Like most people I am impressed at what I see the current wave of chatbots like ChatGPT with their large language models, LLMs, can do. They are ubiquitous and the use of the expression artificial intelligence, AI, has almost become synonymous with them. AI is everywhere and is much more than a chatbot. GPT stands for generative pretrained transformer, which although able to generate novel text, is limited to knowledge that it is trained on.
However I am not amazed and I do not see it as the greatest thing since sliced bread (do people still use that expression?) I’ve described them as a great search engine with a fantastic natural language user interface and when I examine their output, the more that is what I think.
More than that, I see it as a modern-day equivalent of Eliza, the first example of its kind. Created by Joseph Weizenbaum at MIT. And yes named after Eliza Doolittle. The reaction to Eliza, given the technology at the time, was just as jaw-dropping as many of the reactions today. “Many academics believed that the program would be able to positively influence the lives of many people.”
Eliza came out in 1964. Put that in perspective, only 4 years after the first functional integrated circuit consisting of 2 transistors and a few non active components. And 7 years before the first central processing unit CPU, the heart of modern computers, (a modern CPU can contain over 100 billion transistors). So Eliza was the state of the art for the art of the time.
(First Integrated Circuit) (Modern CPU)
At the time Eliza was believed to have understanding and emotions like feelings. Such was the novelty of a human actually interacting with a machine that seemed to respond just like a human to your discussion. Just like modern chatbots seem to have understanding and feelings to some users. Or even consciousness and some sort of agency. Even researchers became attached to it as if it were a human.
There were memory and processing power and processing speed constraints that do not exist today. Memory for instance used to be the critical bottleneck and expense factor in early personal computers. I studied Boolean geometry just so I could reshape circuit designs to optimize and minimize the use of integrated circuits and achieve the maximum use of limited and expensive memory. Today I doubt that many know who George Boole was. Memory today is unlimited. Gigabytes everywhere. The Apollo lunar lander had a tiny computer with a few kilobytes of memory!
So within those restrictions, the world of knowledge that Eliza could refer to in a conversation was restricted. For instance to avoid having to load it up with lots of real-world knowledge, an early implementation limited the conversation to psychotherapy. Such a therapist rarely has knowledge of the patient’s background, their real world history, at least at the beginning of a session, and has to use tricks to get to that. With Eliza that allowed tricks like repeating a user’s question back to them. “My mother takes care of me.” “Why do you say that your mother takes care of you?” Or reversing the question: “Do you have family?” “What makes you say that?”
A main part of Eliza were scripts on how to respond to questions. That was necessary since the real-world knowledge was limited. It is not unlike the scripts that call centre marketers use to respond to all possible responses and questions coming from whoever they have cold called. If you say what they are selling is too expensive, they have a script to reply back to you, to keep you involved.
The user interface was primitive by today’s standards. The standard was to interact with a teletype with a responding printout of the answer.
Keywords were applied. Keywords were preidentified in the code and allocated levels of importance in what was called a “keystack”. They triggered the focus of a question and the appropriate response. Words like “mother”, “same”, “like”. The Transformer part of GPT does essentially the same thing. A different and more effective way of course with modern computing technology, but the same idea.
The difference today is one of scale in my view and not about knowledge understanding by AI systems. The current chatbots have access to the cloud which has access to everything that has ever been published. So are not limited to small real-world domains like a psychotherapists chair. In a chatbot the system is then trained on what they can get access to on the web/cloud, and then that is the limit of its world knowledge forever (or until it is retrained and rereleased as the next version).
And no need for a teletype. It would be 30 years after Eliza before natural language processing appeared. So now the interface with a modern chatbot uses the same advanced English (or other) written and spoken conversation modes as say talking to your cellphone. I am as much impressed with modern natural language processing as anything else in AI.
So to my mind, a modern chatbot is a massively effective search engine (able to search everything in its training) with an amazing user interface, but otherwise much like Eliza.
The same awe of the modern chatbots is seen in many acolytes; a modern version of the “they will change the world scenario”. Sometimes apparently going too far.
Late last year MIT published “Artificial Intelligence, Scientific Discovery, and Product Innovation” by PhD student Aidan Toner-Rodgers about the impact of AI on the workforce. The finding was that scientists aided by AI tools were considerably more productive than their peers working without the technological aid. Last Friday, MIT announced that it reviewed the paper following concerns and determined that it should be “withdrawn from public discourse.” From complaints that the paper did not stand up to scrutiny, MIT conducted a review process stating that MIT “has no confidence in the provenance, reliability or validity of the data and has no confidence in the veracity of the research contained in the paper.” The author is no longer at MIT. See the link below.
Eliza today is considered revolutionary, extraordinary, pioneering. Interestingly Weizenbaum was concerned about the use of AI. He was concerned about the duality of “decision” and choice, the first a computational task suited for computers, the second a matter of judgement suited to humans. In my view, like all emerging technologies, beware.
MIT Backs Away From Paper Claiming Scientists Make More Discoveries with AI
AI is the next step in the Human evolution, AI is the only way that Human intellect will ever reach the stars.
AI isn’t humanity though. It’s a shallow imitation – literally soulless.
Again I agree. We have no idea what a soul is, whether there is such a thing, and even to understand what different people mean when they use the word. “That band has soul.” So there is no way to build it into a machine, even if it were possible. Cannot create something that is as of yet unknown. Most people I think treat “soul” as a spiritual thing, maybe God-given, something that separates us from the animals – and the machines.
I think there is truth in that Arthur. The time to travel long distances interspace are so vast that it is nigh impossible for humans I think. Ideas around sending colonies on space ships that are living quarters for generations are fantastical. So if our intellect is installed in AI machines, and that is sent out, does not matter how long. AI is more than a chatbot though as I say. It includes robotics with all the features such as vision, speech, reasoning, planning, inventing and so on.