HN users question whether post-training makes models dumber
Understand this story
A Hacker News user posted a question: post-training makes AI models stronger on benchmarks and more stable at writing code, but their actual performance is dumber. He said he mainly observed three problems with Claude: the agent is bad at writing natural language, using vague wording to avoid terms like "callee"; it cannot manage encapsulation and always produces spaghetti-style code; it is afraid of breaking existing code and only layers on top of old code rather than refactoring. He asked whether this is how the model has always been, or whether post-training has compressed outputs into predictable failure modes.
AI overview based on source reports · Updated 58 min ago
Coverage timeline
Follow the coverage to explore different perspectives.
- Hacker News AIAsk HN: Is post-training make AI models dumber?
An Ask HN post asks whether post-training, which improves benchmark performance and coding consistency, is making models worse in other ways. The author lists three observations, mostly about Claude: agents write vague natural language to avoid technical terms, produce spaghetti code that ignores encapsulation, and add workaround code rather than restructuring existing code. They ask whether this has always been the case or whether post-training is compressing outputs into predictable failure modes.
Activity for this story
Not enough continuous observations to show a trend yet.