Are LLMs the air fryers of AI?

Do you know someone who got an air fryer for Christmas? Or did you get one yourself?

If you know someone who got an air fryer, then there’s a high chance that you have heard all about it, and how it has been a complete game changer. They can cook things in a fraction of the time it used to take! And it’s not a fryer at all - it’s really a mini-oven! If you got an air fryer yourself, then there’s a chance that you’ve used it for everything, and that, even now, you are thinking about what you could use it to cook next.

I don’t have an air fryer myself, but am old enough to remember when my family first got a microwave. We lived off jacket potatoes for at least a week, and tried microwaving many things that should not be microwaved (there’s a reason that roasts are called roasts). Eventually, we found, just as my friends with air fryers seem to be finding in the weeks after Christmas, that, while the microwave is a useful tool to have in the kitchen, it’s not the only answer, and certainly not the best answer for everything.

It seems to me likely that LLMs are like air fryers, and that we are living in an extended version of the post-purchase euphoric period when we try to use them to do everything - including things that it turns out that they are not very good at. It’s not surprising that this euphoric period has lasted longer than the few weeks after Christmas: producing language is a much more impressive capability than defrosting a ready meal or roasting a potato.

It’s also not surprising that we should attempt to apply the powers of LLMs so widely: language is the tool that we use to understand the world and each other, it is the way we organise our lives, our thoughts and our societies, and it is the nearest we come to knowing what others are thinking. A tool that seems to speak our language seems capable of anything.

However, we are also discovering limits. I think we are having, or are about to have, experiences similar to opening the microwave to find a congealed mess, or finding a tough and shriveled object in the air fryer. For example, it has frequently been suggested that, given that LLMs speak language, and that code is just language, LLMs could do all of our coding for us. It has even been suggested that we could build an entirely LLM based operating system, in which we give the model low level control over a machine’s hardware.

I think that, while there is value in LLM coding assistants, we should recognise when we need a different type of tool to do a different type of job. LLMs are good at generating text in contexts where a degree of variation and unpredictablity is acceptable, and possibly desirable, and when the consequences of variation and unpredictability are low. (I don’t care whether my LLM-powered assistant says, ‘Hello!’ or ‘Good morning!’ at the start of each day.) They are less useful when we require precision (I do care whether my generated code says ‘print(‘Hello!’)’ or ‘print “Good morning!”’.) In coding, particularly complex, novel and low level coding, we require precision: I do not think that we are ready to delegate this to LLMs yet.

This does not mean that LLMs will not have an important and profound place in our workflows. The air fryer and the microwave are often referred to as game changers for a reason: like other modern conveniences, they change the dynamics and economics of cooking, freeing up time to do other things. Humble time-savers can save relationships and make family life more fun. But they are also part of an entire kitchen of tools that enable us to apply other techniques. Importantly, that kitchen (if maintained well) is plumbed, powered and serviced to a well known and agreed set of standards.

It’s fine to be excited about a new air fryer, and it’s fine to be excited about LLMs. Excitement leads to experimentation and experimentation leads to knowledge. But we also have to find them the right space on the kitchen worktop, check that the plug is wired correctly (and that we can afford the electricity bill), and learn which tools are best for cooking pasta, and which tools are best for cooking chips.

(Views in this article are my own.)