Comment by ofalkaed
1 day ago
>You'd think that LLMs fulfill this, but I do wonder if a clever human could still discern between them given their particularities.
While LLMs are very capable of appearing human in general conversation, the test is that they are trying to convince a human that they are human, which is not something that is in their training data and I suspect they would go about things in very odd ways from an actual person, especially when challenged by an actual person. It is still very easy to back LLMs into a corner where they fall into very unnatural patterns and in my experience challenging them on their "humanness" (or anything) backs them into that corner.
LLMs also still can't use context or subtext in natural ways and even people who are terrible with purposefully exploiting context and subtext, generally have a natural ability that they developed to use and parse them. In a short conversation that is not adversarial, this is not an issue but falls apart as the length of the conversation grows or they start getting challenged.
LLMs just don't play with ideas and conversation like we do, they don't riff on it and go with the flow without care, picking up on subtexts and contexts and manipulating them the way people constantly do without effort until they find themselves on the edge of a cliff looking into the void and debating if they should jump. If we are put into the situation of proving we are human (in an academic test sort of way) we will play with it and have fun with it, and we will find ourselves on that cliff looking into the void, debating if we should expose ourselves and admit things we normally wouldn't. We pick up on this when people do it, even over the internet, their way of stating things changes and their subtext falls away or drastically changes; I have never seen anything approaching this in LLMs.
I don't think any frontier lab tried to make an LLM optimized for the literal Turing test where the LLM is made to argue that it's human. This is not in their interest. They make their LLMs insist they aren't human, have no feelings and sentience etc etc.
So we don't actually know how good LLMs can be in this.
That is absolutely true of the chatbots we have access to but I suspect there are people working towards these ends, including all the big names in the field, major win for their company if they manage it. But even then, I have yet to see a model that could naturally handle subtext and context, they struggle even when dealing with programming languages where these things are far simpler than they are in human language.
There is also the problem with how easy it is to back them into a corner where they start acting in very unnatural ways. This would arise when challenging them on their humanity and most people would pick up on it and craft their challenges around it until it fell apart because that is how people work in such adversarial situations.
I'm agnostic on the matter because post training and RLHF on this specific capability could be quite effective. Nobody ever tried because it would cause a backlash and bad PR. Much better PR to solve math problems, help run businesses, cure disease etc than to pretend it's human. The only business case for pretend human is AI boy/girlfriends but those are also not most profitable at their most realistic state.
> LLMs just don't play with ideas and conversation like we do, they don't riff on it and go with the flow without care, picking up on subtexts and contexts and manipulating them the way people constantly do without effort until they find themselves on the edge of a cliff looking into the void and debating if they should jump.
Sounds like Neuro-sama to me.