So when I was a kid, there was this guy called

Uri Geller. He claimed to be a psychic. He was

a big deal. He was on the John Carson show...

Johnny Carson... So like, imagine a show like Jimmy

Fallon

and Jimmy Kimmel and Colbert's audiences all

roll up into one. Uri was, of course, a fraud

because psychic powers aren't a thing, but he

fooled the CIA, professors, scientists,

parapsychologists, physicists, et cetera, et cetera.

He was exposed by a guy named James Randi,

who wrote this video's featured book. Randi

wasn't a scientist. He was a stage magician.

Turns out that the secret to spotting a fraud

isn't asking the experts in the field in

question.

It's asking someone he knows how to reproduce

the result. So let's do a thought experiment

about ChatGPT. So imagine for a moment that

you wanted to build an online service that mimics

what ChatGPT does. For now, just worry about

the text part. Let's not worry about Dall-E or

Sora

or any of that stuff, but we want to do it

without using any neural networks or anything

that people

might normally consider to be AI. This has to

remain a thought experiment for reasons I'll

talk

about later, but let's go through it together.

It'll be an interesting exercise. So you want

to

build a conventional internet software startup

and you've got a handful of years to get a

product to

market. You don't have to care about revenue

and you have, I'm just going to pick a number

out of

the air, say $63 billion, $920 million of

investment money to spend on it. Do you think

that could be

done? Because I do. And I think there's a

decent chance there'd be enough money left over

to feed

all of the hungry people in the US for a decent

fraction of a year. So you start off by

building a

giant index of the web and, what the hell, while

you're at it throw in a few terabytes of pirated

books and journals. Now, we're not talking

about training a model or anything like that.

This is

straightforward text indexing the way Google

used to work back before they killed it. Now

you take

the query text, we're not going to call these

prompts although there would be a lot of

similarity.

And you use the standard text search algorithms

to find the top N similar matches. Might be

100

might be 1000, who knows. We would run experiment

later to see what number gives us the best

results.

Then we take the text that matched and we dump

it into some series of algorithms to extract

the

text, break it up in its parts of speech, find

permutations of it, move words around, replace

words

with synonyms, et cetera, and then decide what

word we ought to output next. There's a lot of

prior

art we could pull from. The act of outputting

similar text goes back at least to Eliza in the

1960s,

then there was a chatbot made by a few people

for a few thousand dollars. It supposedly

succeeded

in passing a Turing test in a competition in

2014. Siri reportedly cost Apple 200

million dollars

and we've got, what, 300 times that much money?

IBM supposedly spent four billion dollars on

its

Jeopardy winning Watson bot. And that wasn't

just to compete on the show. That also includes

all the

money they spent trying and failing to cure

cancer with it after it won Jeopardy. We have

16 times

more money than that. And unlike Watson or Siri

or the Turing test bot, we don't have to care

if it

answers the questions correctly. We'll just

stick a disclaimer at the bottom of the output

page that

says the bot hallucinates and the user is

responsible for verifying everything. How close

do you think

we could get to what ChatGPT could do? It's

hard to say because it could never ever happen,

both because no one would spend that much money

making a chatbot without being convinced that

it

might magically become an equivalent of a PhD

in every subject. And because we get sued into

homelessness by all of the people whose content

we were plagiarizing because we wouldn't have

the

defense of claiming that we "grew" it instead of

building it, and we can't be held responsible

because

no one can understand how it works. But I've

done a lot of data searching projects. And

given that

amount of money, I don't have a difficult time

believing we could get there. And I bet we

could

make it even easier by taking what has been

learned about the attention mechanism since the

2017

"attention is all you need" paper. And then retrofitting

that knowledge back onto non-neural

network vector encoding techniques like latent

semantic analysis or hyperspace analogs to

language.

And what would make it easiest of all is that

the hardest part of any such project,

making sure the answer is correct, no longer

applies, because we can just stick our fingers

in our ears and chant "It's just a hallucination!!!"

over and over anytime somebody complains.

At the very least, I think it'll be much more

technically possible that ChatGPT being PhD

level

smart in every subject. Just something to think

about. Thanks for watching. Let's be careful

out there.