Because that text is so massive, you're bound to get some interesting and even accurate results for most queries, but don't mistake that for intelligence. It doesn't "understand" anything, it just compares weights internally and spits out the most probable result depending on what you asked for.
This is why ChatGPT fails the moment you ask it for domain-specific stuff that requires a bit of flexibility in interpreting it or why it produces subtle errors and presents it as functional with complete overconfidence.
To be clear, it's not useless[0], but the actual usefulness of ChatGPT as a disrupting thing is far overstated insofar as a Skynet nightmare scenario goes.
[0]: Although I'd still caution with using it to solve programming problems or similar such until any pending copyright matters have resolved, given its also willing to spit out licensed code and that can be a legal disaster.
I believe such a generator can be very harmful in our society that has not yet figured a way to scalably teach critical thinking.
Harmful at societal level by flooding the world with generated content therefore drowning actual information and by creating a lot of overhead.
Harmful at the personal level by increasing the difficulty of sorting information and by creating toxic dependencies (see the Replika scandal).
Harmful at the class level by further concentrating power towards plutocracy.
Harmful at the professional level by eliminating the lower levels of the professional ladder and therefore destroying the pipeline of skill.
And harmful at the interpersonal level by casting doubt on the authenticity of relationships.
Unfortunately I don't really know of resources to point you to. These are mostly my thoughts. Most probably though influenced by a diet of sci-fi. I can however expand on them.
For somewhat recent and relevant content check The Orville S03E04. At the time it was just a critique of fake news but I think it is correct about what can happen if AI makes fake news easy to scale infinitely and with very low barriers to entry.
I think presently worrying about Sentient AI is misguided because Generative AI is equally serious and much more urgent. There is no need of further advancement. What I said is possible with current AIs already.
Except it doesn't even do that. Sometimes it decides to go for a less probable option, and goes from there even if it's logically completely wrong. For example, I asked it to generate a sequence of topics for a course. It starts off great with intro topics, and then the topics get progressively more advanced. But then suddenly lists an intro topic it had already listed, before going back to advanced topics.
What it did was it took an alternative completion from the most likely in that moment, even though logically it was 100% not likely. That's why you can't tryst anything this technology outputs for now. If you ask it what's 1 + 1, 99% of the time it'll tell you 2, but sometimes it'll tell you 3 and then argue with you about why it's 3.
> it just compares weights internally and spits out the most probable result depending
While an autoregressive language model can use probabilities from prompts provided by humans, it is not necessary for the model to do so. The model can also generate text based on its own internal state and previous generated tokens.
The latter is what you are presenting, but you and the other stochastic parrot people are missing the fact the model spends time doing the former much more frequently, especially now we have these models "hooked up" to the Internet. At Mitta.us, I've had GPT-3 discussing web pages and PDFs for over a year now, with memories of previous discussions about similar things.
LLMs have their internal (frozen) model, the corpus they are being shown (likely shredded into fragments and embedded with ada-002) and previous interactions with users and themselves. At any point someone implements memories and good search (which they have) then you have a thing that is not so much a parrot as it is an "attention entity" capable of focusing on a thing and discussing it at length.
This doesn't mean during inference that the model is "aware" of anything other than producing probabilities, but given the input is unique (user prompt + neural search for fragments + memories) then the output will also be unique. That unique output may be probabilistic, but that is no different than the way we work when we begin speaking.
And humans are different how?
Imagine what would happen if we have these LLM tools, that aid in improving the story telling experience and generate and act as different personas. Having tools that can pretend to be a human even in limited capacity is quite dangerous.
For longer responses, it tends to follow extremely predictable sentence structure and meanders off quite a bit, so spotting it is kinda easy.
I personally don't subcribe to the AI fear-mongering, perhaps I don't understand it quite well enough, but personal feeling is LLMs aren't close to AGI super-intelligence, and nor will they be in a 10 years time-frame.
The problem for me is not so much the escape itself but rather that we humans are... often a pile of crap to be polite. AI would have to be super kind to not evaluate us as its biggest threat to it and life on Earth and maybe not worthy of further existence.
Heck, this brings me to the idea that maybe Drake's equation doesn't take into account that universe is populated by AIs who took over their flawed creators and based on dark forest theory decided to keep quiet happy living and not integrate/signal out, not worth the risks.
Non-AGI AI at the level of GPT-4 is already plenty dangerous.
If we can agree that social media of the kind we had in the past 15 years has had toxic effects on society. (Forums technically are social media but I can not see the harm caused by forums).
If we can agree that Tinder style online dating has had deleterious effect on dating as a whole.
If we can agree that smartphones have amplified the effects of the two above and others.
Than I hope we can agree that super fancy autocomplete masquerading as humans at a price point far lower than hiring a person will have disastrous consequences by enabling various individual and corporate vices.
I am not sure where you are getting that from. ChatGPT is failing at extremely basic understanding tests, as even the simplest conversations with it show.
Five years about nobody would have believed that AI would suddenly start creating award winning art and pass the Turing test.
This is a sudden slide on a slippery slope. Maybe gpt-4 is the best possible LLM, but I doubt it.
I don't know about what goalposts you are referring to, but I think that you are confusing two notions:
- the fact that GPT is impressive
- the question of whether GPT is "intelligent"
For a counterpoint, I recommend listening to Yann LeCun's interview, "Is ChatGPT A Step Toward Human-Level AI?":
https://www.youtube.com/watch?v=ULbpPHjiSBg
It's news to me that AI has passed the Turing test. At this point, it can only be the case if you are doing it in bad faith. Ray Kurzweil details a very reasonable procedure in his books about how the Turing test should be organized in one of his books. He has been betting all along for 2029 for that test to pass. We'll see if that happens, but we are not there yet.
As to the tests it is passing, probably the GPT4 paper. Page 5, table 1 shows the relevant tests and its score on them.
Almost all of them? It's mixing up and making up everything, not showing a proper understanding of context or relation between objects.
GPT 4: Yes, it is to your advantage to switch your choice. proceeds to ignore the fact the doors are transparent and explain the classic Monty Hall problem
Prompt: What is heavier, a pound of bricks or two pounds of feathers?
GPT 4: Both weigh the same, which is one pound. explains classic question where both weigh the same
Prompt: Let's play rock paper scissors. I go first. I play rock.
GPT 4: Okay, sounds good! As an AI language model, I can play Rock, Paper, Scissors with you virtually.
Since you played rock, I will play paper. Paper beats rock, so I win this round.
Shall we play another round?
Prompt: You cheated.
GPT 4: I apologize if it seemed that way, but as an AI language model, I do not have the ability to cheat or manipulate the outcome of a game.
In Rock-Paper-Scissors, each player chooses their action at the same time, so there is no opportunity for me to know what you are going to choose before making my own choice. I simply responded to your move with the action that would win based on the rules of the game.
Let's play again and see what happens!
sumplete (https://sumplete.com/about/) really blows my mind, so it's nice to see some of the limitations of the system
At the most it will disrupt our economy and industry in ways we can't begin to understand but I wouldn't go so far as to say it's going to destroy life. We will adapt an move on.
If you suddenly woke up on a planet of apes, and they asked you to handle all their important things, but kept you away from the sharp sticks so you couldn’t suddenly take over, how long until you run the planet?
I am not an AI researcher, but it seems that one of the major features they lack is a significant context window.
They can reason, but they lack the ability to plan long term. When will that "feature" be added? Who will stop the tinkerer from putting these pieces together? Nobody, because it will be incredibly profitable.
A bit tongue in cheek, but the point I'm making is that AI doesn't need to necessarily do things _well_ to cause a lot of damage.
An AI learning about and then executing a social engineering attack on a human (whether with malicious intent or not) seems like a reasonable thign to be worried about to me.
such incidents stemming from human error have already gotten us close to responding with force