When Altman celebrates lastly getting GPT to keep away from em dashes, he’s actually celebrating that OpenAI has tuned the newest model of GPT-5.1 (in all probability via reinforcement studying or fine-tuning) to weight customized directions extra closely in its likelihood calculations.
There’s an irony about management right here: Given the probabilistic nature of the difficulty, there’s no assure the difficulty will keep fastened. OpenAI repeatedly updates its fashions behind the scenes, even throughout the similar model quantity, adjusting outputs based mostly on consumer suggestions and new coaching runs. Every replace arrives with completely different output traits that may undo earlier behavioral tuning, a phenomenon researchers name the “alignment tax.”
Exactly tuning a neural community’s conduct isn’t but an actual science. Since all ideas encoded within the community are interconnected by values referred to as weights, adjusting one conduct can alter others in unintended methods. Repair em sprint overuse at this time, and tomorrow’s replace (aimed toward enhancing, say, coding capabilities) would possibly inadvertently convey them again, not as a result of OpenAI desires them there, however as a result of that’s the character of making an attempt to steer a statistical system with thousands and thousands of competing influences.
This will get to an implied query we talked about earlier. If controlling punctuation use continues to be a wrestle which may pop again up at any time, how far are we from AGI? We are able to’t know for certain, nevertheless it appears more and more probably that it gained’t emerge from a big language mannequin alone. That’s as a result of AGI, a expertise that will replicate human basic studying capability, would probably require true understanding and self-reflective intentional motion, not statistical sample matching that typically aligns with directions when you occur to get fortunate.
And talking of getting fortunate, some customers nonetheless aren’t having luck with controlling em sprint use outdoors of the “customized directions” function. Upon being instructed in-chat to not use em dashes inside a chat, ChatGPT up to date a saved reminiscence and replied to at least one X consumer, “Bought it—I’ll stick strictly to brief hyphens any more.”
