AI means more than LLMs

One of the most frustrating things that I find when talking about and listening to people talk about AI at the moment is the reduction of what AI means to just generative AI. In this current wave of AI, generative AI, or LLMs are the all dominant force in the discourse.

It doesn’t matter if it is an executive or an engineer talking AI, in the current climate we almost always assume they mean generative AI and LLMs.

As somebody who was drawn to AI and ML research over 10 years ago, before the advent of generative AI, it frustrates me a lot that the scope of what AI is has been limited to such an extent.

Intelligence is not just language

It’s understandable that as humans we see the way in which LLM are able to generate text that approximates natural language and see that as the peak of artificial intelligence.

It’s no accident that we are like that, as we have learnt to interpret the ways in which people wield their words as a signifier of their intelligence, but for LLMs the construction of language is not an indicator of actual intelligence, instead we are often fooled by it.

So much of what is making these LLM models powerful at the moment is the complex engineering and tooling around these models. The LLM alone cannot do what Claude Code, OpenCode, or Gemini can do.

The narrowing of our future

There is so much fascinating AI and machine learning science out there that is more efficient at particular tasks that people want to use LLMs for. Intent or image classification are functions that LLMs are used for, but they are not really optimised for this. They give us an approximation of classification for an unknown space that can be useful, but they are nowhere near as efficient as other ML models at these tasks when we know the space we want to classify into.

It’s not only existing, more efficient, approaches that are being lost in the discourse, it is also the newer ML techniques that are not getting enough attention. In this sense we are actually limiting what AI can mean, and creating a glass ceiling that we must break through for further improvements in the field.

I really used to love listening to podcasts and news in the ML/AI space, as there was fascinating discourse on so many different approaches, including discussions about BERT and models/methods like that that formed the basis for LLMs. After 2023 they became so obsessed with LLMs, and generally far too hype driven with the influx of capital in the space, that I stopped listening to most of them. Honestly I miss that.

LLMs are data hungry

The current focus on LLMs also pushes us down the path of heavy data consumption. For large language models to be large they need a lot of data. We have seen an erosion of the data privacy norms that have been trying to be put in place world wide, in order to get more data for these companies.

So this means we have seen a number of violations of Intellectual Property and privacy from these large companies, as well as dark patterns defaulting to opt-in rather opt-out for training on your data.

The sheer scale of data required for LLMs has also made us take steps back on transparency about training data, which means we can’t have independent assessment of data for bias or harmful content. That means we are just reproducing the bias of the data (which these companies have admitted).

We have also seen the horrific stories from data workers about the conditions that they work under to further train these models.

All of these leaves a bad taste in my mouth, but also for the general public. The association here is that AI is based on stealing your data, your art and your work. That’s a lesson that will be difficult to unlearn when we want to talk about other AI approaches in the future.

Just one approach we are missing out on

I couldn’t possibly cover all the different kinds of ML/AI approaches that are being drowned out, nor do I have the expertise to do so. Instead I will just focus on one approach that I was really interested in when studying my Masters, Evolutionary algorithms.

Evolutionary algorithms are an approach that to attempt to evolve towards a better solution through iterations of mutation and selection that take inspiration from the process of natural evolution. This can be really effective when we have a good idea of measuring fitness for a problem space, but not a good idea of what the best solution itself might look like.

A lot of AI and ML has taken inspiration from the natural world, including Artificial Neural Networks (which power a lot of AI solutions today) and I have found evolution to be such an interesting process for a long time, so I was drawn to Evolutionary Algorithms as an approach to ML.

Evolutionary algorithms help us break down one of the barriers of LLMs as a tool for AI, and that is that the LLMs are somewhat limited in what they can offer as a solution based on what is in their training data, whereas Evolutionary Algorithms are designed to explore novel solutions. To increase the ceiling of what AI/ML can do, we need tools and models that can create novel solutions too.

The boy(s) who cried AGI

Currently you will hear a lot of leaders of AI companies, and those that benefit from them, that the current wave of AI is so incredibly capable at a wide range of tasks that we are nearing Artificial General Intelligence (AGI). The AGI claim is just the peak of the mountain of claims that these people make.

Setting the expectation that Generative AI and LLMs will herald the coming of a sci-fi like AGI can only end with disaster. Either the apocalyptic claims these people make about job losses and more come true, or the bubble bursts and everyone is left feeling that they have been lied to by the boys who cried AGI.

For the future of LLMs this over-hype is bad news too. LLMs have been a big breakthrough for unlocking capabilities for certain fields, but setting them up as the all conquering AI approach is just setting them up for failure.

This spiralling hype cycle will leave us with a town of people who see any future AI claims and breakthroughs, and just think, “oh here’s another boy crying AGI again, I won’t fall for that”. For a positive future for AI, and what we can build with it, this is a real concern for me.