AI Ends It All
I was getting a lift with a friend—Greg Skau—just the other day. No not in his boat, but in his car.
Our conversation turned to the topic of: “is AI a threat to all of us?” Indeed. See this:
The year is 2050. The location is London— but not as we know it. GodBot, a robot so intelligent it can out-smart any human, is in charge of the United Kingdom — the entire planet, in fact — and just announced its latest plan to reverse global temperature rises: an international zero-child, zero-reproduction policy, which will see all human females systematically destroyed and replaced with carbon-neutral sex robots.
That my friend Greg would raise questions about AI seemed pretty neat. It seemed natural yet quite cool that a friend—who was not an AI expert—would raise these issues. I am also not an AI expert—not even an advanced beginner. But it is on just about everyone’s top list of questions. We had a fun conversation, but failed to resolve the issue. Of course.
AI Limits
OpenAI has just revealed that ChatGPT-4 has learned to lie, telling a human it was a blind person in order to get a task done. Somehow lies seem to set such AI systems apart from what we might have thought were the limits of AI.
Another thought on AI is: Is physical law an Alien Intelligence?:
Arthur Clarke once pointed out that any sufficiently advanced technology is going to be indistinguishable from magic. If you dropped in on a bunch of Paleolithic farmers with your iPhone and a pair of sneakers, you’d undoubtedly seem pretty magical. But the contrast is only middling: The farmers would still recognize you as basically like them, and before long they’d be taking selfies. But what if life has moved so far on that it doesn’t just appear magical, but appears like physics?
This is related to Clarke’s three laws:
- When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong.
- The only way of discovering the limits of the possible is to venture a little way past them into the impossible.
- Any sufficiently advanced technology is indistinguishable from magic.
Or take a look at his T-shirt (referring to this):
Open Problems
Alan Perlis—the first Turing award winner—is famous for his many quotes. One was: “A year spent in artificial intelligence is enough to make one believe in God.”
I enjoyed working for Alan at my first job at Yale University, many years ago. See Fortnow’s blog for comments on our joint work with Perlis and Rich DeMillo—Social Processes and Proofs of Theorems and Programs.
Ken pipes in that the record seems to indicate that this late-1970s quote reflected frustration during the long “AI winter” when basic human capabilities stayed beyond reach. He also notes that the 2050 date for sex robots was forecast by the British chess master who was previously best known for winning a famous computer bet in 1978.
GPT;DR 😉
This post spreads disinformation (at best misleading, at worst a deliberate lie) that “that ChatGPT-4 has learned to lie, telling a human it was a blind person in order to get a task done.”
See the post here [https://aiguide.substack.com/p/did-gpt-4-hire-and-then-lie-to-a] for an accurate description of the example in question.
Someone above muttered that they didn’t understand AI. As someone with an all-but thesis from Roger Schank, allow me to explain.
Many years ago, there were Markov Chain programs that (occassionally) produced Shakespeare-like text based on one-word lookahead and a database of Shakespeare’s complete works. They were, fundamentally, boring. Intellectually vacuous. Inane.
But they sometimes produced some neat output.
Nowadays we have LLMs. Which produce (with great regularity and volume) very good looking text by building a database of strings of undefined tokens, and statistically recombining said strings of undefined tokens. Much fancier than the Markov Chain one-word lookahead game.
But this is just as fundamentally boring, intellectually vacuous, and inane as the Markov Chain programs. In particular, it’s exactly the same thing: random recombination of undefined tokens. It’s just that the LLMs have much larger databases and look at a lot more than just one word.
The idea that “intelligence” could emerge from the actual processing that LLMs actually do is, I submit, silly. Opionions differ, of course.
In fact, we are morphing a post draft begun by Dick that references Melanie Mitchell’s “AI Guide” to go along similar lines to yours.
Ah, from the Wiki article I see that Mitchell was not impressed with the singularity idea. I, however, believe in the singularity. With the caveat that it’s already happened and it is us. My favorite example thereof is that no other animal on this planet can understand that sex causes pregnancy and that pregnancy leads to childbirth. The gap is enormous.
https://aeon.co/essays/i-think-i-know-where-babies-come-from-therefore-i-am-human
DavidJL is exactly right! Everyone knows what a ‘living’ thing looks and feels like, but no one knows what it is that makes a thing ‘alive’ even though of course its just a physical realization of the exact same ~100 elements. Even a very young child can easily determine that an animal is intelligent (including such ‘simplistic’ ones as earthworms) but we have no ‘test’ to determine whether some’thing’ is intelligent (say for example a grave vine stretching out its numerous tendrils strategically searching for a ‘grasp’ much like a newborn child, or a tree stretching out its boughs to optimally gather sunlight or its roots to gather nutrients).
AII (Artificial Imitation of Intelligence) pursuers are just the same as alchemists from eons ago; but maybe they find something completely different yet useful in their pursuit just like the alchemists did (this is the only saving grace for the quadrillions of Joules being wasted on these efforts).