Human-led AI: 4 themes on the present and future of AI

In November last year, Faculty CEO Marc Warner delivered a talk to London’s famous Gresham College examining AI’s opportunities, potential threats, and how to frame regulation of this transformative technology.

Watch the lecture in full through the link here. Or if you’re pressed for time, read on for a digest of four key themes from the talk.

What is AI, and how does it work?

Artificial intelligence (AI) is the field of research that tries to get computers to do tasks that we think of as intelligent when carried out by humans. 

AI’s early roots date back to the middle of the last century, and the broader field that has developed since then encompasses many different techniques for achieving this goal. But the most modern and dominant approach today is machine learning – training software algorithms to learn things from existing data. In most cases, vast reams of data.

To illustrate how this works in practice, Marc shared a simplified example of how you might train an AI model to distinguish between dogs and cats. To pull this off, your model would need to be able to draw a boundary in a data set that can be used to distinguish the two animal types. 

With only a few data points to refer to, your model will make a lot of mistakes in this process. But by training your model with more data points to recognise more of the individual characteristics of both animals (eye shapes, ear shapes, fur types, etc) it will be able to draw a more accurate boundary and deliver more accurate results.

As the tasks you ask an AI model increase in complexity, so do the parameters of training. But fundamentally, the more data points you have to train a model on the more accurate its results will be.

Let’s look at another example. ChatGPT, the now famous generative AI chatbot developed by our partners OpenAI, is a large language model (LLM). An LLM is particularly useful for simulating the way language works and for generating new texts.

To do this, an LLM like ChatGPT basically follows the same data training process described for our cat and dog separator. Except, of course, we cannot explain the meaning of a particular word to an LLM. Instead we train it by feeding it enormous amounts of data. In this case, millions upon millions of written sentences from sources across the entire internet. 

The training process would then mask one of the words in a sentence and ask the LLM to make an accurate prediction on what the next word is most likely to be. In this example, the sentence might look like the below:

Not quite correct. The training then instructs your LLM to adjust its boundaries to a point that it will predict the word ‘legs’ next time. This process is then repeated millions and millions of times on all kinds of different sentences to the point where your LLM learns to predict all words very effectively.

So while AI’s results can appear near-magical at times the techniques used to get them are actually quite simple, in principle at least. The sheer volume of data and the enormous scale of compute needed to train and run these solutions is what makes them so complicated to actually implement. But the mathematical principles beneath them are relatively easier to grasp. 

‘Super intelligence’ – how smart could AI become?

How do we measure how intelligent AI is? A chess computer trained on all of the matches of Russian grandmaster Gary Kasparov might be able to beat you at chess. But it could not walk around, make itself a coffee, or distinguish a cat from a dog. Would these factors then make you more intelligent than a chess computer?

Intelligence is clearly much more multidimensional than singular tasks. So when trying to plot the relative capabilities of AI it is useful to introduce the idea of generality. Specifically, the difference between ‘narrow’ and ‘general’ intelligence.

A chess computer has high capability in a narrow environment. In this case, a single environment. Meanwhile, you are a much more general intelligence. While you may lose the chess match vs the trained AI, you are able to optimise for other goals across a very broad range of environments. On a plotted graph measuring capability vs generality, an LLM like ChatGPT might sit somewhere in the middle.

But what might sit in the top right of this graph? It’s here where many in the AI space start to consider the concept of a ‘super intelligence’ – artificial intelligence that is at least as general as human beings, but much more capable. 

It may be tempting to anthropomorphise this super intelligence as either an all-knowing god (or devil). But a more useful mental model is to think of super intelligence as more like a universal chess computer. In the same way the chess computer has a goal to checkmate the opposition as effectively as possible, the super intelligence would have goals around whatever it has been programmed with by its creators. Or that it has subsequently learned from chasing these goals.

But is a super intelligence even possible? The honest answer at this point is; we don’t know.

To date, nobody has built one. What we can say is that there is no known theory of science that prevents it. In many ways, the computers that we have today are already more capable than our own biological hardware. They can send signals faster and they can retrieve things with higher veracity. 

So at the current rate of technological development, the extraordinary claim that would demand extraordinary evidence would be to suggest that a super intelligence cannot be built, rather than that it can.

How can we constructively frame AI regulation?

The regulation of AI is one of the most pressing challenges humanity faces. But the phrasing of the matter requires some refinement to truly be useful. 

Talking about AI as a single unified field is akin to talking about physics as a single unified field. On one side of physics there are everyday tools, such as radios or microwave cookers. At the other end of the spectrum are nuclear weapons. 

If we were to talk about ‘regulating physics’ it might sound like a nonsensical conversation. But unfortunately, this is the position in which we often find ourselves in the AI debate today. At the very least, we need to start carving out two categories of AI development: the safe and the unknown.

The green ‘Safe’ zone in the above diagram represents AI technologies that have been used safely for decades in the real world. We understand all of the elements that make up these AI solutions and they are completely under our control. They typically land on the narrow generality and high capability end of the spectrum.

These safe AI solutions are being used today, for example, in NHS hospitals around the UK, helping clinical staff to make better decisions around managing bed capacity vs incoming patient demand.

The red ‘Unknown’ zone represents a set of potential technologies that we don’t currently understand well, and by extension, can’t currently control. It is sensible to be cautious when regulating these high generality, high capability AI technologies. Particularly given the natural incentives that nation states, private companies and potential bad actors may have in developing them for self-enrichment.

But it is vital that we don’t let hypothetical concerns about AI technologies that fall under the unknown category slow us down in taking advantage of the tremendous value offered by safe technologies.

Much of the world is still facing a prosperity crisis, with national GDP on a rough flat line since the 2008 financial crash. AI is one path for improving circumstances across the board. It will build great technology companies that will meaningfully change the way we live, across education, security, renewable energy, and much more.

Human-led AI and the future

Human-led AI is the approach taken here at Faculty to ensure AI remains both under control and effective in achieving its goals. In other words, powerful and trusted. 

To be considered human-led, an AI solution must be built according these three principles:

• Safe: From their foundations, every algorithm has a governance mechanism controlling its actions.

• Modular: Each component can be separately tested and understood. Technologists can build solutions with an incremental approach, making sure the systems are safe as they go.

• Human-first: Explainability is built in structurally and algorithmically. Humans can clearly understand what systems are doing and why, in such a way that they can make good choices about whether to implement or not.

The Human-led AI approach represents working within the green safe zone of proveable, reliable, narrow AI models. 

Working with the unknown red zone technologies will require other approaches to complete the picture. Some notable other frameworks in this area include Open Agency Architecture by David Dalrymple, Inverse RL by Stuart Russell, and the Beneficial AI Roadmap by Yoshua Bengio.

Building transformative AI is going to force us to decide what it means to be human. We are going to have to decide what we put into our algorithms, what we care about, and what we’re willing to fight for. 

AI transformation is a joint challenge for national governments, the business world and wider society as a whole. Doing this effectively means managing the outputs of both people and machines. So it’s critical that we build systems that augment human endeavour, not replace it entirely.

Interesting in learning more about Human-led AI and how its principles could supercharge AI transformation within your organisation? Let’s start a conversation.