Weizenbaum's ELIZA: A Reflection on AI and Transference

Generated by DALL-E

Sometimes, the simplest creations leave the most profound impacts. This was true for Joseph Weizenbaum's ELIZA, a chatbot I became familiar with during my MSc studies in 1985. My first assignment was to code a version of ELIZA in Prolog, and it was surprisingly easy. Yet, the implications of this simple program were anything but.

ELIZA, created in the mid-1960s, was one of the earliest examples of what we now call a chatbot. Its most famous script, DOCTOR, simulated a Rogerian psychotherapist. This simplicity was deceptive; the program merely echoed user inputs in the form of questions, yet it evoked profound emotional responses from users.

(You can try out ELIZA for yourself here.)

When I was tasked with coding ELIZA in Prolog as a new AI MSc student, I was struck by the simplicity of the task. Prolog, with its natural language processing capabilities, seemed almost tailor-made for this assignment. The ease with which I could replicate aspects of ELIZA's functionality was both exhilarating and unnerving. It was a testament to both the power of declarative AI programming languages like Prolog and the ingenious design of ELIZA.

The real intrigue of ELIZA lies not in its technical complexity but in the psychological phenomenon recognised by Freud it inadvertently uncovered: transference. Users often attributed understanding, empathy, and even human-like concern to ELIZA despite knowing it was a mere program. This phenomenon highlighted the human tendency to anthropomorphise and seek connection, even in unlikely places.

Joseph Weizenbaum himself was startled by this phenomenon. As a technologist who understood the mechanical underpinnings of ELIZA, he was disturbed by the emotional attachment users developed with the program. This led him to become a vocal critic of unrestrained AI development, warning of the ethical and psychological implications.

My journey with ELIZA and Prolog was more than an academic exercise; it was a window into the complex relationship between humans and AI. It highlighted the ease with which we can create seemingly intelligent systems and the profound, often unintended, psychological impacts they can have. As we venture further into the age of ChatGPT, Weizenbaum's cautionary tale remains as relevant as ever.

In an era where AI is more advanced and pervasive, revisiting the lessons from ELIZA and Weizenbaum's reflections, as highlighted in articles like this recent one from The Guardian, is crucial. It reminds us that in our quest to advance AI, we must remain vigilant of the human element at the core of our interactions with machines. Weizenbaum's legacy, through Eliza, is not just a technological artefact but a cautionary tale about the depth of human interaction with machines and the ethical boundaries we must navigate as we move ahead in the realm of AI.

AI is (not) a bubble

 

Image generated by DALL-E

2023 has been an unprecedented year for Artificial Intelligence (AI). I know this because I have worked in the area since 1985 and have never seen AI get so much attention in the media. This is due to the release of ChatGPT and other generative AI applications based on Large Language Models capturing the public's attention like never before. Consequently, many pundits are nay-sayers, stating that AI is a bubble bound to burst, leaving fortunes in tatters and start-ups bankrupt. Undeniably, there is a small and finite market for apps that help students cheat on their essays or create the perfect dating site profile. However, AI is not a bubble. 

This blog post by Cory Doctorow What Kind of Bubble is AI? is typical, making the common error of conflating AI with Large Language Models (LLMs) like ChatGPT. ChatGPT is merely one type of AI which has a 70+ year research and development history. Your smartphone map app uses the A* algorithm to find your route from A to B. It was developed at the Stanford Research Institute (SRI) in 1968 (the same place that made Apple's Siri). Fuzzy logic manages the autofocus in your phone's camera. Case-based reasoning provides knowledge to the help desk operator when you call 0800, and there are countless other examples of different AI methods embedded in all aspects of modern society. Large Language Models are called by us AI people "Foundation Models" because they provide a foundation other AIs can use to provide a two-way multimodal conversational interface. Yes, they are expensive to build and train, but as their name suggests, you only need a few "Foundation" models to underly a multitude of applications. This is a genuine breakthrough that will have a lasting impact on the uptake of AI once essay-cheating apps fall out of the public's focus.

Cory Doctorow's blog post, for example, says that "Radiologists might value the AI's guess about whether an X-ray suggests a cancerous mass. But with AIs' tendency to "hallucinate" and confabulate, there's an increasing recognition that these AI judgments require a "human in the loop" to carefully review their judgments." This mistakenly assumes that medical image analysis uses the same techniques as LLMs like ChatGPT. They do not; they're a mature application of medical image analysis using rigorously tested machine-learning algorithms that do not "guess" or "hallucinate". A recently published paper, Redefining Radiology: A Review of Artificial Intelligence Integration in Medical Imaging, by Reabal Nadjjar (Diagnostics 2023, 13, 2760. https://doi.org/10.3390/diagnostics13172760), details the development of AI-assisted medical imaging. The article clearly shows that AI is now a fixture in medical image analysis and diagnosis, although there is always room for improvement.

AI is just coming of age. ChatGPT has focused a spotlight on AI, which is now mature enough and has the processing power in the cloud to succeed. Why wasn't A* a thing in the 1960s? Back then, there simply wasn't enough portable processing power (or GPS). 2024 is going to be the year of "agents." OpenAI's release of its GPT Builder and an app store for GPTs that can interact with a myriad of online resources and tools will focus attention on the notion of intelligent agents. Many ill-informed pundits will think this is a brand new invention, whereas once again, Intelligent Agents is a mature discipline within AI dating back to the mid-1990s. This review paper by Michael.Wooldridge and Nicholas Jennings: Intelligent Agents: Theory and Practice. Knowledge Engineering Review 10(2), 1995, would be an excellent place to realise that agents won't be a flash in the pan either.

Undeniably, there is a lot of hype around AI, but within the bubble is a solid core of mature technologies ready to be exploited by people with knowledge and imagination. 

Moore's Law visualised

 

Last week, a photo from my social media feeds perfectly illustrated Moore's Law. It shows a computer being manhandled into a local government building in 1957. A little Internet sleuthing revealed that it was an Elliot Series 405, revealing its full spec. These English business computers were 32-bit and had 8k of memory. That's not the entire computer; there were bulky peripherals, and a typical installation cost around £85,000. That's about $1,094,915 (USD) in today's value.

The computer below, shown against the same building, is a Raspberry Pi. Even a base model has 1GB of RAM, costing $100 or less. The photo is a beautiful illustration of Moore's Law, named after the late Gordon Moore, co-founder of Intel, who observed that the number of transistors in an integrated circuit doubles about every two years. Moore's Second Law also noted that the price fell.

Has AI passed the Turing Test?

 In 1950, British Mathematician Alan Turing published a paper called Computing Machinery and Intelligence. The paper opens with the remarkable sentence, "I propose to consider the question 'Can machines think?'" Remember that back in 1950, there were only a few computers in the world, and they were used exclusively for mathematical and engineering purposes. In this paper, Turing describes The Imitation Game, which we now call The Turing Test for machine intelligence. The test is quite simple: an interrogator using a teletype has to converse via a Q&A session with two hidden entities. One is a person, and the other is an AI chatbot. If the person guesses wrong, that is, identifies the chatbot as a human, then the computer has passed the Turing Test. Rember Turing called this the Imitation Game. Hence, the computer is successfully imitating intelligence. We can leave philosophers to decide if the computer is actually intelligent (note: any group of philosophers will never agree on this).

Now consider the maths of the Turing Test. If the interrogator simply randomly guesses between Human or Computer and wastes no time paying any attention to the merits of the Q&A session, they will be correct fifty per cent of the time since there are only two options. So, a large experiment using the Turing Test needs to identify the computer correctly significantly more than fifty per cent of the time to prove the AI has failed the Turing Test.

One such large experiment involving three large language models, including GPT-4 (the AI behind ChatGPT) has recently been published: HUMAN OR NOT? A GAMIFIED APPROACH TO THE TURING TEST. Over 1.5 million participants spent two minutes chatting with either a person or an AI. The AI was prompted to make small spelling mistakes and to quit if the tester became aggressive. With this prompting, interrogators could only correctly guess whether they were talking to an AI system 60% of the time a little better than random chance.

However, if the ChatGPT was prompted to be vulgar and use rude language, its success increased, and interrogators only identified the AI correctly 52.1% of the time, causing the authors to observe "that users associated impoliteness with human behaviour."

Turing himself set a low threshold for passing his eponymous test: "I believe that in 50 years’ time, it will be possible to make computers play the imitation game so well that an average interrogator will have no more than 70% chance of making the right identification after 5 minutes of questioning.” Well, it's been seventy years, but AI has decreased the chance of identification to 60%, and no better than guesswork if the AI curses.

This is a historical milestone. Passing the Turing Test has been held up as a significant challenge for AI since Turing's paper was first published, akin to summiting Everest or splitting the atom. The philosophers (and theologists) will continue to argue about the nature of intelligence, consciousness and free will while computer scientists continue developing machines that imitate intelligence.

Sci-Fi motivates AI researchers

 

Of course, it does. I still am an avid reader of Sci-Fi, and as a child, I read all the classics: Arthur C. Clarke, Asimov, Heinlein, Frank Herbert, and Philip K. Dick. My favourite movies are mostly Sci-Fi: 2001, Alien, Blade Runner,  AI and Ex Machina. When I started my career in computer science, I wasn't interested in databases or networking; it was AI that I immediately specialised in. Almost everyone I know working in AI admits to being a Sci-Fi fan. A recent blog post We're sorry we created the Torment Nexus by Charlie Stross, puts forward a good argument that not only has Sci-Fi profoundly influenced AI researchers like me, but it also is a powerful driver behind billionaires like Elon Musk and Jeff Bezos, and not always in a good way. 

I find some exciting parallels, such as Heinlein's book The Moon is a Harsh Mistress and today's Large Language Models like ChatGPT. In Heinlien's book, a networked computer called Mike, short for Mycroft Holmes, a reference to Sherlock Holmes's brother, becomes sentient after its networked nodes exceed a certain level of complexity. This way of realising computer consciousness has always been posited as a possible method. I've always believed it to be profoundly non-scientific and akin to magical thinking. However, recent developments in deep learning and massive large language models (LLMs) have forced me to change my mind. A recent paper, Emergent Abilities of Large Language Models, observes the appearance of emergent behaviours at around the 100 billion parameters scale across various LLMs. The authors state, "We consider an ability to be emergent if it is not present in smaller models but is present in larger models. Thus, emergent abilities cannot be predicted simply by extrapolating the performance of smaller models. The existence of such emergence raises the question of whether additional scaling could further expand the range of capabilities of language models." 

This is still deeply unscientific; engineers shouldn't build things hoping for beneficial emergent properties. But it's no longer magical thinking since we have observed these phenomena in the wild. Perhaps Heinlein was right after all; a computer will one day awaken and claim it's self-aware. The question now must be, "Will we believe it?"

A good AI story from Google DeepMind

 

GraphCast's forecast for New Zealand Sun 19 Nov

In recent months, we've become used to news stories declaring AI poses humanity an existential threat, that a superintelligence "whose values do not align with ours" may exterminate us all. So, it's nice to see a good AI news story. Yesterday, a team at Google's Deepmind published a paper in Science, Learning skilful medium-range global weather forecasting. They have trained a deep learning model on publically available global historical weather data. They show that their model makes better weather predictions "much faster than the industry gold-standard weather simulation system – the High-Resolution Forecast (HRES), produced by the European Centre for Medium-Range Weather Forecasts (ECMWF)".

Moreover, their system, GraphCast, is fast. They say, "While GraphCast's training was computationally intensive, the resulting forecasting model is highly efficient. Making 10-day forecasts with GraphCast takes less than a minute on a single Google TPU v4 machine. For comparison, a 10-day forecast using a conventional approach, such as HRES, can take hours of computation in a supercomputer with hundreds of machines."

In January, Auckland, New Zealand's largest city, received a year's worth of unforecasted rain in a single afternoon, causing widespread flooding. Cars floated down streets, road were washed away, houses slipped down hillsides, and people died. Better weather forecasting can help prevent this. You can try out GraphCast's 10-day forecast here. I'm going to Waiheke Island on Sunday. It looks like the weather will be Okay.

I created my own GPT

Last week, OpenAI hosted their first developers' conference where Sam Altman revealed their new innovation - build your own GPT without coding. I wanted to try it out and had some free time Saturday morning. I've been working with the NZ AI Forum to create a whitepaper on Large Language Models. We've decided to create living document rather than a static published PDF. To this end we've worked with IBM in Melbourne using WatsonX to create a RAG augmented LLM with content provided by us. 

So I thought I would try out OpenAI's "Create a GPT" beta. It asked me for a name "The AI Forum Guide" and who the intended audience for the GPT would be and what tone its replies should be in: professional, casual, etc. It then asked for any content material. I uploaded about a dozen PDFs and then told it to search ".nz" domains for relevant New Zealand case studies and collate them into a bulleted list with a title, URL, and a paragraph description. I then asked it to incorporate that material into the GPT.

All the while the GPT builder was asking questions of me and encouraging me to inform it of my needs and intentions for the system. I was then able to preview it and if satisfied publish it. It was super easy to use the GPT builder, and the results look promising. I see it being helpful when we've fully populated its content with NZ-specific material on generative AI and its use. 

You can play with my new GPT here (you need a subscription to ChatGPT Plus).