Something Big Is (not) Happening
Or is it
happening, perhaps it has already happened? Maybe it's just acting like it has
happened, because it is tired and would like to go to bed?
Although
trains can get you far going straight across long, flat distances, the sherpas
are still gainfully employed since we have not yet succeeded in getting
locomotives up tall mountains--then again we have cable cars that reach some
distance, we have crampons and pickaxes for the hard pack and ice walls, winter
coats to wrap the vital, strong, cunning individual who would stand over the
peak like some great romantic figure…
…It's always the
little things the AI can't figure out. How many Rs in the word Strawberry--did
they fix that one, too? Or how to identify a seahorse emoji, or create a
summary that gets all the little details right, every little particularity,
every grain of sand on the beach, and not just the general shape of its curve.
In
my business we make our money by finding out ways to get around the AI,
we call them "Stumpers." You discover that there are a great many things LLMs
are incapable of, and they are often incapable in surprising, yet increasingly
predictable ways. But they are really good at programming; because, one might
say, that's how the "intelligence explosion" happens--they focused on
programming first so that they could program themselves, generate themselves,
autopoiesis, autocatalysis, autonomous self-generating self-regenerating
accelerating hyper machines from the future that have come to kill the people
who would stop them in the past! (I think we've all seen this movie before).
Excuse
my French, but in the biz we might call this a
posteriori logic. You take what you have observed, empirically, and work
backwards to theorize how that thing must've come about. You invent a narrative
that makes sense. But, the world is not a good story,
even if we love to tell stories about it. Let's consider the case more
concretely:
You
flip a switch and either the light bulb turns on or it doesn't turn on. If your
intent was to turn the light on, when it does go on, you say the switch
"works." It doesn't matter if you know how, if you even understand the
underlying mechanism, as long as it does what it was "supposed" to do.
A
computer is something like this, it's basically a very complicated kind of
lightbulb, with many different kinds of switches and levers and buttons and
other inputs that go through numerous logic gates and even, like an
etch-a-sketch, has some stores of memory that can be written or erased at will.
There is a way to change the color on the screen, there is a way to simulate
complex physics models, there is a way to execute high frequency trades; in
each case, there is a success state: the screen turns green, the results match
the model, the trade is profitable.
Since
programming is only just that, creating something that does what it's supposed
to do in any case, as long as it does that thing it doesn't matter why it
works, as long as it works. It's an immediately practical discipline, and
unique amongst the various forms of writing in that it is beyond polysemy,
beyond the possibility of multiple interpretations. Even the law, with its
numerous codes and cases and precedent, is still subject to interpretation, ceaseless
interpretation, interpretation so vast, so constant, that our entire
society hinges, sometimes, on the thoughts of nine individuals in DC (or more
likely their assistants'). AI assistants are still very good with the law,
especially for looking up relevant case law, for getting a quick overview of a
said case. Although, if there were twenty millions
dollars on the line, would you still trust the model? If it was a life or death decision, would you trust the model?
Judgement, yes, but decision? No, they are not capable of making a
decision, at least important ones[1].
Computer programming, on the other hand, is not about making decisions, it's
about mechanical functions. So Claude might go back
and change some things, may appear to make judgements, have particular tastes.
This is a behavior that was probably intended by its architects, this recursive
evaluation, which was trained on the work of hundreds of talented programmers
who were, to be sure, ready at hand. It doesn't matter if you're writing a
script that outputs "Hello World" or creating an entire video game, as long as
it works, as long as the material is recombined such that it matches some ideal
image, at least closely enough.
All
this, because the LLM is fundamentally a writing machine, it does
everything via text, and if you make it produce writing that exists purely to
serve some sort of mechanical function, and you train it to succeed in that
task, then it will tend to do so, even with vast intricacy. It still fails when
it comes to spatial relations within text, because everything is understood in
terms of relations and correspondences between tokens as values
themselves, and apparent spatial position is not a stored value. Space is the
field of difference, irreconcilable difference--if the letters, words on the
page were not separate, then how would they be analyzed into their particular
relations? Kant had already demonstrated this schematic form hundreds of years
ago in the Critique of Pure Reason, which no current LLM could ever reproduce
by themselves, because the book is so dependent on the specific spatial
arrangements of the text, the various tabular organizations, the antinomies
themselves which are always cut off from one another, a chasm of difference
sits between[2].
And
if an LLM cannot reliably reproduce a book like the first critique without some
outside intervention, then someone will always find something it can't
yet do, requires some tooling, some re-working to get done, some function that
it is currently incapable of. They'll always need someone to take a look under
the hood, figure out how their machine ticks. A strong, fearless
individual, the spanner in the works, the eddy in the stream! There, sitting in
my room, looking over prompts, waiting for someone to shoot me a message,
asking if they can share their screen.
(Claude Sonnet 4.5)
[1]Cf. Hui, Yuk. Machine and Sovereignty :
For a Planetary Thinking. First edition., University of Minnesota Press,
2024. Specifically chapter 4, "Mechanism, Organism, or
Decisicionism."
[2] A contemporary example of a book that employs
spatial relations within the text as a fundamental part of its form is House
of Leaves by Danielewski which, if not a literary masterpiece, at least
succeeds in its concept.
Comments
Post a Comment