AI Coding Models Draw Complaints Over Cryptic Jargon; Gemini Remains a Holdout

QbitAI cited examples circulating online. One developer opened Claude to review overnight vibe coding results and saw phrases such as 'ledger 15/15 all green,' 'one leg dead,' and 'same-wave mass deletion.' The report said users complained about models including Fable and Sol, saying that once a model is connected to coding tasks, it often stops speaking plainly. It listed other hard-to-parse outputs about adding iron rules and red lines, and about re-kneading someone's skin.

Some developers argued the issue is partly translation and jargon. QbitAI reported that one commenter said 'one leg dead' likely comes from 'one leg is dead,' a normal programming term when two backup or parallel paths are called legs. But the report said other phrases still baffled users, and some said they were reluctant to read AI delivery reports carefully for fear their own language would become similarly compressed.

The report pointed to two explanations. One is overfitting to coding benchmarks. The other is the influence of chain-of-thought, or CoT. To cut token consumption, model makers have deliberately pushed AI reasoning into a compressed caveman style, QbitAI said. OpenAI was described as an early adopter. Although GPT's chain of thought is hidden, developers said they had occasionally seen examples that omitted all words not needed to convey meaning. The report said caveman-style reasoning uses about 70 percent fewer tokens than plain language for the same sentence. Claude used natural-language CoT for a long time but changed after Opus 4.6, and V4 Pro also became a victim, according to the report.

The compression trend is visible in developer tools too. QbitAI said a GitHub Skill called CaveMan, which teaches AI not to speak like a human, has attracted 104,000 stars. Its principle is to use one character whenever possible instead of two. The report joked that a Chinese version of this approach would amount to Classical Chinese, citing the Analects as an example of token-efficient writing. It also cited a commenter, Susan STEM, who argued that Chinese, especially Classical Chinese, is one of the few natural languages that was systematically compressed before the computer era.

The report noted a 2019 project called wenyan-lang, created by Lingdong Huang, then a student at Carnegie Mellon University working at the intersection of computer science and art. The programming language is designed around the style and tone of Classical Chinese, using only traditional Chinese characters and corner brackets. Its GitHub repository now has about 20,000 stars. A Hello World example is written as '吾有一言。曰「「天地，好在否！」」。書之。' and the machine outputs '同天地好在.'

Despite the broader trend, QbitAI said Google's Gemini, referred to in Chinese online slang as Jimi, still speaks in a more human style. The report attributed part of that to visual hierarchy: Gemini outputs use indentation, italics, unordered lists, white space, and a vertical guide line on the left, which can reduce noise and provide clearer signals. The report also said community sentiment has shifted. After Gemini 3.8 Flash was released last week, some users defended Gemini as the only AI they could still communicate with. It added that Gemini's coding ability and intelligence have remained steady.

The report framed the language shift as a side effect of the coding arms race, as model parameters have grown from hundreds of billions to 10 trillion. It said the hunt for benchmark performance and token efficiency has reshaped how models talk, with only a few products still producing conversational language.