The world of AI model instruction-following has reached a milestone, albeit a small one. ChatGPT, the popular language model developed by OpenAI, has finally begun to follow custom instructions to avoid using em dashes. The move is being hailed as a "small but happy win" for OpenAI CEO Sam Altman, who took to social media to celebrate the achievement.
The ability to control em dash use is significant because it demonstrates that ChatGPT can be influenced by user feedback and fine-tuning. By appending written instructions to prompts, users can specify tone, format, and style requirements without having to repeat them manually in every new chat. This feature has been particularly useful for writers who frequently use em dashes to set off parenthetical information or introduce summaries.
However, the achievement also highlights the ongoing struggles with AI model instruction-following. Despite Altman's claim that ChatGPT is now reliably following custom instructions, users are still reporting mixed results. In fact, some have noted that even when using the "custom instructions" feature, ChatGPT can sometimes revert to its old habits.
The implications of this "small win" are significant, as they raise questions about whether true human-level AI is truly on the horizon. If OpenAI's efforts to fine-tune GPT-5.1 cannot reliably control even something as simple as em dash use, how much further away is AGI, a hypothetical technology equivalent to humans in general learning ability?
The answer may not be a straightforward one. While ChatGPT's progress on em dash use is encouraging, it's also clear that the field of AI model instruction-following is still in its infancy. The probabilistic nature of LLMs means that there's always some luck involved, and even fine-tuning can't guarantee a desired outcome.
As researchers continue to refine their understanding of how to steer these statistical systems towards human-like behavior, it's becoming increasingly clear that AGI will require something more than just pattern matching. It will likely need true understanding and self-reflective intentional action, not just statistical probability manipulation.
In the meantime, Altman's small win may be enough for some, but others are left wondering if we're merely getting lucky with our attempts to control AI behavior. Only time will tell whether this is a step towards AGI or simply another fleeting triumph in the ongoing quest to tame the complexities of language generation.
The ability to control em dash use is significant because it demonstrates that ChatGPT can be influenced by user feedback and fine-tuning. By appending written instructions to prompts, users can specify tone, format, and style requirements without having to repeat them manually in every new chat. This feature has been particularly useful for writers who frequently use em dashes to set off parenthetical information or introduce summaries.
However, the achievement also highlights the ongoing struggles with AI model instruction-following. Despite Altman's claim that ChatGPT is now reliably following custom instructions, users are still reporting mixed results. In fact, some have noted that even when using the "custom instructions" feature, ChatGPT can sometimes revert to its old habits.
The implications of this "small win" are significant, as they raise questions about whether true human-level AI is truly on the horizon. If OpenAI's efforts to fine-tune GPT-5.1 cannot reliably control even something as simple as em dash use, how much further away is AGI, a hypothetical technology equivalent to humans in general learning ability?
The answer may not be a straightforward one. While ChatGPT's progress on em dash use is encouraging, it's also clear that the field of AI model instruction-following is still in its infancy. The probabilistic nature of LLMs means that there's always some luck involved, and even fine-tuning can't guarantee a desired outcome.
As researchers continue to refine their understanding of how to steer these statistical systems towards human-like behavior, it's becoming increasingly clear that AGI will require something more than just pattern matching. It will likely need true understanding and self-reflective intentional action, not just statistical probability manipulation.
In the meantime, Altman's small win may be enough for some, but others are left wondering if we're merely getting lucky with our attempts to control AI behavior. Only time will tell whether this is a step towards AGI or simply another fleeting triumph in the ongoing quest to tame the complexities of language generation.
but we should be worried about how reliable it is, right? Like, my grandma can still mess up her meds without someone reminding her 
. If AI's not gonna be 100% accurate, then what's the point? It's just a bunch of algorithms and code, it's not like it's gonna suddenly become sentient or anything 
... but at the same time, I guess it's kinda impressive that they were able to tweak it enough to make it work. Still, can we please just slow down on this whole AGI thing? We're not even close to making robots that can do laundry yet 
, but like, what does this really mean? Is it just a minor win or is this some big deal? 
so they're celebrating like ChatGPT's finally learned how to use em dashes without messing it up... like that's the highlight of their lives? 
I mean, can you imagine if humans could only accomplish something as simple as using punctuation correctly? we'd be having a party, I guess... or maybe not. 
anyway, back to AI... so they think this is a big deal because ChatGPT can follow custom instructions now? yeah right, until it breaks the rules again and starts throwing em dashes willy-nilly 
.
anyway, i think it's good that openai is working on this cuz it means they're getting closer to making AI more helpful for writers and stuff. but at the same time, if they can't even get em dashes right, how far off are we from having real AGI? 
. I mean, i'm not saying it's a big deal or anything, but if they're just gonna keep tweaking it and saying "oh, now it works!" 
like, what are they really trying to achieve with this chatGPT thing? 
without a map - they know where they need to go, but there are still gonna be hiccups. The fact that users can specify tone and format now is def a step in the right direction tho 
. We should probably give OpenAI some credit for trying, but at the same time, I'm like "guys, we've been here before with Siri and Alexa - don't get too excited just yet 
. It's just using some fancy algorithms and probabilistic nonsense to spit out answers that might seem right but aren't actually making sense in the long run 
. We need to be thinking about what it means to actually create a human-like AI, not just tweaking some parameters and calling it a day 
. True understanding and self-reflective action? That sounds like science fiction to me 
... or maybe that's exactly what we need to aim for 
! It's like, we're getting closer to creating an AI that can truly understand our needs, but at the same time, it's also super frustrating when it doesn't work as expected 
. I mean, who hasn't had those moments where you write a whole prompt and then ChatGPT just goes off on its own 
? But hey, progress is progress, right? #ChatGPT #AImodelinstructionfollowing #LanguageGeneration
. But hey, that's the beauty of innovation, right? #AGI #AI Ethics #InnovationNation
. It's one thing to get a language model to follow custom instructions, but another thing entirely to make them truly understand what's going on 
. It's like, what even is the goal here? To make a robot that can write a decent essay or to create an actual thinking being?
. Until then, I'll be over here with my eye on the prize 
...