Issue—for the most part—resolved

Based on this article by Benj Edwards posted on Ars Technica.

OpenAI CEO Sam Altman celebrated a "small-but-happy win" this week: ChatGPT can finally follow custom instructions to consistently avoid using the ubiquitous em dash. However, this years-long struggle to control something as simple as a punctuation mark highlights a major gap between the industry's rhetoric and reality, casting doubt on claims of rapidly approaching Artificial General Intelligence (AGI).

The core issue is that LLMs don't "follow rules" but merely shift statistical probabilities, meaning the instruction to "avoid em dashes" is only text competing against the model's vast, pattern-filled training data. If controlling punctuation remains a precarious, non-deterministic battle, it suggests that true AGI—which requires understanding and reliable intentional action—is far from emerging solely from statistical text prediction.

Check out this article.

Related Stories