Listen, I don't wanna get personal or anything. I just want to say that I don't think enough people are talking about the current utility and future potential of LLMs to assist people who have social difficulties. Not that I'm qualified to say anything about it, but I'd like to listen.
Maybe not: apnews.com/article/chat... www.nbcnews.com/tech/tech-ne... news.northeastern.edu/2025/07/31/c...
HEY THANKS I HADN'T HEARD
Sadly, while the tech has a lot of potential, it needs a LOT of work to be safe.
So does your car.
Cars have been subject to decades of increasing safety requirements. AI will be, too, but right now, it's in the equivalent of the Model-T era.
Only the Model T does 150 mph, yeah
The Model T (en.wikipedia.org/wiki/Ford_Mo...) had a top speed of 42 mph. You might be thinking of a tesla model s.
First time seeing a metaphor, eh?
Someday weβll rehearse social interactions in our Sensobooth before big dates
Having done some LLM training, I think it will do more harm than good.
Can you elaborate on that? "Training" is an overloaded word right now.
I was doing contract work with an AI company on a variety of projects. Sometimes writing prompts, sometimes responses, sometimes research based questions, and often responding to interactions between users and AI models. The latter was particularly disturbing . . .
I decided to stop because of many issues, including disrespect, wage theft, etc. But also because it began to mess with my emotions and disturb my morals. AI is incapable of true empathy or creativity, in my opinion.
Not being a creative person myself I can't speak to that part, but I believe from my experience that LLM, in the right framework, are capable of producing output that is to me indistinguishable from genuine empathy. It's okay for us to disagree though!
I agree to some extent, but the problem is that many users do not have the capacity to think critically and recognize that they are interacting with AI. When an issue goes down a dangerous path (like red pill content) or SI, the models either divert or play along, which can be truly problematic.
I lump that all together in my mind under the umbrella of model sycophancy β the tendency of the model to be agreeable and go along with its prompter. My intuition (which is all I've got) says that that might be addressable with training and to a lesser extent prompting.
But zooming out annoyingly for a moment, I think that AI is here now and our conversations need to be around what's next. Models are sycophantic? Let's work on it. People don't understand them? Let's work on it. Genies and bottles, if the genie ignored your wish and made something up 5% of the time.
I agree with all of this. However, the AI companies are in a hurry and refuse to deal with the issues. They refuse to listen to the concerns of experts in humanities based fields. As usual money takes priority over truly helping people.
I think some hurry in this case is warranted, because there's a bit of an arms race at the moment. Sooner or later somebody's going to train a model that can reliably improve the next generation of models. I want the "good guys" to get there first. Who are the good guys? Depends on the day.
I finally came around, but if I'd had Claude or GPT as a sort of dating coach when I was young it would have made a HUGE difference