Any of the "grasping context" stuff you're seeing is a combination of a few simple programming techniques to retain samples of what was previously typed by the user and similar techniques to give the illusion of continuity.

All of these different models being released ranging from stable diffusion to chatgpt are essentially just the classic Chinese room. There's nothing happening internally. There's no subjective experience of being ChatGPT. The model isn't updated as it interacts with you.

While the results to queries are interesting and sometimes helpful, it's still fundamentally a system just processing input and giving some kind of generated output.

These models aren't live. They're a snapshot that captures a point in time of the node weights when someone decided "good enough for now". A generous analogy might be for the patient with brain damage that can only remember things for 10 seconds. Think the movie 50 First Dates.

There's no actually learning about the world happening until the next update. Every new request you make to chatgpt is as brand new to the model as the last request.

Yes, apparently if you throw millions or billions of input parameters to a more or less standard gradient descent model you get some interesting and almost convincing results. But there's still the basic issues that these systems don't learn based on interactions.

This is mostly why I think the hype around these systems is hype right now. When an AI system starts asking questions back and the learning from them, that's the revolution.

In the meantime if probably less than 5 years, this is just where a lot of nervous investors are trying to dump money to cover the downturn.

and make no mistake, in context learning can be used in a meaningful way to resolve deficiencies. gpt-3 getting 98.5% accuracy (on even very large numbers) on addition arithmetic by describing the algorithm of addition to be performed on 2 numbers. https://arxiv.org/abs/2211.09066

In context learning is just overpowered. It's essentially implicit fine-tuning https://arxiv.org/abs/2212.10559v2

as for asking questions, llms can do that just fine. anyone who's used bing will tell you as much.

You don't know what subjective experience cGPT may or may not be having. just like i don't actually know for you or you for me. you've simply chosen to assume such.

This point is underappreciated by most people.

We are close to the stage when conversations with AI's will be indistinguishable from conversations with humans, and then philosophical quibbles about whether such beings are intelligent or conscious will be irrelevant to most people, as the proof will be in the pudding.

We're already past that stage in some limited ways, with some people and some AI's, but these limits will likely be overcome soon, and only some luddite holdouts will refuse to believe they're dealing with "real" people/beings.

It can grasp context and be very useful for a lot of things even if it doesn't learn continuously, isn't alive, can be described as a Chinese Room, etc.

So we can't adequately defend our own thought process as being any "better" than an AI except at a functional level, and AI's are quickly catching up to and even exceeding human functions, so the moat humans have drawn to defend themselves as exceptional beings is ever shrinking.

Probably not, although of course you can’t prove that. But how do you think this is relevant to the uses that people are putting it to?

When people talk about grasping context, they’re talking mainly about how relevant the responses are to the prompt. In that respect, these models are a significant improvement on their predecessors, and that’s what makes them useful. We don’t need AI to be conscious to be useful, in fact it’s probably better if its not, since conscious AI opens up a whole can of worms which we may not be ready to deal with.

The can is at least partially open already, and the worms are crawling out.

They can of worms I was alluding to is that it’s easy to justify using non-conscious machines purely as tools. But if they’re conscious, suddenly you have a new species of enslaved beings instead.

https://www.project-syndicate.org/commentary/ai-chatbots-nai...

ChatGPT flips into bullshit mode and interprets it as a straightforward offer; GPT4 grasps the humor and identifies it as a marketing strategy, and proffers several reasons why such a strategy might be effective (icebreaker for customers, encourages social media engagement, promoting a "quirky" laid-back atmosphere). It's fun to watch AI skepticism get rebutted in real time. Frightening too. We really have no idea what's about to hit us, do we...

Seeing this happen to HN is... sad.