There’s plenty of room on the LILO
Photo credit: Tony Cuenca via Unsplash
What are LLMs good for? Everything, if you believe marketing announcements and press releases. Very little, if you believe the most sceptical of sceptics. It sometimes seems that we are still waiting for the crowning innovation, the killer app that will secure the place of LLMs in our enterprise – or the deflating moment when we realise that our expectations cannot be met.
While we are waiting for that moment, I think there is plenty for us to do by putting LLMs to work in relatively mundane contexts. What are they good for? There’s a clue in the name: Large Language Models are good at language, something which traditional computing has been notoriously bad at for most of its existence.
If traditional computing was any good at language, we would spend a lot less time filling in online forms, pushing buttons, repeating ourselves to automated help lines, and otherwise behaving in ways which are convenient for machines rather than convenient for us.
I believe that most enterprises are rich in what we could term Language-In-Language-Out (LILO) problems: situations in which the organisation needs to figure out what the end user wants, and needs to give them an answer in language which they can understand. Sometimes these situations manifest in a phone call, and sometimes they manifest in a letter or an email. Sometimes they may simply manifest in someone reading the information on your website and trying to make sense of it. Frequently, they overwhelm the capacity of the organisation to respond.
I think that there are two types of work that we need to do to address LILO problems.
The first is application and integration. It might seem that we could just drop a chat interface over existing processes and interfaces. But those processes and interfaces have been designed over years to work in a certain way: to interact with structured data, not language. Integrating LLMs, rather than simply creating a language-like veneer, is hard work. As in most technology implementations, deploying the new capability is not hard – but integrating it with everything else is. And that’s before we even start to think about the implications for business processes, operating models, training and culture.
The second type of work we need to do is to ensure that we are deploying language thoughtfully and responsibly. Language is an astonishing superpower: it is one of the things that has made humans the dominant species on the planet. And when we use this superpower, when we communicate using language, we make implicit promises about sincerity, truth and trust. LLMs are not designed to keep these promises: they are designed to be plausible and pleasing. If we want them to communicate on our behalf, we must develop and deploy the controls and methods to prevent them breaking our promises for us.
The frontier of AI products will continue to move: companies will keep innovating, adding new features and new capabilities to their models and products. We should continue to observe them with interest and curiosity. But there is plenty of work to do on this side of the frontier. Traditional computing has been around since the middle of the last century, and we are a long way from exhausting its potential. LLMs (in a commercially usable form) have only been around for a few years, and we have barely begun to figure out what we can do with their facility for language.
While waiting to see what the next wave of innovation brings, I think we should get on with applying LLMs in simple and mundane ways, doing the hard to work to add it to interfaces in contexts where it can be trusted, while developing the practices and controls which allow us to be more ambitious, safely. Diligent, patient work on these problems holds out the hope that we can teach machines to behave in ways which are convenient for us, rather than machines teaching us to behave in ways which are convenient for them.
There are plenty of language problems to solve – and plenty of room on the LILO.
