Since the introduction of ChatGPT last Nov, Artificial Intelligence (AI) development has exploded everywhere in our faces. In a short span of months, its rapid evolution has caused some like Musk to ask for a moratorium or to pause its growth for 6 months. It is simply not possible to stop AI’s progress, as we will explore today.
ChatGPT 3.5 is an LLM (Large Language Model) that was introduced by OpenAI to the world less than 6 months ago. GPT stands for Generative Pre-Trained Transformer and this version had at least 175 Billion parameters that were pre-trained with data up to 2021. Before everyone could get fully overwhelmed and amazed by it, ChapGPT 4 was launched in Mar 2023. It is now multimodal (able to take images and text as input) and even more powerful than anything we have seen before. Don’t you feel that Google and Youtube searches suddenly became dumber overnight? I certainly do.
How does ChatGPT work? Basically, it analyzes and breaks down the input prompt from the user word by word via a process called tokenization. It then searches its pre-trained database for words that can be used for an answer. Thereafter, it ranks the list of words by probability from the highest to the lowest to prepare for the response. Using NLP (Natural Language Processing) techniques, an answer is formed and presented to the user. There is also a level of probability randomness built into this process in order to ensure that a slightly different answer is presented each time. For image input, a similar process is utilized.
In summary, ChatGPT works by attempting to understand your prompt and then spitting out a string of words that it predicts with the best answer to your question, based on the data it was trained on. ChatGPT 4 is pre-trained on an even bigger dataset and is able to recognise images versus version 3.5. It seems like 5.0 is likely to be launched before the end of the year as the AI arms war heats up. Google and Chinese tech firms are racing to match or exceed it to maintain their market share and industry competitiveness.
Based on how they work now, the AI we see is actually only ANI (Artificial Narrow Intelligence) which is good at “TALKing”/generating written and image responses to human prompts. What most fear is AGI (Artificial General Intelligence) which is when the machine is able to THINK for itself. We are far from there and the Terminator is not coming any time soon.
The other area of interest is Generative AI. With a simple prompt, the app is able to generate photo-realistic images that are improving constantly. OpenAI started with DALL-E and introduced DALL-E 2 in Jul 2022. Within months, powerful systems like Mid-Journey and Stable Diffusion stormed the world with images that are only limited by one’s written imagination.
These apps use neural network concepts like the human brain to initially train themselves to analyze and recognise images of animals and humans. Then with more sophisticated processes like GAN (Generative Adversarial Network) and CNN (Convolutional Neural Network) using a police-against-thief setup, they are able to generate even more impressive images.
In recent times, we see unreal viral images of the Pope togged in designer Balenciaga wear. With easily accessible animation and voice generation software, it is now possible to combine with ChatGPT-generated scripts to create videos that were not possible just a year ago. Having Elon Musk discuss AI with Steve Jobs? Sure!
Generative AI still has some buggy issues like having 6 fingers on a hand but the latest versions of Mid-Journey 5.1 that was just released seem to have resolved that. More such images will probably win photo competitions in the near future as it becomes harder and harder to distinguish the fakes from the real ones. Their improvements within months are short of amazing when you view the new ones created every minute on their Discord channels.
Another reason why AI is unstoppable. There are almost an unlimited amount of unstructured data being created now. They are available to be mined by AI models as training datasets. Approximately 90% of all dark web data had only just been created in the last few years alone. According to the latest estimates, 329 million terabytes of data are being created every day. Data is the new “oil” for AI. The more it can analyze, the better it gets and there is no shortage of data supply in the foreseeable future.
A recent article can articulate this data explosion phenomenon much better. The Bloomberg title was “The Future of AI Relies on a High School Teacher’s Free Database”. A German teacher decided to create a dataset to help train image-to-text diffusion models. He felt that the info should not be controlled by a few companies but instead be open source. He started a non-manual automated process that crawls the internet to locate the images on the web and associate them with descriptive text.
As of now, this freely available dataset has 5 billion images for anyone to use to train their AI models. His company is called LAION (Large-scale AI Open Network) and it has enabled firms like Stable Diffusion to offer a superior Generative AI tool to create amazing images just by using written prompts. Anyone in the world with determination like him can do so and create their own dataset. The sky is the limit to creating a new dataset and no one can stop you.
Authorities have also been scrambling to try to make sense of AI and look to form a framework of sorts for it. Europe is in advanced discussions on this topic for their EU AI Act while America is still trying to figure it out. The EU proposal is slowly moving through its approval process but it is unlikely to gather strong support like its previous GDPR data protection regulation.
This is because it is humanly unenforceable. How can one ensure that the huge amount of data available on the web to be harvested and used by the AI model has no copyright data? Even if the companies are made to disclose the materials of the datasets used and then asked to remove some questionable data, they are unable to untrain the models. The neural network of the AI model is already a black box like the human brain. Once the data is fed in, one cannot be sure what sort of output it will produce. Hence we are constantly amazed at what generative AI models can now produce.
There is also the fear that these new AI regulatory frameworks may hamper the growth of this sector in that country. Hence they may fall behind in innovation because of these unclear/unenforceable directives. Would they want China to win this AI arms race? Probably not.
Elon Musk wanted a 6 months pause in all AI development initially. Then at the same time, he selfishly started a new AI firm to race to catch up with the competition. Anyone in the world with an internet connection is able to develop his own dataset to train a new AI model. ChatGPT has spawned newer and better uses every day. Someone on Github just created AutoGPT, a turbocharged version of ChatGPT. It is an AI agent that can be tasked with a main goal broken down into sub-tasks with an API connection to ChatGPT which goes into an autonomous loop in order to optimize a final solution.
Thousands of new AI uses are being formed every day. Many are seeking to monetize it while accelerating job efficiency and productivity. For example, the turnaround time and lifecycle creation of data packages like training modules and marketing strategy presentations have shortened considerably to hours instead of days.
Just a few days ago, Google launched the latest AI salvo that promised to be the ChatGPT killer. The new improved and upgraded Bard can now do much more now after its less-than-ideal launch in Feb. Every mother, father and son tech company is rushing to do a catch-up to launch the next new AI tool.
The advent of AI cannot be slowed down or stopped. The Pandora box has been opened and the pace will accelerate. As Thanos had said before: “I am INEVITABLE”.
Leave a Reply