April 10, 2024 • 30 mins
It’s been just under a year and a half since ChatGPT - an AI-powered chatbot launched by so-called non-profit OpenAI - ushered in a new era of investor and media hype around how artificial intelligence would change the world. But what if this we're actually at the peak of what generative AI can do? In this episode, Ed Zitron walks you through the four intractable problems that are stopping Large Language Models like ChatGPT in their tracks - and why they're all-but-impossible to overcome.
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:02):
Fall Zone Media.
Speaker 2 (00:05):
Hello and welcome to Better Offline. I'm your host ed
z tron. It's been just under a year and a
half since chat gpt, an AI powered chatbot launched by
(00:26):
so called nonprofit open Ai, ushered in a new investor
in media hype cycle around how AI would change the world.
Chat GPT's instant success made both open Ai and its CEO,
Sam Mortman overnight celebrities. As a result of chat GPT's
alleged intelligence, which seemingly allowed you to do everything from
generate an entire essay from a simple prompt to writing
(00:48):
entire reams of software code. You can theoretically ask it
anything and it will spit out an intelligent sounding response
thanks to being trained on terabides of text data, like
a search engine that's able to think for itself. Ah.
Big problem is chat gpt doesn't think at all. It
doesn't know anything. Chat gpt is actually probabilistic. It uses
(01:09):
a statistical model to generate the next piece of information
in a sequence. If you ask it what an elephant is,
it'll guess that the most likely answer to that prompt
is that an elephant is a large mammal, and then
perhaps describe its features, such as a long trunk. Chat
GPT doesn't know what an elephant is, or what a
trunk is, or what the alefant Haantai family is. It's
simply ingested enough information to reliably guess what an elephant
(01:32):
is meant to be, or indeed know that that's what
you're asking it. This is the technology underpinning the latest
artificial intelligence boom. It's called generative artificial intelligence, and it's
powered by large language models, and they underpin tools like
open ais, chat, GPT, Anthropics, claude x dot COM's horrifying
(01:52):
chatbot Grock, and of course Google's Gemini. Essentially, they're AI
systems that ingest vast quantities of written text or other data,
and then, through mathematics, try and identify patterns of relationships
between words or symbols, or basically any meaning from the
text or thing they're being fed. And it almost seems
like magic because it's able to generate this plausible, seeming
(02:15):
almost human content at this remarkable speed. These models are
now capable of generating text, images, and even video in
response to simple chat prompts or by learning the patterns
and structures of their data. Yet underneath the hood, there's
always something a bit wrong. Generative AI, at times authoritatively
spits out incorrect information, which can range from funny like
(02:37):
telling you that you can melt an egg, to outright dangerous,
like when the recently launched AI powered New York City
chatbot for small business owners started telling them that it
was legal to fire somebody for refusing to cut their dreadlocks.
This is why you'll see strange glitches in images generated
by AI hands with too many fingers, horrifying looking people
in the back of realistic looking photos, and so on
(02:59):
and so far. Because these models don't actually know what
anything is. They don't have meaning, they don't have consciousness
or intelligence. They're guessing, and when they guess, they sometimes hallucinate,
which I'll get too soon. And while they might be
really really good at guessing, there are effectively a very
(03:22):
very powerful version of auto complete. I don't know anything.
I really mean that these things aren't even intelligent. But
because these models seem like they know stuff, and they
seem to be able to do stuff, and the things
that they create almost seem right. The media and the
Vocal Investor class on Twitter have declared that large language
models would change everything to them. Llms like chat GPT
(03:44):
would upend entire business models, render once unassailable tech giants nunerable,
and rewrite our entire economic playbook by turning entire industries
into something you tell a chatbot to do. In a sentence,
you know, it doesn't really matter that genera ive AA
is mostly good at reams of generic slop, and that
it's also clogging services like Amazon's Kindle eBookstore. And I
(04:05):
guess the rest of the Internet with generative content that
doesn't matter at all, because it's kind of good. Obviously,
I'm being sarcastic. This is all very, very bad. In
the last year, there have been hundreds of muling articles
about how AI will replace everything from drive through workers
to medical professionals. Theoretically, you could just feed whatever information
(04:27):
a potential customer could ask for into a vast database
and have an AI chew it up, and then they
could just generate exactly the answer you'd need. AI would
just naturally slip into areas of disorganization and inefficiency and
spit out remarkable new ideas, all with minimum human input.
In every one of these stories carries with them a
(04:48):
shared belief one might even call it a shared hallucination.
And they all believe that generative AI will actually be
able to do these things, that it will actually be
able to replace people. And what we're seeing today is
just the beginning of our glorious automated future. But what
if it's not. What if generative AI can't actually do
(05:09):
much more than it can today? What if we're actually
at peak AI. In the next two episodes, I'm going
to tell you why I think that is. I'm going
to tell you how I think this whole thing falls apart.
(05:33):
AI's media hype train has been fairly relentless since November
twenty twenty two, when chat GBT launched, and AI champions
like open Ai CEO Sam Mortman have proven all too
willing to grease its wheels, making bold promises about what
AI could do and how today's problems are so easily overcome.
It's also helped that the tech media has largely accepted
(05:53):
these promises without asking basic questions like how and when
will it do this stuff? And can it do this stuff?
Don't believe me, Go and look at any interview with
Sam Moltman from the last few years, watch any of them.
In fact, just look at any AI figurehead getting interviewed
and count the amount of times they've actually received any
pushback or been asked to elaborate on any specific issue.
(06:17):
It's actually very rare. Let me play you one of
the few times that anyone's actually interrogated an AI person.
Specifically Joanna Stern of The Wall Street Journal, who you
might remember from the Vision Pro episode A Better Offline.
She interviewed open AI's Chief Technology Officer, Mirror Marathi about Sora,
which is open aiy's video based version of chat GPT
(06:37):
where you can theoretically ask it to generate videos. Just
to be clear, it's unreleased and unclear whether it'll ever
actually get released, and the videos look good at first,
then they look really weird. But just listen to this
particular question. It's Joanna asking Mirror, the CTO of open
ai and eighty billion dollar AI company, Hey, did you
train on YouTube? What data was.
Speaker 3 (06:59):
Used to train?
Speaker 1 (07:01):
We used publicly available data and licensed data, so videos
on YouTube?
Speaker 2 (07:09):
Now, I encourage you to go and look up this
clip because at this point Maurti makes the strangest face
I've ever seen in a tech interview.
Speaker 1 (07:19):
I'm actually not.
Speaker 3 (07:19):
Sure about that okay, videos from Facebook, Instagram.
Speaker 1 (07:26):
You know, if they were publicly available available, publicly available
to use, there might be the data, but I'm not sure.
I'm not confident about it.
Speaker 3 (07:39):
What about Shutterstock, I know you guys have a deal
with them.
Speaker 1 (07:44):
I'm just not going to go into the details of
the data that was that was used, but it was
publicly available or licensed data.
Speaker 2 (07:52):
The remarkable part about this interview is that it's a
relatively simple question. You, as the CTO of an eighty
billion dollar AI company, what training data did you use
to train your model? Did you use YouTube? So yes
or no? Question? Mirror mirror, answer the bloody question, mirror
all right, right? The answer is, of course Open Ai
(08:14):
likely trained it's video generating model, Sora on YouTube videos,
which might be why they're yet to launch it. And
the videos generated by Soora also feature some remarkably similar
images to say, SpongeBob SquarePants, and I wouldn't be surprised
if they carry with them multiple weird biases about race
and gender that we'll see in the future. But also
(08:37):
when you watch these videos, much like most generative AI content,
there's something a bit off about them. In The Wall
Street Journal's interview. You get to see some of the
prompts that were used and some of the videos that
came out, and you see crazy things happening, like a
robot completely changing shape as it turns, cars disappearing and
appearing behind the robot. It's not very good. It seems
(09:00):
cool at first. If you squint really hard, it looks real,
but there's always something off. And that's because, as I've
said before, these models don't know anything and don't know
what at robot looks. They can make a really good
guess though. Anyway, Sterne's interview with Marati of Open AI
is a great example of how the entire AI artifst
falls apart at the slightest touch, because it's fundamentally flawed
(09:23):
and not actually able to deliver the society defining promises
that Sam Morltman and the venture capital sect would have
you believe. In a year and a half, despite billions
of dollars of investment, despite every major media outlet claiming otherwise,
generative artificial intelligence has proven itself incapable of replacing or
even meaningfully enhancing human work. And the thing is, all
(09:46):
of these problems I'm talking about with generative AI all
of these hallucinations, all of these weird artifacts that are
popping up throughout these videos, the weird mistakes that the
texts that are popped out by chat GPT have, all
of these. The problems are problems that aren't necessarily just technological.
Their physics, they're mathematics. These aren't things you can just outrun.
(10:11):
And I believe that there are four intractable problems that
will stop generative AI from progressing much further than it
is today. The first is, of course, its energy demands,
the massive amounts of power requires. The second are its
computational demands, the amount of compute power it requires to
even crunch the simplest things out of chat GPT, It's hallucinations,
(10:34):
the authoritative failures it makes when it spits out nonsense
or creates a human hand with eighteen fingers, and of
course the fact that these large language models have an
insatiable hunger for more training data. Now, let me break
that down. Large language models are extremely technologically and environmentally demanding.
(10:55):
The New York Are reported in March twenty twenty four
the CHATGBT uses more than half a million kis what
hours of electricity to respond to the two hundred million
requests it receives in a day, or seventeen thousand times
the amount that the average American household uses in a day,
and others have suggested it might be as high as
thirty three thousand households worth. Generative AI models demand specialist
(11:18):
chips called graphics processing units, typically a souped up version
of the technology used to drive the graphics in a
gaming console, albeit at a much higher cost, with each
one costing tens of thousands of dollars each. They do
this because large language models like chat GPT are highly
computationally intensive. I'm going to break that down. Don't worry.
(11:38):
When you ask chat GPT a question, it tokenizes it,
breaking it down into smaller parts for the model to understand.
It then feeds these tokens into various mechanisms that help
it understand the meaning of the thing you asked it
to do based on the parameters that it learned in training.
Chat GPT generates a response by predicting the most likely
(11:58):
sequence of things that you might want it to do.
An answer to a question, an image, so on, and
so forth. Each one of these steps is extremely demanding,
processing hundreds of billions of these parameters learned patterns from
ingesting training data such as how the English language works
or what a dog looks like to produce even the
simplest thing. Training these models is equally intensive, requiring chat
(12:21):
GPT to process massive amounts of data. Another problem I'll
get to in a bit adjusting those hundreds of billions
of parameters and developing new ones based on what the
data says as it quote unquote learns more. Though as
we're clear, chap GPT doesn't learn anything. It just makes
new parameters to read things. A model like chat GBT grows,
making it more complex, which in turn requires more data
(12:44):
to train on and more compute power to both ingest
the data, create more parameters and turn it into something
resembling an answer, And because it doesn't know anything, it's
suggesting the most likely to be correct answer, which leads
it to hallucinating and correct things that, based on proper ability,
kind of seem like the right thing to say. These
hallucinations are the dirty little secret of generative AI, and
(13:07):
are impossible to avoid thanks to the fact that every
single thing these models say is a mathematical equation rather
than any kind of intellectual exercise. If you ask CHATGPT
how many days there are in a week. It doesn't
know that there are seven days, but it's been trained
on patterns of language and generates a result based on
those patterns, which at times can be correct and can
(13:27):
also be wrong. There's no way of fixing this problem.
You can mitigate it, you can make it less likely
it will mess up, but hallucinations will happen because there
is no consciousness. It is not learning anything. This thing
has no knowledge. More computing power would allow it more
(13:48):
parameters to give it more rules, so that a generative
AI will be more likely to give a correct answer,
but there's no eliminating them, and doing so may require
more computing power than actually exists or is possible without
an AI of consciousness, an impossible dream known as average
generalized intelligence that Sam Mortman would have you believe is imminent.
(14:08):
There's really no solving hallucinations. When you answer questions using probability,
you're always going to have mistakes because you're not actually
answering them using knowledge, intellect, or experience. You're using dice rolls.
It's a bloody game of dungeons and dragons. We turned
in Carter into dungeons and dragons anyway. A newly published
(14:30):
paper by Tepo Felon and Matthias Holweg of the University
of Oxford agrees funding that large language models like chat
GPT are incapable of generating new knowledge. It's a remarkably
in depth rundown of the fundamental differences between a large
language model and a human brain, and it combines both
psychological and mathematical research going back to child psychology as
(14:51):
well the basic building blocks of how we consume and
learn things and how we make decisions. As a result,
the paper title theory is or You Need AI Human
Cognition and Decision Making argues that AI's data and prediction
based orientation is an incomplete view of human cognition and
the forward thinking theorizing of the human mind. In Layman's terms,
(15:13):
the mess of the information we've learned over our lives,
our experiences, and our ability to look forward and consider
the future is just fundamentally different to a model that
predicts things only based off of past data. Think of
it like this, If you've read a book and you
might think about writing a new book based on those ideas,
you're not remembering every part of the book. You don't
(15:36):
have a perfect memory, and you're also constantly thinking about
things as your day goes on. The human brain is
a goddamn mess. Generative AI is in some level stuck
in amber. Though the billions of parameters might change, the
data never does the way it consumes the data. May
be that the data doesn't change. In essence, generative AI
(15:58):
is held back by the fact that it can't consider
the future and is actually permanently mired in the data
of the past. Their largest problem might be a fast,
simpler one, a farcillio, or a kind of an ironic one.
There might not be enough data for these bloody things
to actually train on. While the Internet may at times
(16:31):
feel limitless, a researcher recently told The Wall Street Journal
that only a tenth of the most commonly used web
data set, the common Crawl, are freely available. Two hundred
and fifty billion page dump of the web's information is
actually of high enough quality data for large language models
like CHATGBT to actually train on. Putting aside the fact
that I can't find a single definition of what high
(16:52):
quality actually means, the researcher pat Blow Vilobos suggested that
the next version of chat GBT would require more than
five times the amount of data it took to train
in its previous version, GPT four. The new one is
called GPT five. By the way, and other researchers have
suggested that AI companies are going to run out of
training data in the next two years. Now. That sounds dire,
(17:18):
but don't worry. They've come up with a very funny
and extremely stupid idea to fix it. One specifically posed
by The Wall Street Journal is that the AR companies
are going to create their own synthetic data to train
their models, a computer science version of inbreeding. The researcher
Jason Stadowski calls habsburg AI. This is, of course, an
(17:38):
absolutely terrible idea. A research paper from last year found
that feeding model generated data into models to train them
creates something called model collapse, a degenerative learning process where
models start forgetting improbable events over time as the model
becomes poison with its own projection of reality. The paper,
called the Curse of recursion. Training on generated data makes
(18:00):
models forgam highlights an example where feeding a generative AI
its own data eventually destroys its ability to answer questions,
and within nine generations. One answered a simple prompt about
architecture with an insane screed about jack rabbits full of
at symbols and weird characters. So not to worry again.
The tech overlords have come up with a great idea
(18:22):
to fix this problem. Their common retort to the problem
of synthetic data is that you could use another generative
AI to monitor the synthetic data being fed into all
model to make sure it's right. At this point, I'd
like to get slightly angry. Are you kidding me? Are
you fucking kidding me? You're saying that the way to
make sure the data generated by an unreliable generative AI
(18:45):
is to use another generative AI, one with the same
goddamn problems, which also hallucinates information that knows nothing. You're
going to use that AI to monitor whether the data
that is created by an AI is any good? Are
you completely insane? You insane? You're going to feed the
crap from the crap machine into another crap machine to
make it not make crap? Why am I reading journalists
(19:07):
credulously printing this ridiculous solution in the New York Goddamn
Times every time? Every time, these bubbles are inflated because
tech executives are able to get their half ass, half
baked solutions parroted by reporters who should know better. You
don't have to give them the better affair of the
goddamn fucking doubt. This is how we got the bloody
meta us. Pardon me, I've calmed down now. Anyway. Anyway,
(19:31):
if you're worried about model collapse, you're already too late,
as these models are likely already being fed their own data.
You see, these models are trained on the web, as
I previously told you, and they're desperate. They need data.
They need more stuff. They need more stuff to ingest
so they can spit out more stuff. The problem is
that these machines are purpose built to make a lot
(19:54):
of content, and so the Web's already being filled with
generative AI. Generative AI is already spamming the Internet. A
report from four oh four Media from last week said
that Google Books has already started to index several different
works that were potentially written by AI, featuring the hallmark
(20:14):
generic writing tropes of these models. Four or four Media
also reports that the same thing is happening over at
Google Scholar their index of scientific papers with one hundred
and fifteen different articles featuring the phrase as of my
last knowledge update a specific phrase spat out by generative models.
This is really bad, by the way, and this is
(20:34):
only going to get worse. When you have an Internet
economy that is built so that the people that can
put the most out there will probably get the most traffic.
They're going to use these tools. These tools are great
for that. If you don't give a rap fuck about
the quality, this is the best thing in the world
for you. And that's the thing. This is a problem
(20:54):
both created and caused by these models. You see. The
other dirty little secret of generative II is that these
models unashamedly plagiarize the entire web, leading outlets like The
New York Times and authors like John Grisham to sue
open ai for plagiarism. While open ai won't reveal exactly
what their training data is, the New York Times is
able to successfully make chat gpt reproduce content from the newspaper,
(21:18):
and the company has repeatedly said that it trains on
publicly available data from the Internet, which will naturally include
things like Google scholar and Google Books. The Times also
reports that open ai has become so desperate for data
that they've used their whisper tool to transcribe YouTube videos
into texts to feed into chat GPT's training data. Pretty sure,
(21:39):
that's plagiarism, but who am I to tell you? And
as the web gets increasingly pumped full of this generative content,
these models are going to just start eating their own swill,
slowly corrupting themselves in a kind of ironic death. According
to Zakhar Schumelov, one of the authors of the model
collapsed paper at the University of Cambridge, the unique problem
(22:01):
that synthetic data creates is that it lacks human errors.
Human made training data, by the nature of it being
written by a human, includes errors and imperfections, and models
need to be robust to such errors. So what do
we do if models are trained off of content created
without them? Do we introduce the errors ourselves? How many
errors are there? How do we can introduce them more?
(22:22):
And indeed, what are the errors? What do they look like?
Do we even know? Are we conscious of the errors
in the human language that make us human? The models aren't? Well,
maybe they are. It's kind of unclear. It's kind of
tough to express how deeply dangerous the synthetic data idea
is for AI models like chat, GPT and Claude are
(22:45):
deeply dependent on training data to improve their outputs, and
their very existence is actively impeding the creation of the
very thing they need to survive. While publishers like Axel
Springer have cut deals to license their companies data to
chat GPT for training purposes, this money is flowing to
the writers that create the content that open Ai and
Anthropic need to grow their models much further. In fact,
(23:07):
I don't think you're going to see more journalists get
hired as a result of these deals, which kind of
makes them a little bit stupid. This puts AI companies
in a kind of Kafka esque bind, but they can't
really improve a tool for automating the creation of content
without human beings creating more content than they've ever created before,
(23:28):
just as said tool actively crowds out human made data.
It's a little silly. The solution to these problems, if
you ask open AI's Sam Altman, is always more money
and power, which is why the information reports he is
trying to convince open ai investor Microsoft to build him
and I'm not kidding, and one hundred billion dollars supercomputer
(23:49):
called Stargate. This massive series of interconnected machines will require
entirely new ways to mountain cool processing units and is
entirely contingent on pen AI's ability to meaningfully improve chet GPT,
something Sam Mortman claims isn't possible without more computing power.
(24:09):
To be clear, open Air already failed to build a
more efficient model, dubbed Arakis, which ended up getting mothballed
because it wasn't more efficient. It's also important to note
that every major cloud company now has inextricably tied themselves
to the generative AI movement. Google and Amazon have invested
(24:32):
billions into chat GPT competitor Anthropic, and both claim to
be Anthropic's primary cloud provider, though isn't really obvious which
one is. In doing so, they've guaranteed, according to a
source of mind, about seven hundred and fifty million dollars
a year of revenue for Google's cloud and eight hundred
million dollars a year of revenue for Amazon Web Services
(24:53):
the cloud service from Amazon by mandating that Anthropic uses
their services to power their clawed model. This similar to
the thirteen billion dollar investment that Microsoft gave open AI
last year, most of which was made up of credits
for Microsoft's cy or cloud, And I somehow doubt that
Microsoft is going to be the noble party that goes
on the earnings and says, well, we don't want to
(25:15):
count the credits that we gave open out want to
be fair. No, they're going to mash that shit right
back into their revenue. Kind of a con kind of
makes me angry when I think about it too. Anyway,
let me just put that aside and not going to
get pissed off again. Look, I'm surprised more people aren't
really upset about this very incestuous relationship between big tech
and this supposedly independent generative AI movement. Microsoft, Google, and
(25:40):
Amazon have effectively handed cash to one or two companies
that will eventually hand the cash back to them in
exchange for cloud services that are necessary to make their
companies work, and all three big tech firms are spending
billions to expand their data center operations to capture this
theoretical demand from generative AI. Every penny the open AI
(26:02):
or anthropic makes will now flow back to one of
three big tech firms, even more so in the case
of open AI, because Microsoft's investment entitles Microsoft to a
share of any future profits from open ai and chat GBT. Yeah,
it doesn't even really matter if they make one, because
Big tech wins either way. Anthropic has to use Google
(26:23):
Cloud and Amazon Web services, open Ai has to use
Microsoft's z your Cloud, and Microsoft is actively selling open
AI's models to their zero cloud customers. And every time
somebody uses open AI's models, that model is being run
on a zero cloud, generating revenue for Microsoft. This is
the rot economy in action, by the way, Big tech
(26:45):
has funded its biggest customers for their next growth revenue stream,
justifying this massive expansion of their data center operations because
ai is the future and they're telegraphing growth to these
brainless drones in the market who will buy any thing,
who never think too hard about what they're actually investing in. Hey,
i's this big, sexy, exciting and theoretically powerful way to
(27:08):
centralized labor. And it's innovative sounding enough that it allows
people to dream big about how it might change their lives. Now,
it might help them not pay real people to do shit. Yeah,
here's the biggest worry I have. Here's the real pickle,
here's the thing that keeps me up at night. None
of these companies seem to have appeared to consider something.
(27:31):
What if generative AI can't actually do any of the
things they're excited about. What if genera avii's content, as
you probably seen from anything chet GBT spits out, isn't
really good enough. Hey, is anyone checked if anyone's actually
using these tools, if they're helpful to anyone? Is this
(27:53):
actually replacing anyone's work? Huh, that's a bit worrying, mate,
I didn't think about that before. Just kidding, I've been
thinking about it for months. Look, here's the thing. I
think that the big problem here is that Sam Altman
and his cronies have allowed the media, the markets, and
(28:13):
big tech to fill in the gaps of their specious messaging.
They've allowed everybody to think that open AI can do
whatever anyone dreams. Yeah, I don't think that generative AI
can do much more than it is today. And also,
from what I've seen, none of these generative AI companies
(28:34):
actually make a profit, and with each new model they
become less profitable, and I don't see that changing in
the future. And so I've dug in a little more
looking under the hood. All the demand that's spurring Microsoft, Google,
and Amazon's data center operations might not actually be there.
(28:57):
My friends, I think we're in the next tech bubble,
and the next episode, I'm gonna walk you through how
I think it might pop. Thank you for listening to
Better Offline.
Speaker 4 (29:15):
The editor and composer of the Better Offline theme song
is Matasowski. You can check out more of his music
and audio projects at Matasowski dot com, M A T
T O, S O W s ki dot com. You
can email me at easy at Better offline dot com,
or check out Better Offline dot com to find my
newsletter and more links to this podcast. Thank you so
(29:35):
much for listening.
Speaker 2 (29:37):
Better Offline is a production of cool Zone Media.
Speaker 4 (29:40):
For more from cool Zone Media, visit our website cool
Zonemedia dot com, or check us out on the iHeartRadio
Speaker 3 (29:46):
App, Apple Podcasts, or wherever you get your podcasts.
Fall Zone Media.
Speaker 2 (00:05):
Hello and welcome to Better Offline. I'm your host ed
z tron. It's been just under a year and a
half since chat gpt, an AI powered chatbot launched by
(00:26):
so called nonprofit open Ai, ushered in a new investor
in media hype cycle around how AI would change the world.
Chat GPT's instant success made both open Ai and its CEO,
Sam Mortman overnight celebrities. As a result of chat GPT's
alleged intelligence, which seemingly allowed you to do everything from
generate an entire essay from a simple prompt to writing
(00:48):
entire reams of software code. You can theoretically ask it
anything and it will spit out an intelligent sounding response
thanks to being trained on terabides of text data, like
a search engine that's able to think for itself. Ah.
Big problem is chat gpt doesn't think at all. It
doesn't know anything. Chat gpt is actually probabilistic. It uses
(01:09):
a statistical model to generate the next piece of information
in a sequence. If you ask it what an elephant is,
it'll guess that the most likely answer to that prompt
is that an elephant is a large mammal, and then
perhaps describe its features, such as a long trunk. Chat
GPT doesn't know what an elephant is, or what a
trunk is, or what the alefant Haantai family is. It's
simply ingested enough information to reliably guess what an elephant
(01:32):
is meant to be, or indeed know that that's what
you're asking it. This is the technology underpinning the latest
artificial intelligence boom. It's called generative artificial intelligence, and it's
powered by large language models, and they underpin tools like
open ais, chat, GPT, Anthropics, claude x dot COM's horrifying
(01:52):
chatbot Grock, and of course Google's Gemini. Essentially, they're AI
systems that ingest vast quantities of written text or other data,
and then, through mathematics, try and identify patterns of relationships
between words or symbols, or basically any meaning from the
text or thing they're being fed. And it almost seems
like magic because it's able to generate this plausible, seeming
(02:15):
almost human content at this remarkable speed. These models are
now capable of generating text, images, and even video in
response to simple chat prompts or by learning the patterns
and structures of their data. Yet underneath the hood, there's
always something a bit wrong. Generative AI, at times authoritatively
spits out incorrect information, which can range from funny like
(02:37):
telling you that you can melt an egg, to outright dangerous,
like when the recently launched AI powered New York City
chatbot for small business owners started telling them that it
was legal to fire somebody for refusing to cut their dreadlocks.
This is why you'll see strange glitches in images generated
by AI hands with too many fingers, horrifying looking people
in the back of realistic looking photos, and so on
(02:59):
and so far. Because these models don't actually know what
anything is. They don't have meaning, they don't have consciousness
or intelligence. They're guessing, and when they guess, they sometimes hallucinate,
which I'll get too soon. And while they might be
really really good at guessing, there are effectively a very
(03:22):
very powerful version of auto complete. I don't know anything.
I really mean that these things aren't even intelligent. But
because these models seem like they know stuff, and they
seem to be able to do stuff, and the things
that they create almost seem right. The media and the
Vocal Investor class on Twitter have declared that large language
models would change everything to them. Llms like chat GPT
(03:44):
would upend entire business models, render once unassailable tech giants nunerable,
and rewrite our entire economic playbook by turning entire industries
into something you tell a chatbot to do. In a sentence,
you know, it doesn't really matter that genera ive AA
is mostly good at reams of generic slop, and that
it's also clogging services like Amazon's Kindle eBookstore. And I
(04:05):
guess the rest of the Internet with generative content that
doesn't matter at all, because it's kind of good. Obviously,
I'm being sarcastic. This is all very, very bad. In
the last year, there have been hundreds of muling articles
about how AI will replace everything from drive through workers
to medical professionals. Theoretically, you could just feed whatever information
(04:27):
a potential customer could ask for into a vast database
and have an AI chew it up, and then they
could just generate exactly the answer you'd need. AI would
just naturally slip into areas of disorganization and inefficiency and
spit out remarkable new ideas, all with minimum human input.
In every one of these stories carries with them a
(04:48):
shared belief one might even call it a shared hallucination.
And they all believe that generative AI will actually be
able to do these things, that it will actually be
able to replace people. And what we're seeing today is
just the beginning of our glorious automated future. But what
if it's not. What if generative AI can't actually do
(05:09):
much more than it can today? What if we're actually
at peak AI. In the next two episodes, I'm going
to tell you why I think that is. I'm going
to tell you how I think this whole thing falls apart.
(05:33):
AI's media hype train has been fairly relentless since November
twenty twenty two, when chat GBT launched, and AI champions
like open Ai CEO Sam Mortman have proven all too
willing to grease its wheels, making bold promises about what
AI could do and how today's problems are so easily overcome.
It's also helped that the tech media has largely accepted
(05:53):
these promises without asking basic questions like how and when
will it do this stuff? And can it do this stuff?
Don't believe me, Go and look at any interview with
Sam Moltman from the last few years, watch any of them.
In fact, just look at any AI figurehead getting interviewed
and count the amount of times they've actually received any
pushback or been asked to elaborate on any specific issue.
(06:17):
It's actually very rare. Let me play you one of
the few times that anyone's actually interrogated an AI person.
Specifically Joanna Stern of The Wall Street Journal, who you
might remember from the Vision Pro episode A Better Offline.
She interviewed open AI's Chief Technology Officer, Mirror Marathi about Sora,
which is open aiy's video based version of chat GPT
(06:37):
where you can theoretically ask it to generate videos. Just
to be clear, it's unreleased and unclear whether it'll ever
actually get released, and the videos look good at first,
then they look really weird. But just listen to this
particular question. It's Joanna asking Mirror, the CTO of open
ai and eighty billion dollar AI company, Hey, did you
train on YouTube? What data was.
Speaker 3 (06:59):
Used to train?
Speaker 1 (07:01):
We used publicly available data and licensed data, so videos
on YouTube?
Speaker 2 (07:09):
Now, I encourage you to go and look up this
clip because at this point Maurti makes the strangest face
I've ever seen in a tech interview.
Speaker 1 (07:19):
I'm actually not.
Speaker 3 (07:19):
Sure about that okay, videos from Facebook, Instagram.
Speaker 1 (07:26):
You know, if they were publicly available available, publicly available
to use, there might be the data, but I'm not sure.
I'm not confident about it.
Speaker 3 (07:39):
What about Shutterstock, I know you guys have a deal
with them.
Speaker 1 (07:44):
I'm just not going to go into the details of
the data that was that was used, but it was
publicly available or licensed data.
Speaker 2 (07:52):
The remarkable part about this interview is that it's a
relatively simple question. You, as the CTO of an eighty
billion dollar AI company, what training data did you use
to train your model? Did you use YouTube? So yes
or no? Question? Mirror mirror, answer the bloody question, mirror
all right, right? The answer is, of course Open Ai
(08:14):
likely trained it's video generating model, Sora on YouTube videos,
which might be why they're yet to launch it. And
the videos generated by Soora also feature some remarkably similar
images to say, SpongeBob SquarePants, and I wouldn't be surprised
if they carry with them multiple weird biases about race
and gender that we'll see in the future. But also
(08:37):
when you watch these videos, much like most generative AI content,
there's something a bit off about them. In The Wall
Street Journal's interview. You get to see some of the
prompts that were used and some of the videos that
came out, and you see crazy things happening, like a
robot completely changing shape as it turns, cars disappearing and
appearing behind the robot. It's not very good. It seems
(09:00):
cool at first. If you squint really hard, it looks real,
but there's always something off. And that's because, as I've
said before, these models don't know anything and don't know
what at robot looks. They can make a really good
guess though. Anyway, Sterne's interview with Marati of Open AI
is a great example of how the entire AI artifst
falls apart at the slightest touch, because it's fundamentally flawed
(09:23):
and not actually able to deliver the society defining promises
that Sam Morltman and the venture capital sect would have
you believe. In a year and a half, despite billions
of dollars of investment, despite every major media outlet claiming otherwise,
generative artificial intelligence has proven itself incapable of replacing or
even meaningfully enhancing human work. And the thing is, all
(09:46):
of these problems I'm talking about with generative AI all
of these hallucinations, all of these weird artifacts that are
popping up throughout these videos, the weird mistakes that the
texts that are popped out by chat GPT have, all
of these. The problems are problems that aren't necessarily just technological.
Their physics, they're mathematics. These aren't things you can just outrun.
(10:11):
And I believe that there are four intractable problems that
will stop generative AI from progressing much further than it
is today. The first is, of course, its energy demands,
the massive amounts of power requires. The second are its
computational demands, the amount of compute power it requires to
even crunch the simplest things out of chat GPT, It's hallucinations,
(10:34):
the authoritative failures it makes when it spits out nonsense
or creates a human hand with eighteen fingers, and of
course the fact that these large language models have an
insatiable hunger for more training data. Now, let me break
that down. Large language models are extremely technologically and environmentally demanding.
(10:55):
The New York Are reported in March twenty twenty four
the CHATGBT uses more than half a million kis what
hours of electricity to respond to the two hundred million
requests it receives in a day, or seventeen thousand times
the amount that the average American household uses in a day,
and others have suggested it might be as high as
thirty three thousand households worth. Generative AI models demand specialist
(11:18):
chips called graphics processing units, typically a souped up version
of the technology used to drive the graphics in a
gaming console, albeit at a much higher cost, with each
one costing tens of thousands of dollars each. They do
this because large language models like chat GPT are highly
computationally intensive. I'm going to break that down. Don't worry.
(11:38):
When you ask chat GPT a question, it tokenizes it,
breaking it down into smaller parts for the model to understand.
It then feeds these tokens into various mechanisms that help
it understand the meaning of the thing you asked it
to do based on the parameters that it learned in training.
Chat GPT generates a response by predicting the most likely
(11:58):
sequence of things that you might want it to do.
An answer to a question, an image, so on, and
so forth. Each one of these steps is extremely demanding,
processing hundreds of billions of these parameters learned patterns from
ingesting training data such as how the English language works
or what a dog looks like to produce even the
simplest thing. Training these models is equally intensive, requiring chat
(12:21):
GPT to process massive amounts of data. Another problem I'll
get to in a bit adjusting those hundreds of billions
of parameters and developing new ones based on what the
data says as it quote unquote learns more. Though as
we're clear, chap GPT doesn't learn anything. It just makes
new parameters to read things. A model like chat GBT grows,
making it more complex, which in turn requires more data
(12:44):
to train on and more compute power to both ingest
the data, create more parameters and turn it into something
resembling an answer, And because it doesn't know anything, it's
suggesting the most likely to be correct answer, which leads
it to hallucinating and correct things that, based on proper ability,
kind of seem like the right thing to say. These
hallucinations are the dirty little secret of generative AI, and
(13:07):
are impossible to avoid thanks to the fact that every
single thing these models say is a mathematical equation rather
than any kind of intellectual exercise. If you ask CHATGPT
how many days there are in a week. It doesn't
know that there are seven days, but it's been trained
on patterns of language and generates a result based on
those patterns, which at times can be correct and can
(13:27):
also be wrong. There's no way of fixing this problem.
You can mitigate it, you can make it less likely
it will mess up, but hallucinations will happen because there
is no consciousness. It is not learning anything. This thing
has no knowledge. More computing power would allow it more
(13:48):
parameters to give it more rules, so that a generative
AI will be more likely to give a correct answer,
but there's no eliminating them, and doing so may require
more computing power than actually exists or is possible without
an AI of consciousness, an impossible dream known as average
generalized intelligence that Sam Mortman would have you believe is imminent.
(14:08):
There's really no solving hallucinations. When you answer questions using probability,
you're always going to have mistakes because you're not actually
answering them using knowledge, intellect, or experience. You're using dice rolls.
It's a bloody game of dungeons and dragons. We turned
in Carter into dungeons and dragons anyway. A newly published
(14:30):
paper by Tepo Felon and Matthias Holweg of the University
of Oxford agrees funding that large language models like chat
GPT are incapable of generating new knowledge. It's a remarkably
in depth rundown of the fundamental differences between a large
language model and a human brain, and it combines both
psychological and mathematical research going back to child psychology as
(14:51):
well the basic building blocks of how we consume and
learn things and how we make decisions. As a result,
the paper title theory is or You Need AI Human
Cognition and Decision Making argues that AI's data and prediction
based orientation is an incomplete view of human cognition and
the forward thinking theorizing of the human mind. In Layman's terms,
(15:13):
the mess of the information we've learned over our lives,
our experiences, and our ability to look forward and consider
the future is just fundamentally different to a model that
predicts things only based off of past data. Think of
it like this, If you've read a book and you
might think about writing a new book based on those ideas,
you're not remembering every part of the book. You don't
(15:36):
have a perfect memory, and you're also constantly thinking about
things as your day goes on. The human brain is
a goddamn mess. Generative AI is in some level stuck
in amber. Though the billions of parameters might change, the
data never does the way it consumes the data. May
be that the data doesn't change. In essence, generative AI
(15:58):
is held back by the fact that it can't consider
the future and is actually permanently mired in the data
of the past. Their largest problem might be a fast,
simpler one, a farcillio, or a kind of an ironic one.
There might not be enough data for these bloody things
to actually train on. While the Internet may at times
(16:31):
feel limitless, a researcher recently told The Wall Street Journal
that only a tenth of the most commonly used web
data set, the common Crawl, are freely available. Two hundred
and fifty billion page dump of the web's information is
actually of high enough quality data for large language models
like CHATGBT to actually train on. Putting aside the fact
that I can't find a single definition of what high
(16:52):
quality actually means, the researcher pat Blow Vilobos suggested that
the next version of chat GBT would require more than
five times the amount of data it took to train
in its previous version, GPT four. The new one is
called GPT five. By the way, and other researchers have
suggested that AI companies are going to run out of
training data in the next two years. Now. That sounds dire,
(17:18):
but don't worry. They've come up with a very funny
and extremely stupid idea to fix it. One specifically posed
by The Wall Street Journal is that the AR companies
are going to create their own synthetic data to train
their models, a computer science version of inbreeding. The researcher
Jason Stadowski calls habsburg AI. This is, of course, an
(17:38):
absolutely terrible idea. A research paper from last year found
that feeding model generated data into models to train them
creates something called model collapse, a degenerative learning process where
models start forgetting improbable events over time as the model
becomes poison with its own projection of reality. The paper,
called the Curse of recursion. Training on generated data makes
(18:00):
models forgam highlights an example where feeding a generative AI
its own data eventually destroys its ability to answer questions,
and within nine generations. One answered a simple prompt about
architecture with an insane screed about jack rabbits full of
at symbols and weird characters. So not to worry again.
The tech overlords have come up with a great idea
(18:22):
to fix this problem. Their common retort to the problem
of synthetic data is that you could use another generative
AI to monitor the synthetic data being fed into all
model to make sure it's right. At this point, I'd
like to get slightly angry. Are you kidding me? Are
you fucking kidding me? You're saying that the way to
make sure the data generated by an unreliable generative AI
(18:45):
is to use another generative AI, one with the same
goddamn problems, which also hallucinates information that knows nothing. You're
going to use that AI to monitor whether the data
that is created by an AI is any good? Are
you completely insane? You insane? You're going to feed the
crap from the crap machine into another crap machine to
make it not make crap? Why am I reading journalists
(19:07):
credulously printing this ridiculous solution in the New York Goddamn
Times every time? Every time, these bubbles are inflated because
tech executives are able to get their half ass, half
baked solutions parroted by reporters who should know better. You
don't have to give them the better affair of the
goddamn fucking doubt. This is how we got the bloody
meta us. Pardon me, I've calmed down now. Anyway. Anyway,
(19:31):
if you're worried about model collapse, you're already too late,
as these models are likely already being fed their own data.
You see, these models are trained on the web, as
I previously told you, and they're desperate. They need data.
They need more stuff. They need more stuff to ingest
so they can spit out more stuff. The problem is
that these machines are purpose built to make a lot
(19:54):
of content, and so the Web's already being filled with
generative AI. Generative AI is already spamming the Internet. A
report from four oh four Media from last week said
that Google Books has already started to index several different
works that were potentially written by AI, featuring the hallmark
(20:14):
generic writing tropes of these models. Four or four Media
also reports that the same thing is happening over at
Google Scholar their index of scientific papers with one hundred
and fifteen different articles featuring the phrase as of my
last knowledge update a specific phrase spat out by generative models.
This is really bad, by the way, and this is
(20:34):
only going to get worse. When you have an Internet
economy that is built so that the people that can
put the most out there will probably get the most traffic.
They're going to use these tools. These tools are great
for that. If you don't give a rap fuck about
the quality, this is the best thing in the world
for you. And that's the thing. This is a problem
(20:54):
both created and caused by these models. You see. The
other dirty little secret of generative II is that these
models unashamedly plagiarize the entire web, leading outlets like The
New York Times and authors like John Grisham to sue
open ai for plagiarism. While open ai won't reveal exactly
what their training data is, the New York Times is
able to successfully make chat gpt reproduce content from the newspaper,
(21:18):
and the company has repeatedly said that it trains on
publicly available data from the Internet, which will naturally include
things like Google scholar and Google Books. The Times also
reports that open ai has become so desperate for data
that they've used their whisper tool to transcribe YouTube videos
into texts to feed into chat GPT's training data. Pretty sure,
(21:39):
that's plagiarism, but who am I to tell you? And
as the web gets increasingly pumped full of this generative content,
these models are going to just start eating their own swill,
slowly corrupting themselves in a kind of ironic death. According
to Zakhar Schumelov, one of the authors of the model
collapsed paper at the University of Cambridge, the unique problem
(22:01):
that synthetic data creates is that it lacks human errors.
Human made training data, by the nature of it being
written by a human, includes errors and imperfections, and models
need to be robust to such errors. So what do
we do if models are trained off of content created
without them? Do we introduce the errors ourselves? How many
errors are there? How do we can introduce them more?
(22:22):
And indeed, what are the errors? What do they look like?
Do we even know? Are we conscious of the errors
in the human language that make us human? The models aren't? Well,
maybe they are. It's kind of unclear. It's kind of
tough to express how deeply dangerous the synthetic data idea
is for AI models like chat, GPT and Claude are
(22:45):
deeply dependent on training data to improve their outputs, and
their very existence is actively impeding the creation of the
very thing they need to survive. While publishers like Axel
Springer have cut deals to license their companies data to
chat GPT for training purposes, this money is flowing to
the writers that create the content that open Ai and
Anthropic need to grow their models much further. In fact,
(23:07):
I don't think you're going to see more journalists get
hired as a result of these deals, which kind of
makes them a little bit stupid. This puts AI companies
in a kind of Kafka esque bind, but they can't
really improve a tool for automating the creation of content
without human beings creating more content than they've ever created before,
(23:28):
just as said tool actively crowds out human made data.
It's a little silly. The solution to these problems, if
you ask open AI's Sam Altman, is always more money
and power, which is why the information reports he is
trying to convince open ai investor Microsoft to build him
and I'm not kidding, and one hundred billion dollars supercomputer
(23:49):
called Stargate. This massive series of interconnected machines will require
entirely new ways to mountain cool processing units and is
entirely contingent on pen AI's ability to meaningfully improve chet GPT,
something Sam Mortman claims isn't possible without more computing power.
(24:09):
To be clear, open Air already failed to build a
more efficient model, dubbed Arakis, which ended up getting mothballed
because it wasn't more efficient. It's also important to note
that every major cloud company now has inextricably tied themselves
to the generative AI movement. Google and Amazon have invested
(24:32):
billions into chat GPT competitor Anthropic, and both claim to
be Anthropic's primary cloud provider, though isn't really obvious which
one is. In doing so, they've guaranteed, according to a
source of mind, about seven hundred and fifty million dollars
a year of revenue for Google's cloud and eight hundred
million dollars a year of revenue for Amazon Web Services
(24:53):
the cloud service from Amazon by mandating that Anthropic uses
their services to power their clawed model. This similar to
the thirteen billion dollar investment that Microsoft gave open AI
last year, most of which was made up of credits
for Microsoft's cy or cloud, And I somehow doubt that
Microsoft is going to be the noble party that goes
on the earnings and says, well, we don't want to
(25:15):
count the credits that we gave open out want to
be fair. No, they're going to mash that shit right
back into their revenue. Kind of a con kind of
makes me angry when I think about it too. Anyway,
let me just put that aside and not going to
get pissed off again. Look, I'm surprised more people aren't
really upset about this very incestuous relationship between big tech
and this supposedly independent generative AI movement. Microsoft, Google, and
(25:40):
Amazon have effectively handed cash to one or two companies
that will eventually hand the cash back to them in
exchange for cloud services that are necessary to make their
companies work, and all three big tech firms are spending
billions to expand their data center operations to capture this
theoretical demand from generative AI. Every penny the open AI
(26:02):
or anthropic makes will now flow back to one of
three big tech firms, even more so in the case
of open AI, because Microsoft's investment entitles Microsoft to a
share of any future profits from open ai and chat GBT. Yeah,
it doesn't even really matter if they make one, because
Big tech wins either way. Anthropic has to use Google
(26:23):
Cloud and Amazon Web services, open Ai has to use
Microsoft's z your Cloud, and Microsoft is actively selling open
AI's models to their zero cloud customers. And every time
somebody uses open AI's models, that model is being run
on a zero cloud, generating revenue for Microsoft. This is
the rot economy in action, by the way, Big tech
(26:45):
has funded its biggest customers for their next growth revenue stream,
justifying this massive expansion of their data center operations because
ai is the future and they're telegraphing growth to these
brainless drones in the market who will buy any thing,
who never think too hard about what they're actually investing in. Hey,
i's this big, sexy, exciting and theoretically powerful way to
(27:08):
centralized labor. And it's innovative sounding enough that it allows
people to dream big about how it might change their lives. Now,
it might help them not pay real people to do shit. Yeah,
here's the biggest worry I have. Here's the real pickle,
here's the thing that keeps me up at night. None
of these companies seem to have appeared to consider something.
(27:31):
What if generative AI can't actually do any of the
things they're excited about. What if genera avii's content, as
you probably seen from anything chet GBT spits out, isn't
really good enough. Hey, is anyone checked if anyone's actually
using these tools, if they're helpful to anyone? Is this
(27:53):
actually replacing anyone's work? Huh, that's a bit worrying, mate,
I didn't think about that before. Just kidding, I've been
thinking about it for months. Look, here's the thing. I
think that the big problem here is that Sam Altman
and his cronies have allowed the media, the markets, and
(28:13):
big tech to fill in the gaps of their specious messaging.
They've allowed everybody to think that open AI can do
whatever anyone dreams. Yeah, I don't think that generative AI
can do much more than it is today. And also,
from what I've seen, none of these generative AI companies
(28:34):
actually make a profit, and with each new model they
become less profitable, and I don't see that changing in
the future. And so I've dug in a little more
looking under the hood. All the demand that's spurring Microsoft, Google,
and Amazon's data center operations might not actually be there.
(28:57):
My friends, I think we're in the next tech bubble,
and the next episode, I'm gonna walk you through how
I think it might pop. Thank you for listening to
Better Offline.
Speaker 4 (29:15):
The editor and composer of the Better Offline theme song
is Matasowski. You can check out more of his music
and audio projects at Matasowski dot com, M A T
T O, S O W s ki dot com. You
can email me at easy at Better offline dot com,
or check out Better Offline dot com to find my
newsletter and more links to this podcast. Thank you so
(29:35):
much for listening.
Speaker 2 (29:37):
Better Offline is a production of cool Zone Media.
Speaker 4 (29:40):
For more from cool Zone Media, visit our website cool
Zonemedia dot com, or check us out on the iHeartRadio
Speaker 3 (29:46):
App, Apple Podcasts, or wherever you get your podcasts.
Host
Ed Zitron
Popular Podcasts
24/7 News: The Latest
Today’s Latest News In 4 Minutes. Updated Hourly.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations.
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.