By Richard Anderson / Co-Founder, Nuon AI
Remember the first time you saw Google Maps satellite view? Using ChatGPT for the first time was similar – you immediately know that this is something special.
Technology doesn’t smoothly advance, although it sometimes feels that way. Ray Kurzweil describes it as a series of S-curves. The arrival of smartphones and social media was probably the most recent steep section of the curve when innovations came thick and fast. Since then there’s been considerable refinement and consolidation, but less innovation. ChatGPT feels like the start of the next phase of rapid innovation.
Over the last few months, it seems that not a day has gone by without ChatGPT hitting the headlines. From the recently announced UK competitions authority’s review of AI models, to the tool finally answering the age-old question of whether you should put jam or cream on your scone first, it’s clear that AI has hit the mainstream.
As an AI company ourselves, we have watched the rise and rise of ChatGPT with interest. And of course, have kicked its tyres to see what it can do.
Being a developer, the first thing I asked ChatGPT for was some code. A class in Java to represent a matrix. Bingo, there it was. But you could google that, so no big deal. “Write me a unit test for that, could you?”: bingo again, there it was. That’s more impressive. But wait, there was no multiply method in the matrix class. “Add a multiply method to that matrix class”. Of course, it did. “Add a unit test for it too.”… and it did.
Two things impressed me. Firstly, the code was pretty good. I’ve certainly seen junior, and even more senior, developers do a worse job. Of course, this is a pretty simple test. Maybe the kind of test you’d place in front of a candidate for a job interview. Which is interesting in itself. The second thing is more impressive to me, the references back to the context of the conversation. I’ve not encountered anything in the past that billed itself as an AI chatbot which did anything close to this.
Pair programming with an AI anyone?
Then I thought to try something a bit less techy: “Give me a few paragraphs on the Mad Hatter’s tea party as an allegory for British imperialism” – a pretty clever question I thought. The answer related the Mad Hatter’s condescending indifference to his guests to the indifference shown by the British to the suffering brought upon the people of the empire. And also how the sense of monotony and timeless repetition felt by the guests at the party reflected the effect of industrialisation on working people.
Maybe that’s a subject which is well written about, but I was slightly stunned.
A deeper look at the tech behind ChatGPT
If you’re reading this blog, you’ve probably read others on this subject. ChatGPT is a Generative AI. It is built on top of Open AI’s large language models (GPT3 specifically). At a really fundamental level, it can be thought of as being similar to a predictive text system.
So, if ChatGPT is based on Open AI’s GPT3 model then, you might ask, what happened to GPT1 & GPT2? That reveals something quite interesting.
Probably the most significant difference between GPTs 1, 2 & 3 is the number of parameters (a parameter being an element of the model which is learnt from historical data sets). GPT1: had 117 million parameters, GPT2: had 1.5 billion parameters, GPT3: has 175 billion parameters, and GPT4?: 100 trillion.
What Open AI (and indeed others) have shown is that there is a relationship between how realistic a conversational AI is and the number of parameters. Emergent behaviours, for example ChatGPT4’s theory of mind capabilities, have also started to show up as the number of parameters increases.
The point is that this technology scales, and there is no indication that we are nearing a limit.
Looking at both ChatGPT and Nuon AI, they are both fundamentally predictive. The most significant difference is that Nuon’s AI learns on the job, it isn’t pre-trained, it learns in real-time. GPT stands for Generative Pre-trained Transformer, so the clue is right there in the name: Pre-trained. Also, Nuon’s AI hasn’t yet benefited from $10b of investment from Microsoft.
Our short-term roadmap includes more real-time products, so it’s probably unlikely that a Generative AI would be directly useful. But you can certainly see applications for it in the longer term as we move toward our north star of bringing capital and risk ever closer together.
What are the risks…or dangers?
People tend to fall into two camps. Those like Bill Gates, Elon Musk, and Geoffery Hinton worry about the AI future, while people like Ray Kersweil are much more optimistic. Interestingly, both groups agree on what AI will be capable of doing. But, those on the worrying side fear a future where humans have given the AI the wrong objectives. Like Aladdin’s Genie, you need to be careful what you wish for. The optimists’ camp sees AI very quickly solving problems which humans find intractable.
But that’s longer term. Though it will be this decade if Ray Kerzweil’s predictions are correct, and he has a good track record! In the short term, what risks do ChatGPT-like technologies present today?
Imagine a conversational AI that’s assisting you every day. You use it many times a day, as you would use Google now. It knows you, it talks to you (a spoken interface is a small step away from where we are today). You’ll spend much less time searching for information, so it’s good right? But, could it persuade you? After all, one of the best ways to persuade anyone of a viewpoint is through discussion – in person, one-to-one. The AI finds itself in a perfect position to persuade individuals in very highly targeted ways. In fact, specifically targeted at you.
Now roll the calendar back to the dawn of social media. As people rushed to join their friends and family to share news and pictures, few anticipated how effective it would be at dividing people and spreading misinformation. The technologies of organisations like Cambridge Analytica are a crude scatter gun compared to a conversational AI which can adjust its arguments in real-time as it detects which argument is most effective with you specifically, and also learns from conversations it is having with millions of other people at the same instant.
Fad, or real change?
The answer to that question comes back to the S curve. Whether ChatGPT is the game changer is open to question. Competition in the space is red-hot. New and more capable systems will emerge whether from Google, Open AI or some other player.
Clearly though, it has already changed the landscape. There was a time before the advent of touch screens, or trackpads and mice, when interacting with a computer was very different. At the very least being able to hold a conversation with a computer represents a sea change.