“A” is for Alien: Analogies Between Modern AI and Richard Feynman’s “The Character of Physical Law”
What I learned from my research into adversarial and generative deep learning: AI is not Artificial Intelligence — it’s Alien Intelligence
Back in the Jurassic Period of AI (aka 2019), my colleague Jason Moore and I published the following opinion piece in Newsweek:
The article was not about space critters. “Artificial Intelligences ,” we wrote, “are the ‘aliens’ of which we speak”.
Fast forward four years and the alien nature of AI, in particular deep networks, has become painfully manifest. I’ve personally come across this alienness time and again in my own recent research into adversarial attacks, where networks are made to swerve from their expected (and desired) behavior.
For example, an undetectable (to us) change in an image can cause a misclassification in a well-trained deep network.
Take a gander at these three pictures (the original is from the well-known ImageNet dataset):
They all look the same to us.
But, fascinatingly, they’re not at all the same as far as a neural network is concerned: “Successful attack” means the network misclassified the image.
Want to go beyond images — say, both text and pictures? Once again, you’ll encounter weird stuff.
Look at these two pictures:
Spot any differences? Doubtful. Yet a neural network captioned the left one as “A man and a woman posing for a picture”, while the right one was captioned as “A table with a bunch of stuffed animals on top of it.”
Get outta here!
And let me herald one final example of alienness, which came out of our research just a few days ago: Large language models (LLMs) can be duped into giving away harmful stuff they’re not supposed to.
On the left you can see how the LLM refuses to “talk” to me, but on the right you’ll notice that with the right prodding (and a kooky prodding at that!)— the LLM becomes quite garrulous:
Alien with a capital A. 👽
These examples show how differently they think. Oops, I said “think” — but, hey, forget about it, I’m not going to fall into the bottomless pit of the debate on whether machines really think or not!
But all of this got me thinking (hopefully my own meager propensity to cogitate once in a while is not up for debate). Specifically, I mulled over the famous quote by renowned physicist Richard Feynman: “I think I can safely say that nobody understands AI.” Ahem, he actually said quantum mechanic, not AI…
Mull, mull, mull — until I finally decided to do a little digging about the quote’s contexture, only to find parallels with the current AI scene.
The quote appears in the book The Character of Physical Law, published in 1965, which is an edited version of a series of lectures Feynman gave the previous year (the 1964 Messenger Lectures at Cornell University).
It appears in Chapter 6, “Probability and Uncertainty — the Quantum Mechanical view of Nature”. What I found interesting in the context of AI is what led up to this quote — and how it resonates with today’s deep networks.
The chapter opens by stating that experimental observation begins with intuition, based on simple experience with everyday objects. Slowly, the range of observed phenomena extends and explanations become so-called laws. Now here’s the thing about laws, says Feynman:
One odd characteristic is that they often seem to become more and more unreasonable and more and more intuitively far from obvious.
“More and more unreasonable”, “far from obvious” — sound familiar?
Moving on.
As we continue to experiment, says Feynman:
… we see unexpected things: we see things that are far from what we would guess — far from what we could have imagined. Our imagination is stretched to the utmost, not, as in fiction, to imagine things which are not really there, but just to comprehend those things which are there.
“To comprehend those things which are there” describes our whole struggle with understanding AI (witness, among others, the important, growing subfield of XAI — Explainable AI).
Feynman goes on to exemplify this stretching of imagination and the attempts to comprehend extant phenomena with the history of light: At first thought to be particles, then waves, and, finally, with the advent of quantum mechanics — both, behaving “in their own inimitable way, which technically could be called a quantum mechanical way”:
They behave in a way that is like nothing that you have ever seen before.
Ooh. I am so reading AI into that.
Electrons behave in this respect in exactly the same way as photons; they are both screwy, but in exactly the same way.
I just love his use of “screwy”: It’s exactly how I’d describe the deep network behaviors I presented earlier.
How they behave, therefore, takes a great deal of imagination to appreciate, because we are going to describe something which is different from anything you know about.
“Different from anything you know about” — I’ll say!
It will be difficult. But the difficulty really is psychological and exists in the perpetual torment that results from your saying to yourself, ‘But how can it be like that?’ which is a reflection of uncontrolled but utterly vain desire to see it in terms of something familiar.
The desire to see something familiar is even stronger, in my opinion, where intelligence is concerned. We keep trying to find the familiar in our AI creations — but this familiarity is slippery. Very slippery.
And then Feynman goes on to state:
I will not describe it in terms of an analogy with something familiar; I will simply describe it.
Maybe it is indeed time to abandon any pretext of familiarity!
A bit later comes that celebrated “I think I can safely say that nobody understands quantum mechanics.”
Again, Feynman:
Do not keep saying to yourself, if you can possibly avoid it, ‘But how can it be like that?’ because you will get ‘down the drain’, into a blind alley from which nobody has yet escaped. Nobody knows how it can be like that.
Yup. Nobody knows how it can be like that. And while Feynman talked of quantum mechanics — for me, it’s AI:
Alien Intelligence.
(Oh, and I do have some free associations regarding that “down the drain” bit, but I’ll spare you…😏)
