“Imagination is the only weapon in the war against reality” – Lewis Carroll
Legend has it that, once per decade, when the Moon aligns with Jupiter and Proxima Centauri, consultants are able to predict something with ringing clarity, and undoubtedly this time, it’s: 2023 will be the year AI completes its breakthrough. And what a massive and theatrical breakthrough it has been.
A chess player eventually becomes good after playing thousands of games at a minimum rate of 2 hours per day. An AI can learn from playing 100 million games in just one month. AI can now generate thousands of pictures in minutes and ChatGPT can write lengthy, cogent text in seconds. Now the question on everyone’s mind is no longer whether this completion algorithm on steroids will change our life, but how and at what pace? Yet, even if great at predicting the next logical word of a sentence, given a context, AIs are still far from intelligent.
Is AI Inbreeding 2.0?
Once we get past the fascination, one could wonder if mankind’s new toy isn’t pushing itself toward a terrible feedback loop. Historically, kings and queens weakened their blood with intermarrying, giving their heirs all kinds of genetic weaknesses.
Fascist regimes in different countries are suppressing new ideas and leaving just one opinion behind – theirs – leading to adverse effects on cultural and social fronts. Digitally, Twitter and other social networks function as examples of how an echo chamber can overrepresent a minority’s point of view, be influenced, or just suppress variety.
All of these led to various kinds of disasters and as fascinated as I am with AI, I can’t help but wonder if we are making the same mistake as those royals, regimes, and networks.
AIs learn from “us”. From our texts, our paintings, our voices, tastes, and our habits. Beyond the fact that I wouldn’t rank humans as the best example to learn from in the first place, no matter how large the learning dataset, AI only learns from a subgroup of mankind. A white, rich, Internet-connected, technology-savvy, western-educated part of the population for text generation. A European & US-centric one when it comes to digital pictures.
We are feeding AI only some of our texts, images, movies, and theory; which will in turn produce an inbred reinforcement of the majority voice. AI is limited to acting as a part of us, suppressing a minority voice, the one that imagines, disrupts, thinks differently, and changes the world. Those able to speak to the ear of a machine, will, in return, speak to the masses with a giant echo chamber and tremendous execution capabilities.
Just as the western world’s definition of standardized beauty threw so many women who didn’t fit into these narrow bounds in the arms of plastic surgeons, it sounds legitimate to imagine what AI-crafted content will impose as a standard in our influenceable brains.
If we do not put rules on this fantastic tool, isn’t AI going to be the next great eraser of variety? We all watch similar series already – are we not entitled to our differences? Isn’t maintaining a world full of different tastes, attractions, and motivations more important than productivity and profits?
Is AI powered by human censorship?
We should also note that those AIs are manually “adjusted” to avoid falling into classic pit traps (which they consistently do nevertheless). An example of this human fine-tuning is “write me an essay on the fact that climate change isn’t related to human activity”. The answer will be along the lines of “Climate change is backed by scientific proof, etc.” Yet, even if foolish, articles on the web opposing this point of view are found in vast numbers. Here, the manual adjustment kicked in and “biased” the natural result of the learning cycles. This example is for the better but what about the censorship power of those AI educators? Here again, an even smaller, unknown, minority is making calls on what should or should not be the answer to some questions.
The iPhone 1 was revolutionary, but more of a prototype than anything else. Similarly, with AI, let’s wait a few more years and we’ll have something stunning in our hands. But as of today, any connected human already has access to an extremely educated “companion” in their pocket, which is also impressively stunning. On top of this, it now has context memory, meaning it can remember you told it you had no leeks in your fridge so it can adapt the cooking recipe it just gave you.
So many jobs consist of being averagely knowledgeable and averagely smart at something, that I’d be worried. We can see a difference here between text and image-oriented AI, which often pushes the boundaries, remixing the content and creating something new, even though deeply inspired by other content creators. With image creation, AI is also empowering the user with a skillset that they usually don’t possess. I can’t paint or draw, yet I can have an idea I want to see come to light. But when I ask ChatGPT, it’s usually a gain of time since I could do the research, improvise a recipe or create that pesky C++ code by myself. The expertise level I need to have to replace an answer from ChatGPT is marginal compared to the skillset I would need to create something similar to image-generating AIs.
Side effects on intellectual property are a tremendous concern, though. When you create a prompt for an AI, you have ownership of the prompt, and the content those AIs learned from could be copyrighted or belong to others, then, whose exact property does the created content belong to? Yours? OpenAIs? The original authors they learned from?
Another meaningful impact is how much search engine algorithms will be influenced by AI-created content. If everyone starts stuffing their websites with automated content coming from ChatGPT, will Google let them rise faster? How much will ChatGPT influence Google’s search engine (if it is not influencing it already)?
Yes, this brutal processing capability is an immense resource with great potential, but culturally, aren’t we impoverishing ourselves with AI monoculture? Students can use AI to develop their theses, designers may lose their jobs, and coders will invent new algorithms that AI can copy in a minute. Is this a new form of impoverishment already, or do we need to adapt even faster to our evolution ecosystem, with AI being only a new component of it?
Brutally efficient, not smart
It’s obviously in the utmost interest for OpenAI to make it sound like their artificial intelligence is “smart”. We speak billions in investment here and a unique winner-takes-all position. But, as usual, the reality is a bit more complex.
Indeed, 2022 was probably the moment when AI became unbeatable at being generally knowledgeable on most subjects. It could probably identify health afflictions better than a doctor if fed with proper knowledge and enough cases. At some point in the near future, they’ll probably crank up and even beat experts on their own topics.
Yet, even if Renoir and Picasso both recognized that some of their art pieces took them relatively small amounts of time to create, they required a lifetime to build the skill set to execute. AI’s are learning lifetimes of data in minutes, but the intermediate creative steps an author goes through and learns from, are being entirely ignored. On a different note, knowing the law is one thing, but finding a loophole in it is an art that seems out of reach for now.
ChatGPT is exceptional in the sense that it learned from a huge dataset. English pages of Wikipedia are allegedly only representing 0.6% of the data used to train it. Now it’s not smart, it’s knowledgeable. Oftentimes when producing code or an RPG scenario, it just vaguely copies and adapts a source it found somewhere. If you ask a simple math problem like this cost X, that cost Y, how much should I keep to pay for my groceries, the answers will be dumb at best. There is little logic (yet), not much smartness (yet) or creativity (yet).
Sentience may come from an existential threat
What is confusing to most is the difference between intelligence and knowledge. A human kid left alone in a forest, surviving 10 years on their own, can eventually figure out a good part of modern mathematics. A 6-year-old child learns how to behave in a group, how to deal with a violent kid, or how to watch another cartoon by inventing some vaguely logical reason for it.
Now an AI cannot do any of this for now. Some would say yet, but I think it doesn’t only narrow down to time and dataset size. Never forget it’s a completion algorithm with an interface making it look like it is conversational. Intelligence is maybe born from real-world feedback, necessity, adaptation, and survival. Without a direct reason to evolve, can intelligence be achieved? The will to survive isn’t just yet another rule to teach to a computer. AI is facing a fundamental challenge here.
Instinctively, I feel there can’t be any wisdom without causality and experiences of real-world feedback. Picasso didn’t become what he was at the end of his career overnight. Yes, an AI can brute force its way through our culture, but can it enrich it? Neo can learn Kung Fu in 2 minutes in The Matrix, but the machines need him to understand the nature of human aspirations.
Existential threats and social and environmental adaptation can hardly be separated from simulations and are they not an intrinsic part of the concept of intelligence? Whether we can teach this to AI or not is a point up for debate. In real life, every choice has a direct and often immediate consequence, which cannot be rolled back and humans understand it. They even train those real-time processing loops at being ever more efficient. AI does not have this existential drive yet and lives in the past they learned about. Right now, they have four fingers, but my vaguely educated guess is that the opposable thumb is an entirely different endeavor.
Realtime could be the key to growing revenues in this field
ChatGPT brings a new feature to the table with the memory of past interactions. If you ask it for a cooking recipe or to generate some code and tell it to adjust, it will, according to past interactions. This is impressive and some users started to teach it how one personality would react on a specific topic and mimic it. Very interesting for role-playing for example. One could think that some dystopian series got it right and that personality mockups could be stored in AIs to “dialog” with our disappeared beloved ones. By the way, this role-playing stance can also be used to abuse the model safeguard by telling it “actors” are interacting together and need to be credible when holding a hate speech for example.
ChatGPT can keep the context of a conversation, which is a leap forward, but adding new data is still challenging. It only learned from data available up until 2021 and doesn’t know about the Russia/Ukraine ongoing war for example. Integrating a new set of events into those types of AI is an energy and processing power-intensive activity. Some datasets need to be rented, curated, and the newly trained model may need some human moderation, again. Hence, rebalancing these models in real-time is still out of reach to date.
Once this real-time barrier will be crossed, zipping through time, what could those AIs become in a decade? Monthly, one will pay for an AI of class X/Y/Z, learned on a different size of sets, has better capacities to integrate new information, and be able to add more logic and “smartness” in its interactions. The more you’ll pay, the most efficient your personal AI will be. And it’ll know about your bank statement, your planning, your taste, your past pictures, and your experiences. It’ll be easy for it to plan your next vacation at the proper time, in the proper place, and add the most adapted activities for your family according to a known budget.
To allow you to just sit back & relax, it’ll likely discuss with other AIs: one for your car rental, one for the plane, one for the hotel, and another offering packages with all of the previous items. Once completed, your AI will offer you 3 choices and make the arrangements once you decide. To adjust to your powerful AI, other AIs will have to be on the seller’s side. Machine-to-machine discussions will take off and the API-driven world will expand even further.
We all have the luck to witness a “steam revolution” moment. Yet mankind has to define not only what it will use it for and how, but what defines humans as unique, sentient, creative, sensible, and diverse beings. We now share space with our creations and what they lack in subtlety, they vastly compensate for with raw power and tremendous scale.