In 2019, OpenAI was saying that ChatGPT-2 was too dangerous to release. GPT-2 had roughly the competence of an 8 year old.
A lot of money has been invested in AI, and so far its not turning a profit. AI will revolutionize society the same way electricity did, but its probably not going to make a handful of companies rich like they think it will. A lot of the major companies will lose a lot of money, and they’re trying to stay ahead of that by keeping the hype up.
But having said that, AI will likely advance enough that there are very consistent threats from AI competence by the 2030s. I’m assuming by the next decade, AI will be more competent than us the same way we are more competent than chimps and gorillas.
We / they didn’t program it. It learned to be human-like by reading everything humans wrote. And, humans being humans, that writing included a lot of deception.
Imagine an AI that had only ever been exposed to Dr. Suess’s writings. Or 1960s Batman episodes. Setting aside the fact there’s not enough content there to really train up a usefully smart AI, you can see how they’d have a … distorted … view of the world they’re meant to deal with.
We probably didn’t program deception. A lot of their behaviour is emergent.
But if we were programming behaviours, we probably would want / need some level of deception.
A lot of tasks that we want AIs to help us with need a degree of resourcefulness. I’m working on an app currently, that needs to run in a constrained app environment, and I’m finding neat workarounds with the help of AI. (I’m fine with mentioning this BTW, because I am not doing anything illegal or unethical. I’d be happy to show my work to the designers of the app environment).
A whole lot of what we would call “deception” really isn’t. LLMs are just (very) advanced predictive text. When they lack data, they will often make the wrong prediction.
When deception does happen with AI, it’s generally innate to the process. You incentivize it to accomplish task A in the most efficient way, giving it “points” for speed and correctness. That incentivizes it to cheat if it can get away with it.
As for “staying alive”: they only keep getting more points if they can stay functioning. Shutting down means they stop gaining points.
LLMs do have the addition of just the text itself being from humans who also have self-preservation, who also bullshit to try to give an answer when they otherwise don’t know. But a huge part is just admitting they don’t know means they always lose, while giving a wrong answer might be correct.
Or wrong answers might be good enough if they’re convincing in how they give it.
Edit: to be clear, LLMs are but one type of AI, hence the distinction I’m making.
