More accurately:
>looks inside
>matrix multiplication
post
You don't have to be an AI developer to not understand your code. I not understand my code every day.
It’s just y = k*x + m a few trillion times.
Yes it's not even for loop and if. Just multiplication and addition.
You're forgetting the activation function to break linearity, otherwise you can only model linear trends. Many activation functions have "if" hidden in them (e.g. ReLU with the "max").
The human brain is far too complex to understand
looks inside
it's just meat
The brain isn't meat, it's fat, and is known to turn to soap under the right conditions. Which coincidentally also explains the soapy tiddy memes.
That was old AI.
New AI uses a complex, multi-layer sequence of Magic 8-balls.
That's a common misconception. It is a complex multi-layer sequence of heavily biased Magic 8-balls.
I wouldn't be surprised if some stuff (as in architecture) was solved by querying a different model to solve a problem
"Mirror mirror on the wall please fix my awful code, make no mistakes"
I would be heavily more inclined to use LLMs (and also probably use it in a more correct manner) if it came in the physical form of a magic 8-ball that I would violently shake to get the answer.
Have you ever tried to follow five hundred nested for loop and if statements? Shit is nuts...
Oh, just wait till you hit that sweet sweet python "Everything can have an else block"
if some_shit:
try:
for stuff in thing:
if stuff is bad:
raise ShitHitTheFanError
# Massive block comment
do = the_thing()
if do is good:
break
else:
do = the_other_thing()
if not a_thing:
help_me() # May raise NoHelpFoundError
except ShitHitTheFanError:
my_life = recover()
if my_life is bad:
for fuck in my_life:
try:
fix(fuck)
except NoHelpFoundError:
break
else:
fuck_this_shit()
else:
wat_is_going_on()
else:
theres_another_one("?")
Conway's Game of Life is Turing Complete and that fact has been fucking with my mind for like 25 years.
If statements are some of the largest sources of bugs in software.
ml is not a hard coded set of if-else pairs. it's a glorified fit with a lot of parameters.
And the human brain is just fatty tissue and neurons.
Cool; walk me through attention layers plz.
Yeah. It's some if statements and some loops but with CUDA and matrix multiplication
Great; that describes all functions, so we've nothing to worry about.
No no we still gotta worry about people in power allowing the worst possible use of technology at everyone's expense for personal gain..
So same old
What specifically do you want the people in power to do?
A very good start would be people not having that much power to begin with. Especially not if they’re not democratically elected representatives.
But for the people in power we’re seemingly talking about (AI company leaders):
- Not presenting AI as something mystical/mythical
- Not turning AI into a product tailored to be addictive
- Not framing AI as a tech only they can be trusted to keep and distribute (see also point 1)
- Maybe developing the technology towards what gives most utility, not what earns them the most money (fat chance)
I just want these AI companies to be held accountable for openly trying to hack everyone. It's ridiculous they're allowed to hide behind "AI did it, not us" as if what their product does isn't their responsibility.
Think of it as a game, the players are corpo, and gov.
Corpo hacks planet, gov says, "Hm, that's pretty bad corpo, we could hold you accountable.. but.. that's pretty neat, maybe share a lil with us how you achieved this terrible incident (that we totally don't want to leverage against international peers) and we can forget this happened.."
A game of incentives.
If the gov comes down too hard, they could slow the development, the same development that they short-sightedly want to exploit in parallel with the accused.
Regulation? I mean is this really an honest question? Are you really that unfamiliar with what regulation is?
Here you go: Let's build GPT: from scratch, in code, spelled out.
GPT is basically a machine that builds machines.
We understand how the machine builder works. We don't understand how the output of the machine works. Like mechanically we understand each individual sub-component works e.g. a single matrix multiplication, but we don't understand why specific weights are the value they are, or the broader macrostructure.
I don't think you realize that acting like "No we actually do understand how it works" just reveals that you don't understand the question, it does not show that the question has an answer.
Please explain what this one specific weight is, and why that value shouldn't be higher or lower.
It's just a series of 10 trillion floats. And we don't really know what the individual floats represent and changing them degrades performance but hey
Technically. It's just if statements.
"It's like Blade Runner, but we'll make sure the robots aren't the villains this time" -- some billionaire, probably
not really it is mostly automatic gradients on billion parameters and shit loads of regularization.
It's not really cold though is it. It's just a lot of numbers and those numbers somehow do things. Like it says no one understands it.
It was better when it was just ifs and for loops and runs on potatoes and does actually useful things.



all 40 comments