• Paul Henninger, Partner |
5 min read

Large language models (LLMs) like ChatGPT aren’t just an incremental technology improvement. They’re the most important innovation since fire or microprocessors or the internet itself.

If that sounds like hype, consider the fact that in conventional data science of around ten years ago, to have more than 100 parameters in a model was usually problematic. At 200, it became unworkable. The model simply confused itself. The biggest LLMs to date have nearly one trillion parameters. Let that sink in…

It’s a bit like coming up with a new design for an aircraft and giving it 75,000 wings. That sounds like lunacy. But it turns out that not only does the design work and it means the plane can fly with high efficiency, it means it can teleport too.

The breakthrough in LLMs has come about because data scientists have completely rethought models and how they store information about the progress they’re making when they learn. In fact, there are things about how they store all the information they gather that we still don’t fully understand, but that will come in time.

LLM use cases

So where will all this take us? It’s a question I was delighted to participate in a panel discussion about at this year’s London Tech Week, where there was a huge level of interest in generative AI.

It’s important to appreciate that ‘generative AI’ as it’s popularly called is not just generative, as important as that is. The ability to generate content - writing documents, poems, stories, code, creating images, or almost anything else we want it to – opens up huge possibilities.

But LLMs are also ‘comparative AI’. That is to say, they are incredibly powerful at comparing documents, images or any kind of source information to analyse the differences and draw conclusions for us. This has immediate commercial applications. For example, imagine a company that wants to develop the best possible coffee machine. An LLM could compare 20 different designs and the documentation about them (or 100, or 1 million), evaluate them and make recommendations about what the new design should look like. You could apply this to almost any design, engineering, risk management or manufacturing problem.

Then there is what I call ‘expansive AI’. This is where the LLM takes in vast expanses of information to find specific things in an incredibly short period of time. Take lawsuits. In the US, companies defending class actions will sometimes dump huge swathes of information in the discovery process, essentially to overwhelm the opposition. They may also call a ‘surprise’ witness at short notice to wrongfoot the other side – who then only has hours to trawl through all the evidence to find information relevant to that individual. Law firms may employ 200-500 people to sift through the evidence in preparation and take weeks doing it. But with LLMs, it can find everything within minutes. The law firm can probably cut their human team down to a handful, supplemented by the LLM. The LLM can not only pinpoint every named reference to an individual, but identify parts of the evidence that are relevant even where the individual is not actually named; it can also search the evidence according to specific lines of argument and highlight the parts relevant to that.

These kind of advances aren’t just ‘productivity increases’. They’re new capabilities that completely change the game. 

Integrating AI and human in the future

There are challenges, however. AI advances could enable adversaries and bad actors too, including the generation of harmful or fake content. But what I see as the biggest challenge is how we can integrate the speed and rate of information of AI with the speed at which humans do things. It’s like light speed compared to walking pace.

As an example, early iterations of AI have had the ability to look at MRI scans and predict for cancer with greater accuracy than clinicians for probably 15 years. But we haven’t really realised the potential from this capability fully over that time. That’s because medicine and caring for patients is about much more than interpreting images. There is a whole care pathway, human and emotional issues, understanding the patient’s state of mind, their situation and life circumstances, etcetera.

As a result, we haven’t really yet defined the optimal way of using AI in medicine. The same applies everywhere else. We now need work out how to integrate the two processes – the two radically different speeds – in parallel to get the best and most useful outcomes.

Right now, LLMs are functioning mainly as individual assistants where we all more or less separately ask it to do things that will help us whether that’s a work task or something outside work. We experiment and see where the value lies. This will then graduate to LLMs as enterprise assistants – assisting whole departments or functions in specific ways that have become acknowledged use cases on a systematic, organised basis. This is where it could radically disrupt and change how we work and what we as humans spend our time doing. The next stage would be LLMs as cognitive assistants – where the LLM is so intelligent that we press the button in the morning and the LLM does… who knows quite what as yet.

Business needs to get started

My message for businesses is – don’t ignore this. Explore. Get started. There are highly valuable potential use cases for just about any business in any sector. Think about what your key business objectives and strategic aims are – and where you’re struggling. Articulate your risk appetite. Map out your processes within your operating model and value chain to identify the areas that are ripest for value creation – where are the ‘green flags’? Then, partner with a technology provider or advisor to start prototyping some LLM-based solutions. Think carefully about your data strategy and which models you expose what data to.

It won’t be long before everyone is doing this. So get started now and get ahead (at least for now).

The Chat conversation is only just beginning – we may be speaking in a completely different language in just a matter of years!