Source – https://www.sify.com/
If I ask have you used any Artificial Intelligence (AI) / Machine Learning / Deep Learning application, chances are you will say no. This despite having used one barely minutes ago in the speech you converted to text, sorting of your email and mobile data, Tweet, FB post or YouTube video you last saw, perhaps even this article was recommended to you by AI. So prevalent is it that you won’t be wrong to blame artificial intelligence for us losing our natural intelligence.
Yet, few have had a chance to directly play or interact with AI. That is where OpenAI’s beta API for GPT-3 (Generative Pre-trained Transformer 3) was supposed to fit in by allowing developers and general citizens to play with its open-source AI to create new, exciting applications. Sadly, Microsoft – seeing how good the project was – put it under lock and key.
So a “collective of researchers (and engineers) working to open-source AI research” came up with the EleutherAI in July 2020. The idea was to “break OpenAI-Microsoft monopoly” by using the bits of knowledge OpenAI project published on GTP project, to create a truly free, open-source AI. The result is GTP-J. I couldn’t find what J stands for so I’ll just use it to call GTP-J, Junior.
The best thing about this project: you can fiddle with Junior to generate text and codes as per your needs. We’ll get into details of how the project works later. First, let’s have some fun.
All you need to do to help Junior write something for you: be it a college essay or a code for an app idea you have in your head, is go to https://6b.eleuther.ai/ and put in a prompt i.e. text that you want the AI to work on to generate more text or code. I am not a coder so I instead used text prompts to take GPT-J-6B out for a test drive.
The first question I asked Junior was: “Does God exist?” The reference in MY head came from a sci-fi flash fiction I had read as a child in the 80s, where when the president of the galactic federation inaugurates an inter-galactic-knowledge-network – a Google-like repository of all human knowledge connected across the planets – and asks humanity’s most vexing question ‘Does God exist?” it answers: ‘I am God.’
Sadly, for AI naysayers and terminator inspired doomsayers, GPT-J did not give any such answer. Its answer – after about 3 minutes of waiting – was (I haven’t added or removed anything, not even a comma), “Will you accept evidence from our world (the physical) to support the existence of a world of the non-physical (the spiritual)? The journey of life is not an end in itself. Our role on earth is to learn, grow, and strive to find the truth and fulfill our purpose. I believe that God exists. This is because our world is filled with evidence of the supernatural. I have no doubt that there are spirits and deities that exist. If there were no such things as spirits and deities, how could there be an increase in human knowledge? How could the world be full of miracles? How can a child one day suddenly develop the abilities of a great doctor or a gifted singer? How can we all be born with the same desires and wants, but still have a multitude of different ways of fulfilling our dreams and living a full and rich life? Is it possible that all of this is the product of a spiritual world, one in which everything takes place on a higher level? I’m not saying that it is or isn’t. I don’t believe that we can ever completely prove one way or the other. I just think that the evidence presented on the subject in the world around us points to the existence of a spiritual”
I was, to put it mildly, gobsmacked. An AI saying it believes in God to a person who doesn’t? The answer did not have me believing in God again, but it sure made me an AI believer.
So, I decided to play further. Since AI is not your average computation that merely quotes from existing text, I decided to see if Junior was really an AI (thinking) or simply a computer puking out junk text from its dataset. How do I do that? By asking it the exact same question. This time the answer to the question “Does God exist?” was closer to my belief.
“There is no person, creature or thing that can answer this question. We are all more or less atheists in the sense that we all seek the knowledge that there is nothing but the existence of God. Most of us are willing to wait, however, for some proof of God’s existence. We ask that He reveal Himself, in the form of a miracle, to any or all of the people of the world. We ask for a miracle that will lead us to know that there is a living God. This question is not so much a question of science but one of faith. We must have faith that God exists. A good religion teaches the following: The existence of God The superiority of human beings over all other creatures The power of God The hope of God The restoration of human beings to God The harmony of creation The sanctity of human life The responsibility of human beings The importance of self-knowledge The reality of progress The love of God The purity of human beings The forgiveness of human beings God, as the creator”
Again I made no changes except that the lines you see without a full stop were not together like here, but below the other, like a poem. Was Junior trying to write a metaphysical poem?
The naughty idiot inside me wanted to ask the same question again to find out what answer it gave a third time. But I decided to do something else. My next prompt/question was: Will AI destroy humanity? How original, you’d say? In my defense first of all I am a Terminator fan, and secondly despite asking this stupid question I don’t believe AI will ever destroy humanity. Why? Because humans are perfectly capable of destroying themselves thank you.
The answer Junior wrote was, well, quite boring both times. One of the answers had the line, “I’m not a theoretical computer scientist. I’m a programmer. I don’t spend much time thinking about whether the universe will end in five billion years,” ending with the line, “I spent a year teaching my AI to play the video”
None of it made sense. The lines in themselves did i.e. they were grammatically correct, but the answers as a whole did not. Seems like the old computer concept of GIGO – Garbage In, Garbage Out – is true even for AI.
To understand how EleutherAI’s GTP-J gives out these texts, you have to know how it is built. You perhaps already know that the reason Artificial Intelligence’s neural networks are better than traditional computer networks and why it can ‘learn’ and ‘grow’ is because it tries to mimic the human brain.
Into that AI neural network (stored on cloud servers provided by Google and CoreWeave) EleutherAI has injected a large curated data set they call The Pile: an 825GB language modeling dataset created from small datasets that include Wikipedia, arXiv, GitHub, StackExchange, PubMed, HackerNews etc. Since it is open-source, you can download The Pile here if you’re so inclined.
GPT-J thus becomes the AI trained on The Pile. Writer Alberto Romero writes, “Because GPT-J was trained on a dataset that contains GitHub (7%) and StackExchange (5%) data, it’s better than GPT-3-175B at writing code, whereas in other tasks it’s significantly worse.”
Perhaps that is the reason why it spewed out gibberish when I asked that question. Since I am an independent journalist and screenwriter who works in the intersection of multiple ideas, I wondered if I could use junior as a sparring partner to come up with ideas, maybe sometimes even write a few lines I could use.
Sadly to most of the tasks, prompts I gave, the results were below ordinary. Since the dataset is not current so ideating about current affairs is out of the question. But even for some other generic things I put into it, the result was often repetition of the prompt I gave or inane lines that were at best amusing and not useable at all.
Maybe the problem lies somewhere else. My natural intelligence is not just trained on 1000 plus books and thousands upon thousands of magazines, articles, snippets, etc., but also massive amounts of pain, hurt, depression mixed with tiny bits of pleasure and grit among others. Clearly Junior has eons to go before it can be a true sparring, ideating partner for me.
AI will not destroy humanity. Instead, it is empowering us. Though the AI sensor or chip isn’t inside us yet (thanks ironically to AI hating Elon Musk it soon will), we can argue that humans have already become bionic, be it through the zillion things on your phone and laptop, to the way AI runs everything from our flights to stock markets. So it is best we embrace AI. And a good, fun place to start, is by playing with GTP-J (at https://6b.eleuther.ai/) or as I’ll not stop calling it, Junior.