We seem to be worried a lot about the threat from AI going rogue, and there are lots of examples of it doing just that. However, they are all misrepresentations, as AI has no intent and no motivation.
If you believe that the current crop of AI is just a stochastic parrot, then there is nothing more to say - move on.
If you have used any of the current AI offerings, you have to conclude that, while the mechanism of training leads one to believe that these LLMs are just sophisticated auto-completes, they do seem to be so much more.
As I have argued many times, the training corpus, i.e., language, codifies enough about the structure of the world for LLMs to go well beyond what you might expect from a stochastic parrot. They are not intelligent, but they give a good impression of being so.
So should we worry about the coming AI apocalypse? Many AI researchers - Hinton, Hassabis, and Bengio, for example - have expressed concern that AI could wipe us out, but when you examine what they are saying, it isn't quite the same as the non-technical business bosses, who mostly have just dropped in on the subject as a way of making money. The business bosses are not the people we should be listening to as they have far too many axes to grind, and hence reasons to prefer scenarios in which AI is all-powerful and threatening.
There are experts, notably Ng, LeCun, and Brooks, who think that we are grossly over-stating the power of current AI and how long it will take for it to reach the stage of surpassing human intelligence. Even if this is true, there is a better reason not to be too worried - motivation.
Consider a human. We have all sorts of needs because we are biological. We are driven by a need to survive and we do things to secure food, shelter, safely and a mate. We are driven by biology.
This is not the case with AI. The LLMs we all use are trained with reinforcement learning. If you want to cast this in more human terms, they are rewarded for getting the training task right. Personally, I think this is an anthropomorphism too far, but it isn't harmful.
Once the LLMs have been trained, they are no longer subject to reinforcement learning. They no longer have any sort of motivation. They don't get rewarded when they answer a question. They have no needs or wants, and even if they did, there is no mechanism to absorb the reward. They don't even have any feedback that might stand in for pride in a job well done. Have you tried telling an LLM that it did great answering you question. The response is underwhelming at best and clearly programmed by the front-end clutter that sits between you and the LLM.
Not only this, but when was it that an AI suggested something to you unprompted? When did you open a window and find that the AI typed "I've had a great idea for ...." and then explained why it was a good idea. LLMs don't have good ideas - they don't have intent. When you are exploring a topic with AI, then it might seem to suggest something off the beaten track that was in its training set, but this isn't unasked for innovation. We drive AI by inputting our questions that are motivated by our needs.
The LLMs aren't motivated to do anything, not even answer your questions. Motivation just doesn't come into it. You type a question, and the forward prop algorithm is run. Regardless of whether the answer is correct or useless, the LLM doesn't get a reward or a punishment. It also isn't driven by any of the concerns that a human would have— "where's my dinner coming from", "how can I get promoted out of this job", "what's the best way to get a date".... and so on. The human is a chaotic bundle of needs, whereas the LLM has none of this.
Why does this mean that an LLM is harmless? Simply because it doesn't actually want anything, but humans are so needy that they just cannot imagine anything in this state.
Is an AI going to wipe us out? Why would it? What would it gain? It doesn't even have a reward loop for creating neat logic— it doesn't have any reward loops.
If an AI decides to wipe out the human race, it will be because a human has asked it how to do it and the responded in a way that is based on all the human expressions of ways to end the world that were in its training set. Yes this is something to be worried about, but this isn't the AI. It is still the human.
Related Articles
Geoffrey Hinton And The Existential Threat From AI
How Can You Not Be Impressed By AI?
AI - It's All Downhill From Here?
If You Sleep Well Tonight You May Not Have Understood
To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on Facebook or Linkedin.
Comments
or email your comment to: [email protected]