AI: Think of the Possibilities
This is an introductory article to what I expect will become a series of articles summarizing my thoughts (and the thoughts of other computing and machine learning professionals) on the latest developments in artificial intelligence. Make no mistake: this technology is revolutionary in the same way that other technologies were, like the printing press, the steam engine, or the digital electronic computer.
List of articles
Trying and failing to use ChatGPT for search — I experiment with ChatGPT to see if it could be useful as an assistant to me in the work I do as a software engineer. It does not go well.
App Idea: Upload Your Brain to a Neural Network — a lighthearted article on an obvious idea I had for neural networks like GPT.
Netflix Showrunner On-Demand — taking recent advances in the algorithm known as Stable Diffusion to it's logical extreme.
Think of the possibilities
Artificial intelligence, more technically
called machine
learning,
and even more
technically, artificial
neural networks, are becoming more and more amazing in their
capabilities. They first started doing some truly disturbing things
a few years ago with
"deep
fakes," creating images that look absolutely real of people that
never existed by sort-of averaging
the images of millions of
other people. They can even
create animated images,
swapping your face with that of
a celebrity or
politician, allowing you
to make a video of you doing or saying things as that celebrity or
politician that they never actually did or said. Then came
the deep
fakes for peoples voices, so with enough samples of a persons
voice, you could translate your voice into the voice of a celebrity
or politician.
People immediately saw the potential for abuse, but their fears were placated when it was shown that you could train other neural networks to detect deep fakes. Of course, I think what ordinary people don't yet grasp is that Generative Adversarial Networks (GANs) are created by putting two neural networks in competition with each other and letting them learn how to best each other. So it is possible to train an AI that learns how to beat the deep fake detectors, and then you have to re-train the deep fake detectors to beat the deep fake AIs. However it seems to me that for the time being, people are not quite as concerned about deep fakes anymore.
But then the GPT
algorithm
was developed. The various GPT networks, suffixed
by a number (GPT-2, GPT-3, GPT-4), each successive network being
bigger and more capable than the previous, have got everyone's
attention now. The reason is that these neural networks actually
seem to understand
things. You can ask it questions,
and it gives you a reasonable answer.
It passes
the Turing
Test
with flying colors (an informal, unscientific test that
simply asks, "can this machine trick people into thinking they are
talking with a human?"). It can probably play and win the game of
Jeopardy, or at least, come close to
what the
IBM Watson supercomputer achieved 10 years ago but with a
fraction of the computing power (and electricity) that Watson
required. It
can probably
correctly diagnose some diseases by reading a person's
description of their symptoms. It
can pass
the Bar Exam, and probably offer sort-of reasonable
legal advice for very simple cases. It
can write
simple computer programs for you by just you explaining what you
want your program to do. (See
this video demo.)
How does GPT work?
GPT is a class of modern neural network architecture called
language
transformers
. A simple (perhaps too simple) way of
describing how these AIs work is that they begin with a massive
chunk of text called a corpus
. In the case of GPT, the
corpus is called The
Pile
which is around 825 gigabytes in size. All of it is
written by humans, and has been taken from places on Internet that
are free to use and permissive about copyright. Through a process
called training
this mind-bending ream of data is compressed
into a massive chunk of statistical information, analogous to how
video or image
data is
compressed (such as with ZIP or MP3 files).
These chunks of raw statistical data called neural network
models
begin to compress the corpus more and more, and it begins
to encode intricate patterns in the data that loosely approximates
the grammar and semantics of human language, quantifying
relationships between similar words (even between different human
languages), and rating the emotional sentiment of clusters of
words. And I must stress, it does all this
without any human intervention to guide the formation of
the neural network model, a technique
called unsupervised
learning. The only human intervention involved was during the
process of collected the training text.
When the neural network model completes training, you can use the
model by taking an ordinary human sentence, encoding
it as a very big number, and transforming
this number by
applying it to the statistical data of the neural network model. A
new very big number appears at the other end of the computation, and
after decoding this new very big number using the
reverse of the process that you used to encode your sentence, a new
sentence of words is formed. Somewhat magically, this sentence that
is decoded contains words and sentences that appear to be written by
a human.
Since the neural network model has statistical information about the similarity between words encoded within them, these neural network models are very good at human language translation (e.g. English to Japanese). They are good at summarizing larger pieces of text, and they can also report whether a paragraph has a positive or negative emotional tone to it — advertising companies who do market research such as Google are very excited by this. And, most impressively, the model can take a sentence and predict what sentences might most likely come after that sentence. The GPT chat bots you may have seen use this mode of predicting the next sentence to create impressive the results that are gaining them so much attention.
The neural network model is a massive chunk of
statistical data, tens of gigabytes in size, but still considerably
smaller than the training
data that it has compressed. Just
for comparison, a BluRay disc holds roughly 45 gigabytes, a modern
cell phone has about 8 gigabytes of working memory
available. (Again, the training data is typically around 825
gigabytes.) Obviously these neural network models are too large for
ordinary personal computers to use, let alone build from the
training data. So they have to run on large server computers and be
used over the Internet. That introduces the first
problem.
The public has no control over this technology
The individuals who use this AI, or the people who contributed training data do not control the model, the company that built it controls it. It is not something you yourself can own, or something you can do yourself in your own home, unless you happen to own a and operate a computer that costs more than your whole house. You are dependent on a for-profit company selling you access to this technology as a service, even though they exploit a free public resource to build their service. And you have no guarantee whether the company selling this service to you has constructed an AI that is free from the political biases of the proprietors of that company.
The trained models then can do these amazing things that we have been seeing. But as any professional doctor, lawyer, or engineer, will tell you, the answers that GPT gives you are wrong more often than they are right. And even when it is wrong, it tells you with an air of supreme confidence, mimicking the style in which all articles on the Internet are written (including this article that you are reading now). It even offers to you a plausible-sounding explanation for why what it said is true.
If the Internet were full of people always being very cautious
and uncertain about everything they said, the GPT chat bot would
probably mimic that sentiment. Instead, the neural network
speaks
with the tone of an amateur who believes they know
much more about a subject than they actually do — it did take
the training data from places on the Internet not guarded by pay
walls, after all. This introduces the second problem:
People may begin to trust and rely on it, even if they shouldn't
Suppose AI becomes much more accurate, and if people come to rely on these AIs, who is to blame when the AI leads you astray? For now, web apps like ChatGPT have a disclaimer warning users that the information produced by their AI is not to be taken as accurate or authoritative, but as these AIs improve in accuracy, people may begin to put too much faith into them. We have already seen this with self-driving cars leading to fatal accidents.
I tried using a GPT model myself (ChatGPT from the OpenAI company). After spending some 15 minutes trying to explain to the computer what I wanted it to do, and after many completely wrong answers, it only managed to write the smallest, simplest program that could satisfy my prompt, a program I could have written myself in 30 seconds. My second attempt, I used it to search for a technology that I thought may exist. The machine first told me it existed and what it was named. But after I pressed it to tell me where I could read more about this technology, it admitted to me that the technology they named didn't actually exist, but if it did that is what it would be called. So it is not quite ready as a personal assistant to any professional. Not just yet, anyway.
But the technology will improve, and so will the accuracy of the
results. Computer hardware manufacturers, especially companies who
make GPUs
(graphics
cards
, which let you play games that look absolutely real)
are now
creating Tensor
Processing Units
(TPUs) for training and running artificial
neural networks that power modern artificial intelligence like
GPT. So it is only a matter of time before the cost of creating
larger and larger GPT networks comes down, and these larger neural
networks won't get the answer wrong quite as often. People may
start using them to do real work. But who is responsible
when the machine makes mistakes? Should we let these machines
completely replace humans? That introduces the third
problem.
Society is not ready to let people without jobs live peaceful, happy lives
The implications of an AI that is getting increasingly better at tasks that require expert knowledge, obviously, is that a lot of white-collar jobs are probably in danger of disappearing. From my own professional perspective, I can see my job as software engineer as probably safe.
But the job of the computer programmer is not safe anymore (which I am saying is a different job from that of software engineer). Anyone going through coding boot camp to learn how to develop apps for your iPhone should reconsider their future career path. I am fairly certain all apps that you find on the app store, even games, will be developed by AI in the near future. Not a big loss, since those jobs were horrible to begin with — lots of busywork, not the slightest bit of creativity required to do those jobs. That said, those jobs do pay the bills for lots of people, unfortunately not for much longer. I can think of some other examples...
The job of the lawyer is probably safe, the job of the paralegal is probably not.
The job of journalist is probably safe, the job of writing breaking news stories, or of writing news briefs, is not. The Associated Press has used AI to generate news briefs for over 10 years now.
The job of insurance agent is probably not safe.
The job of advertising design artist is probably not safe.
The job of filming or creating music for TV and film is probably not safe anymore.
The job of actor is probably safe, but will change from what it is today — someone who stands in front of green screens with motion capture devices attached to their body — to a person who sells images of their body for training AI to generate films with characters that vaguely share their likeness and voice. The H&M fashion company has used AI-generated models to model their clothes for over 10 years now.
The job of adult entertainer is not safe. Many people have made good money selling images of their bodies to specific audiences. But now, so much well labeled image and video data exists for free on the Internet, greedy companies will make a fortune selling subscription services to generate a video of any pornographic scenario requested by the customer. And unfortunately, there is not yet any law that might force a portion of these profits be paid to the people who's bodies were filmed to create the training data.
It really would not be a terrible problem if some of these jobs disappeared but ONLY if there were an abundance of other jobs that anyone less skilled than a doctor, lawyer, or engineer, might still be able to do. That, or we would need to completely reorganize society around the idea that not everyone needs to have a job, and people without jobs could still live peaceful, happy lives doing whatever they wanted to occupy their time (make art, solve puzzles, play games). Perhaps there should be a social contract where people live to train AI, and in exchange, businesses using AI could provide goods and services to people for free.
If only everyone ran a small businesses using AI to do the job for them
If it were possible for an individual professional to buy a computer, train the computer to do their job, and then keep pulling a salary or wages for the job their computer does for them, then I would be an enthusiastic supporter of this technology. This would be a fulfillment of the promises of modern technology that I heard as a kid — that it will increase our leisure time and improve the quality of our lives.
But that is not how things work right now. On the contrary it seems to be getting harder year after year for small business owners to sustain profitable businesses. With many corporations using AI to automate the jobs that many of these small businesses now do, it will only get harder for small business owners. In the US, it is getting harder and harder to find any jobs that pay a living wage at all, or any jobs that provide useful medical insurance, regardless of who you work for.
There are technical challenges as well. The neural networks we train require so much data, and so much computing power to create, that only the worlds wealthiest companies (Google/Alphabet, Meta/Facebook, Amazon, Microsoft, Apple, IBM, Oracle), have any chance of controlling this technology and profiting off of it. All the rest of us lose our jobs, and the corporation keeps all the of profits. They of course also pocket the savings they get from firing people. With this possibility in mind, the future looks bleak and dystopian.
Many have said it, and it is absolutely true: neural networks like GPT are truly revolutionary. And with technological revolutions, political turmoil, and political revolutions, soon follow. The US Civil War, and the first two world wars, were the wars that book-ended the second industrial revolution. We are starting to see how AI plays out in modern warfare, especially with armed drones, in conflicts like in Turkey vs. Iraqi Kurds. Us humans are in for a rough ride these next few decades
...And I haven't even begun to talk about what might happen if we build machines that are more intelligent than humans.
Learn from the mistakes of history
Let's just say that the kind of slave uprising story of machines becoming sentient and overthrowing their human overlords, the kind of story we mostly saw in science fiction films of the 20th century (e.g. James Cameron's The Terminator) are probably the least of our concerns at this point in history. In the long term, we are probably more in danger of becoming slaves ourselves, not to artificially intelligent machines, but to the people who own these machines and use them to give themselves superhuman abilities to exert their will on anyone who did not win the lottery of birth into a wealthy family.
But my biggest fear right now, in the short term, is that the pace of technological evolution is far outpacing our human ability to develop a system of ethics around it. It is time to really start thinking about what the future of humanity is going to look like. With challenges like:
AI now becoming powerful enough to eliminate millions of jobs
climate change,
the still-present threat of nuclear war,
and the new potential threat of large-scale drone wars,
...we cannot afford to just wait and see what happens. We have seen it all happen before, when a new technology gradually shifts the balance of power from the many to the few until world-wide conflagrations like World War 1 and 2 occur. In my opinion, I believe our current historical situation is more similar to that of development of the printing press rather than the machine gun. Like the printing press, unlike the machine gun, more people having control over this technology will be better. Though currently only a very small number of people control this technology.
Regardless of what your opinions are of the communist revolutions of the late 19th and early 20th centuries, they happened for the very practical reason that people could no longer live under the conditions that had evolved at that time as a result of the industrial revolution. And we are dangerously close to approaching that tipping point once more in the year 2023. We need to begin developing a system of ethics regarding the use of AI, and codifying these ethics the into law, immediately. We especially need some guarantee that all people will benefit from this technology, and not just the few people who are wealthy enough to own the worlds most powerful supercomputer.
Sources:
Sasse, Ben. Washington Post 2018-10-19.
Opinion: This new technology could send American politics into a tailspin By Ben Sasse
Burgess, Matt. Wired Magazine 2021-12-15.
The Biggest Deepfake Abuse Site Is Growing in Disturbing Ways
Villasenor, John. Brookings Institute 2019-02-14.
Artificial intelligence, deepfakes, and the uncertain future of truth
Samay Pashine, Sagar Mandiya, Praveen Gupta, Rashid Sheikh. ArXiv 2021-06-23.
Deep Fake Detection: Survey of Facial Manipulation Detection Solutions
Warzel, Charlie. BuzzFeed News 2018-08-28.
I Used AI To Clone My Voice And Trick My Mom Into Thinking It Was Me
Gabbot, Adam. The Guardian 2011-02-17.
IBM computer Watson wins Jeopardy clash
Yirka, Bob. Science-X Tech Xplor 2023-01-24.
ChatGPT found to be capable of passing exams for MBA and Medical Licensing Exam
Sloan, Karen. Reuters 2023-01-26.
ChatGPT passes law school exams despite 'mediocre' performance
Tung, Liam. ZD-Net 2023-01-26.
ChatGPT can write code. Now researchers say it's good at fixing bugs too
Goodwins, Rupert. The Register 2022-12-12.
ChatGPT has mastered the confidence trick, and that's a terrible look for AIAssociated Press 2023-01-05.
Nearly 400 car crashes in 11 months involved automated tech, companies tell regulators
Darcy, Oliver. CNN Business 2023-01-26.
BuzzFeed says it will use AI to help create content, stock jumps 150%
Pous, Terri. Time Magazine, 2011-07-12.
H&M Admits to Using Computer-Generated Bodies for Models
Engler, Alex. Brookings Institute 2019-11-14.
Fighting deepfakes when detection fails
Covington, Taylor. The Zebra 2023-01-05.
Small Business Statistics
Zahn, Max. ABC News 2023-01-10.
Here's the difference between a 'minimum wage' and 'living wage,' and why it matters
Farooq, Umar. The Intercept 2019-05-14.
The Second Drone Age How Turkey Defied the U.S. and Became a Killer Drone Power