Ever since the dawn of Deep Learning, the complexity and sheer scale of AI systems has far exceeded any human’s ability to comprehend their inner workings. Welcome to Black Box AI. Welcome to the AI of Mysteries. But, why? How? Very simply: because modern AI systems are all based, to varying degrees, on a very very simple codebase that is iterated upon an ubergazillion times upon a truly massive training dataset… the result of which is an “AI brain” stored across a few measly terabytes.
a Black Box AI in 700-dimensional space
The thing is, this brain is a 768-dimensional neural net — think a multidimensional tangle of wires far exceeding the complexity of the billions of neural connections in our own pink brains — that is comprised solely of 0s and 1s. It isn’t human readable… it’s only human probe-able. And those probes… are discovering… some very odd peculiarities. (not to mention, some tragic flaws)
“I cannot emphasize this enough:
We do not understand these systems,
and it’s not clear we even can.
I don’t mean that we cannot offer
a high-level account
of the basic functions:
These are typically
probabilistic algorithms
trained on digital information
that make predictions
about the next word in a sentence,
or an image in a sequence,
or some other relationship
between abstractions
that it can statistically model.
…But zoom into specifics
and the picture dissolves
into computational static.”
— Ezra Klein,
“This Changes Everything”
The New York Times — March 14, 2023
tweet/meme by @anthrupad
Watkins on the Ignorance of the Experts:
“I find that
we’re rushing ahead
and we don’t have the wisdom
to deal with this technology,” said Watkins.
“So the sooner people can realize that
the people who seemingly know what they’re doing
don’t actually understand what they’re dealing with,
maybe that will help sort of put the cultural brakes on a bit and everyone go, ‘Oh, maybe we just need to slow down a bit.’ We don’t need to rush into this. It’s getting kind of dangerous now.”
— Vice News interview
with Matthew Watkins, of SERI-MAT
regarding his report with Jessica Rumbelow
on “SolidGoldMagikarp” and other AI anomaliles
The Closed Nature of “Open” AI.
OpenAI did not respond to a request for comment.
It doesn’t help that OpenAI, the billionaire geniuses behind ChatGPT, aren’t open at all. They started as a non-profit, then quickly rethought that in the name of global security (no joke. the AI arms race is real. and lethal). Initially they published a lot of research papers explaining what they were doing. And the global AI development community was thankful.
But recently, with the success of GPT-3, and now with the insane global phenomenon that is ChatGPT, OpenAI has clammed up… completely. So while no-one knows anything about the inner workings of any of these models, now additionally no-one outside the technorati has any idea of even the broad strokes of these black box AI models.
Further Thoughts on Black Box AI:
Don’t know what’s going on? Do some Computational Neuroscience.
Probing the Deep Core of the AI Brain: It has Turrets?
The real problem: Mislaignment of AI and human destiny.
The Endgame Scenario: One way this all plays out (written by ChatGPT)