Not Utilizing the Em Sprint

OpenAI has not achieved its objective of growing a superintelligence or synthetic basic intelligence, nor has it cracked its deliberate building of an autonomous “AI researcher.” However it has discovered tips on how to get ChatGPT to cease misusing the em sprint. So, that’s one thing.

In a post on X, CEO Sam Altman introduced, “In the event you inform ChatGPT to not use em-dashes in your customized directions, it lastly does what it’s imagined to do!” He referred to as the event a “small-but-happy win.” The corporate confirmed the potential to chop off the chatbot’s reliance on the punctuation mark in a post on Threads, the place it made ChatGPT write a proper apology for “ruining the em sprint.” Notably, the chatbot was not capable of write the apology with out utilizing an em sprint.

To that time, it looks like there’s a noteworthy distinction to be made right here. OpenAI has not discovered tips on how to get ChatGPT to make use of the em sprint in a extra applicable method or to deploy it extra sparingly by default. As an alternative, it has merely given customers the flexibility to inform ChatGPT to not use it, a change that may be made inside the chatbot’s personalization settings.

That capacity follows the release of GPT-5.1, the newest mannequin from OpenAI. One of many main factors of enchancment that the corporate hammered in its rollout of the brand new mannequin was the truth that GPT-5.1 is outwardly higher at following directions and presents extra personalization options. So the em sprint clampdown seems to only be one instance of how customers could make use of the mannequin’s extra compliant sensibilities relatively than some repair to the underlying mannequin’s general output.

The truth that the em sprint repair is one thing that has to occur on a user-by-user foundation most likely speaks to only how a lot of a black field most LLMs are. In reality, there are customers in Altman’s replies on X displaying that, regardless of the instruction, their occasion of ChatGPT continues to spit out em dashes. OpenAI’s presentation of personalization as a clear up would appear to recommend that discovering an answer at scale continues to be actually, actually arduous.

The corporate has seemingly discovered a option to weight customized directions extra closely inside its calculations when producing a response to a immediate, which may produce a outcome like an individual’s model of ChatGPT not utilizing em dashes. However it nonetheless looks like the corporate can’t determine why the issue occurred within the first place or persists. No surprise the corporate is leaning heavily into personalization and speaking much less about AGI currently.