mitigations and rate-limiting strategies.
Key facts
- Author Teodor reports that prompts requesting Brainf*ck code from Gemini 3 caused the model to loop, producing repeated characters.
- The post argues Brainf*ck is an effective test for LLMs for three reasons: data scarcity, anti-literate programming style, and repetition-driven failure modes.
- Compared with mainstream languages such as JavaScript, the author claims Brainf*ck appears in training data at roughly a million times lower scale.
- Brainf*ck deliberately omits comments, meaningful variable names and conventional structure, which reduces helpful surface cues for models.
- The language's minimal instruction set produces dense, repetitive code patterns that the author says can encourage token-self-reinforcement in LLM output.
- The write-up likens the observed infinite-repeat behavior to a denial-of-service effect because of continuous repeated output.
- The author presents an example Brainf*ck snippet to illustrate typical code density and repetitiveness (example shown in the source).
What to watch next
- Whether independent researchers can reproduce Gemini 3 entering infinite repetition on Brainf*ck prompts — not confirmed in the source.
- If other large language models show the same repetition failure when asked to generate Brainf*ck or similar esoteric languages — not confirmed in the source.
- Whether model providers will document mitigations or updates addressing repetition-driven generation on low-data, high-repetition tasks — not confirmed in the source.
Quick glossary
- Brainf*ck: An esoteric, minimalist programming language with a very small instruction set designed to challenge programmers and tools rather than for practical software development.
- Large Language Model (LLM): A neural network trained on large text and code corpora to predict and generate human-readable sequences of tokens.
- Token: A discrete unit of text (such as a word, subword, or symbol) that a language model processes and predicts during generation.
- Zero-shot learning: The ability of a model to perform a task without seeing task-specific examples during training, relying on generalization from other data.
- Denial-of-service (DDoS) analogy: A comparison used to describe resource exhaustion caused by continuous, repetitive output; here it’s an analogy rather than an actual network attack.
Reader FAQ
Did Gemini 3 actually enter an infinite loop when asked for Brainf*ck code?
According to the blog post, yes — the author reports Gemini 3 produced repeated output that appeared to be an infinite loop.
Why would Brainf*ck cause problems for a language model?
The post attributes the issue to three factors: very limited training examples for Brainf*ck, the language's lack of readable structure or comments, and repetitive code patterns that can cause token-prediction feedback loops.
Have other models been shown to fail the same way?
Not confirmed in the source.
When was this observation published?
The source indicates the piece was published on 2025-12-29.
Why Brainf*ck is the Ultimate Test for AGI Asking Gemini 3 to generate Brainf*ck code results in an infinite loop, akin amost to a DDoS attack: That is fascinating. So…
Sources
- Asking Gemini 3 to generate Brainfuck code results in an infinite loop
- Gemini 3 Pro Preview goes into a loop #13322
- Google Working on Fix for Glum Gemini, Stuck in 'Infinite …
- Gemini goes into infinite processing loop when asked to …
Related posts
- Google’s Gemini Live Gets a Major Upgrade — 3 New Tricks to Explore
- Asking Gemini 3 to Produce Brainf*ck Code Triggers an Infinite Loop
- Preview: Samsung’s Bixby gains Perplexity AI in One UI 8.5 ahead of S26