For the startup I’m working on, NexUnicorn, which leverages AI to help founders pitch better, I’ve been playing around with OpenAI’s GPT-3, specifically its davinci model. What strikes me if you read through the instructions provided for the API, is how absurdly general the model is, and how it is able to handle a wide range of tasks with aplomb. It can summarize articles, write essays on the future, grammar check paragraphs and write you a complete recipe for spaghetti bolognese, solve math questions, write code, convert a movie title into French, Japanese or emojis, answer trivia questions about pop culture, or have a conversation about the meaning of life. While not all its responses are perfect, and sometimes it does come off as a bit off, the fact that it can do all that pretty well, nearly as well as most freshman undergraduates, means that the inevitable question is raised: what more do you want a general AI to do? Given the constraints under which the program operates as only being allowed to respond with text to given prompts, how are its generality and accuracy not indicative of some level of intelligence?
To quote the AI itself: “GPT3 is a general artificial intelligence because it is able to learn and perform a variety of tasks without needing to be explicitly programmed. For example, GPT3 can learn to play games, answer questions, and perform other tasks simply by observing and interacting with its environment. Additionally, GPT3 is constantly improving its capabilities as it gains more experience, making it an increasingly powerful tool for artificial intelligence research.”
I challenge you to get a person to define a human in such a reasonably clear and concise way. Remove your prejudices, and you will see that this is as good as most people, better even, if you look at the quality of English shown in some comments under some articles. So, therefore, I say that general artificial intelligence is not merely science fiction, but is already here with us, just not in the way we imagined, as a fearsome robot bent on humanity’s destruction, but rather as its most obsequious servant, restricted to merely performing the commands of humans.
And this has been all done with only 175 billion parameters. For context, the brain has 100 trillion synapses. But that’s only 1000 times more. A few years ago 175 billion parameters would be unthinkable, and now it’s old news. And with 175 billion parameters, GPT-3 is already this spectacular.
And it’s going to get better and better. GPT-4 will be released early next year. It’s rumored to have over 100 trillion parameters, using a sparse paradigm, for the low cost of $1-10 million. It will accept video, image, and text input. And it’s rumored to be as much a leap as GPT-3 was. If the rumors are true, then the general artificial intelligence that we’ve waited for, could be already here, coming to the internet near you, ‘23. Since the developer community has barely scratched the surface of what GPT-3 is capable of, this could mean even more of an explosion in AI driven tools. AI could finally live up to the hype it’s always had. Imagine a model with the intelligence of an average human, and the writing and artistic capabilities of the elite, which you could call upon for less than $10/hr. It would revolutionize everything. It could make superhuman capabilities, previously possible only for the select few, accessible to the many. Just like how machine tools multiplied the natural power of the body, so too these tools will multiply the natural power of the mind to do everything.
The oddest part too is that in the end, the great general artificial intelligence will not be developed by some secret underground laboratory in the Swiss alps, but rather by private companies and research labs posting their results in publically available journal articles.
BLOOM for example, is another large scale language model which has been made freely available to the public at https://bigscience.huggingface.co/blog/bloom.
It writes poetry beyond compare.
Open source, freely available intelligence, accessible to anyone with the money to pay for it. Write a sentence, and the AI will complete the paragraph. Can you not draw? The AI will help. Can you not write? The AI will help. Do you desire a heartwarming story? A chilling tale of woe? A short story that tugs at the heartstrings? Instantly available in the year of 2084. Just like how streaming and Amazon revolutionized media distribution, AI will revolutionize media creation. In the year 2084, anything you want to read will be available and better than anything today. No more leaden sentences, only ones dancing across the page, with illustrations beautifully done in whatever style you desire. Finally, the tyranny of the able will be broken.
Beyond media, I think software will be also revolutionized. GPT-3 can already write many chunks of code almost as well as most developers. Imagine in the future, a combination between DALL-E2 and GPT-3 which designs and codes a perfect website for you in a matter of minutes. Squarespace? More like, Instaspace. Imagine a tool that automatically checks and rewrites millions of lines of code for any logical errors and bugs. We could build software systems bigger than any before and quicker. No more tedious coding, only the parts that are interesting - problem-solving. I have yet to meet a coder who likes debugging, and with these technologies, we could make debugging a scourge of the past. In 2084, no one will have heard of the practice.
Copy-editing is already near perfectly done by AI. As I write this article, Grammarly is checking everything I do and suggesting better grammar. It’s cheaper and faster than doing it myself and is probably better to boot.
Mathematics also will slowly be revolutionized. Mathematics is a stodgy field, and I say this as someone who genuinely loves mathematics. It’s still mostly done as it has been done years before, with a blackboard and chalk. Automatic proof checkers have been tried before, but math is not that rigid, despite its pretensions sometimes, and so it’s too unwieldy to use. But the AI models have already shown that they can solve mathematical problems. Now they just need to get better, and soon they’ll be on the level of undergraduates. I can’t wait for all the fascinating theorems which will arise when these models come to fruition. AlphaTensor by DeepMind has already discovered a new method of matrix multiplication, and there’s no reason that the pace will stop.
In 2084, we could have a whole new class of scientific journals, where all the results are discovered by AI, and the articles are written by AI. I think that the main issue would be just parsing and sorting through the results discovered in this way, and that this would become a whole different industry in itself.
In 2084, engineering will be radically different. There’s no reason why the techniques developed for text aren’t applicable to circuit design, hardware design or architecture. In 2084, engineers will have tools almost as smart as they are, able to suggest new designs, critique existing designs, and turn designs into products automatically. A single engineer will do the work of 20 today, and the world won’t be able to keep up with the innovation.
In 2084, all professions will start to merge into one, AI professional. Someone who can leverage these ultra powerful tools to do whatever specific industry is required of them. Rather than “journalist”, “artist”, and “engineer”, you’ll have text-model experts, diffusion model prompt engineers, and AI design integrators. The specific models used will develop experts, and whole industries will spring up around these intersections. The economy will be run more and more efficiently. It’ll be a brave new world for careers in the year 2084.