Gergely Orosz Tests New Model Writing Capabilities with Multiple Personal Interviews, Including Fable and GPT-5.6, All Resulting in Failure
The models output a large amount of text but lack deep understanding, generating "word salads" that fail to capture their professional style and thematic insights.
Orosz believes this shows that current SOTA models' "thinking" ability in expert domains is still far from meeting standards, despite being advertised as proficient in writing.
Source: Public Information
ABAB AI Insight
Gergely Orosz, founder of Pragmatic Engineer, has long produced high-quality technical content through blogs and interviews, and has previously tested the limitations of AI in programming and writing, similar to his critiques of the GPT series between 2023 and 2025.
AI companies advance models through larger parameters and training data, but actual outputs remain at the level of pattern matching, failing to truly capture expert tacit knowledge and style, with resource investment primarily focused on scale expansion rather than architectural breakthroughs.
This phenomenon is similar to the performance of early automated writing tools in journalism, or how Go AI after AlphaGo still requires human intuition; current generative AI is at an early bottleneck stage transitioning from statistical prediction to reliable cognitive agents.
Essentially, this is a technological substitution: existing paradigms fail to replace humans in creative and expert writing because "understanding" requires real-world models and feedback loops, while the current training data has a structural gap with expert tacit knowledge, leading to outputs that seem fluent but are actually hollow.
ABAB News · Cognitive Laws
Word piling is easy, real understanding is hard; scale does not equal intelligence.
Experts see through it at a glance, AI remains at the surface of imitation.
Thinking is not generating, but judging right from wrong; this is where human advantage lies.