What's the Deal With Exploiting Outsourced Labor Markets?
A baseless prediction on the future of AI and Publishing (Part 2 of 3)
(Note: the following contains minor spoilers for Guardians of the Galaxy: Volume 3)
I cannot think of a single piece of media that has captivated me this year quite like Nothing, Forever. Aside from having a title that sounded like the next great emo LP, Nothing, Forever was comedic gold on existential levels. It wanted to be an easily accessible sitcom, built upon the principles of established joke structures and comedic timing. However the humor was not in its success, but rather its failures.
Nothing, Forever was a self-scripting and self-animating stream on Twitch. You may have heard it referred to by its more colloquial name, “AI Seinfeld.” The Twitch stream replicated the classic sitcom Seinfeld through pastiches like Jerry’s stand up sets and conversations in Jerry’s apartment. The conversations almost always involved talking about a new restaurant, though on occasions it talked about other things which were equally not all that important. While the technology to produce a show on the fly is impressive, it was not what its creators hoped it to be.
The scenes were oddly paced, the laugh track often came in at the wrong time or not and all, and the characters frequently contorted their bodies while attempting to sit down. However these bugs which made the show hilarious, were lost on its creators who aimed to create a show that people could turn on wherever, whenever and immediately enjoy. They genuinely believed, and likely still believe, that the future of entertainment is auto-generated.
The Reality of Where We Stand
Many writers have expressed concerns that they will soon be replaced by AI programs which produce automated scripts and novels, eliminating the need for human writers. These fears are not unfounded, but they are often spoken of in dire and fatalist terms, as if the human race is actively at being exterminated by sentient, self-aware robots. While I will concede that it is possible for AI programs to one day replace the creative expression products which humans currently produce, the notion that this is near and inevitable is, like a short lived comedy game show, Bunk.
While programs like ChatGPT may be impressive, they are at this time word calculators. They review the data given to them and predict the words needed to produce the request. You input words, it outputs words. It does this not through free thinking, but rather through probability.
As computational linguist Emily Bender and Alexander Kollar argue in their paper Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data, Large Language Models (LLM) will struggle to achieve Natural Language Understanding (NLU) because the model is built on probability and data inputs, thus limiting its knowledge. This can be best understood by the Hyper Intelligent Deep Sea Octopus thought experiment in the Climbing towards NLU paper. I will save myself some time and provide the summary which New York Magazine wrote in their feature on Emily Bender.
Say that A and B, both fluent speakers of English, are independently stranded on two uninhabited islands. They soon discover that previous visitors to these islands have left behind telegraphs and that they can communicate with each other via an underwater cable. A and B start happily typing messages to each other.
Meanwhile, O, a hyperintelligent deep-sea octopus who is unable to visit or observe the two islands, discovers a way to tap into the underwater cable and listen in on A and B’s conversations. O knows nothing about English initially but is very good at detecting statistical patterns. Over time, O learns to predict with great accuracy how B will respond to each of A’s utterances.
Soon, the octopus enters the conversation and starts impersonating B and replying to A. This ruse works for a while, and A believes that O communicates as both she and B do — with meaning and intent. Then one day A calls out: “I’m being attacked by an angry bear. Help me figure out how to defend myself. I’ve got some sticks.” The octopus, impersonating B, fails to help. How could it succeed? The octopus has no referents, no idea what bears or sticks are. No way to give relevant instructions, like to go grab some coconuts and rope and build a catapult. A is in trouble and feels duped. The octopus is exposed as a fraud.
In order for a machine to truly understand the world in the ways which humans do, it needs massive amounts of data. We are talking immeasurable amounts of information input into the system and then classified so the program then knows what to make of it. It would then need the ability to take in and process new information on its own, without the need for classification of materials, lest it fall behind in our ever changing world.
However even if it could obtain and process all of that information, the machine itself would not understand the emotions and struggles of humans. It would only understand how to describe the emotions and struggles of humans, and even then, not very well. It would just be in the general ballpark. This ballpark is the reason AI text is actually identifiable.
On the Dangers of Stochastic Parrots and Talking Raccoons
At this time, these LLM’s just can’t replicate human dialogue and syntax with the accuracy needed to be accepted as human, and given the design of these programs, they will never be able to create anything new. They can only replicate the things they have been shown. They cannot problem solve or create a new thought.
This is the reason I have no worries about an Artificial General Intelligence gaining sentience. I have worries about tech bros poorly designing a computer program and plugging it into critical medical machinery or national defense equipment, but I do not fear HAL. The type of world ending AI talked about in Sci-Fi films has never been scoped, and the components required for dreaming of robotic sheep are still theoretical. Many assume modern AI will progress beyond what humans can keep up with because we’ve read novels and seen films where AI did just that, but to believe that is to assume that because we saw it in a futuristic movie, it is destined to happen in real life. Those novels and movies were speculative works based in a version of reality imagined by the author.
It is important to realize LLM’s like ChatGPT are not the AI or AGI of science fiction. They are their own program, and there is no guarantee that their capacity is what we’ve always imagined they would be. In fact, I’m personally inclined to believe that we are nearing the end of the current wave of AI advancements (more will come later, but there will be a cool down for now).
How could I think such a thing? After all, didn’t tech leaders just call for a pause on AI development to establish policies which protected mankind?
Yes. They did. And that’s exactly why I’m not so concerned. DeepMind Senior Research Scientists Zachary Kenton signed the letter claiming a pause was needed. However if he was really concerned, maybe he could just stop working on AI projects. Since he’s a Senior Research Scientist for Google’s AI project DeepMind after all. Elon Musk signed the letter, which is funny considering he continues to pour money into AI research. And if Encultured AI co-founder and CTO Nick Hay is so concerned about AI’s advancement, maybe he shouldn’t be hiring LLM Specialists right now.
This letter was not a call for a pause. It was an advertisement for further investors. They want this technology to be seen as a hyperintelligent being so investors stake their claim in the eventual profit, but it’s not that. It’s a probability machine combining English words (LLM’s do not currently work well with other languages). They want investors to think they have created Rocket Raccoon, but in the end they’ve just made another meth dealing octopus.
Yep. That’s right. Guardians of the Galaxy Volume 3 was, in my opinion, largely about the separation between man and machine outputs. The High Evolutionary wanted to make something with the ability to create and problem solve, but all of creations (with the exception of Rocket Raccoon) were limited to only doing the things they’d been taught to do. They could not look at a new problem and find a solution unless they already had the answer to reference.
That is where AI is now. It cannot create, just regurgitate.
A Show About Belittling People Who Aren’t Like You
You may be looking at this and thinking, “That’s where AI is now, but soon it will progress beyond that.” As I’ve conceded, I could see the existence of such a program some day, but I believe such a program is 30 - 40 years away if it is to occur at all. There is far too much manual labor needed before the program itself can become easily reproduceable and affordable to prospective purchasing publishers.
Even if they were affordable and publishers desired them, they would require extensive editorial oversight, because these Large Language Models lack discernment. They take the data given, process it according to what they’ve been told, and then spit out the result. In order to do this the system needs large amounts of training data. So researchers pull from the internet.
Now, if you’ve ever been on the internet, you know it sucks. It’s kind of racist… and misogynist… and homophobic… and horny… very horny. It’s also not always the bastion of intellectual debate some had hoped it would be, with political factions dividing themselves into online communities and repeating talking points till they became memes to insult their opponents.
To feed all that into a machine without denoting the content to avoid would create a monster, but how do you review all that content to classify whether the system should or shouldn’t reproduce it? The same way western countries have been addressing mass labor needs for decades. You outsource it!
That’s exactly what OpenAI did for ChatGPT. They paid Kenyan works $2/hr to review content and classify whether it was violent, hateful, sexually explicit abusive, or any other type of content that’s bad for the public image. This information was to be fed into program which would detect and deter the creation of such content by ChatGPT.
The laborers essentially worked for a psychological sweatshop, however. They were forced to meet content review quotas while reading fan fiction of violence, rape, incest, drug abuse, child abuse, and more. They were supposed to have psychiatric care available to them at all times, but workers reported this was not made available to them. The project was shut down by the Kenyan company when OpenAI allegedly sent images for tagging that included child sexual abuse imagery. Even sweatshops have higher standard than techbros.
However even with the designed content deterrent, programs like ChatGPT aren’t perfect. In fact, they’re concerningly imperfect. That is what led Timnit Gebru, Google’s AI ethicist, to co-write On the Dangers of Stochastic Parrots. AI lacks the sufficient guardrails to not repeat terrible things.
And that brings us back to Nothing, Forever. The show was rolling along, until all of a sudden Jerry went right-wing. Not small government and low taxation right wing. I mean angry uncle right-wing.
Nothing, Forever was suspended for weeks after that. It’s back up, but redesigned and lacks the charm of the original. Perhaps because it now has less freedom. There are more guardrails in place to prevent it from transbashing again. I’m glad those guardrails are in place, but I believe it reveals a fundamental truth about creativity.
In order to create, you must be able to discern. You don’t have to be perfect, but it is impossible to create an impactful story without some understanding of human behavior beyond the binary good and bad. AI, in its current form, lacks the ability to do that, and such a program would still require major editorial oversight to ensure the n-word doesn’t leak through.
So, I don’t personally see AI as a major threat to creators. I think there should still be pushback, to pre-empt any attempts to move from humans to AI, but I don’t think the tech is there yet.
But we must ask ourselves now, is it really that important to save the human generated arts? Tune in next week, because you probably will be surprised with where this goes.
Media Recommendations for You to Ignore
Film: Saw The Flash yesterday. It’s fun, but not very good. You should probably just rewatch Wall-E. Wall-E slaps.
I’m tired and going back to bed.