Brains!
As Halloween approaches, the zombies come out. For most people, their one universal trait is their desire to eat (human) brains, but this feature is neither universal nor particularly old.
The modern zombie age began with the classic Night of the Living Dead, and director George A Romero is one of the most influential figures in the development of zombie film and literature. It’s interesting to note, however, that Romero’s zombies were the slow, shambling sort. And, while they ate human flesh, they had no particular interest in eating brains.
In fact, the 1985 film The Return of the Living Dead appears to be the first film in which zombies with an obsession for human brains appeared. This was a classic horror-comedy, where the zombies were created by a toxic gas created by the US military. Unlike most zombies, though, neither dismemberment nor destruction of the brain was sufficient, so one zombie was cremated... and the smoke released the chemicals into the atmosphere, causing a toxic rain that created more zombies.
To control the outbreak, the town was destroyed with a nuclear device, which released vastly more chemicals into the atmosphere, and the film ends with more toxic rain over a much wider area. It’s a wonderful film – scary and silly in equal measure.
Wherever the idea of zombies eating brains came from, it’s become an extremely common component in zombie films and literature, and a number of people have suggested explanations for why zombies find brains irresistible.
One ‘theory’ is that eating brains eases the pain the zombie experiences, with the suggestion that brains ‘provide zombies with the necessary endorphins to dull the pain of Rigor Mortis brought about by decomposition’. This is utterly ridiculous, but consistent with the silliness of Return of the Living Dead.
In the series Z Nation, eating brains appears to allow zombies to retain some degree of sentience, while The Murphy appears to be able to acquire a person’s knowledge from eating their brain.
Cannibalism has a long and complex history, and has been practised in various forms by various cultures throughout human history. I won’t go into it, except to note that a key hazard associated with cannibalism is tied to the risk of prion diseases such as Kuru, transmitted mainly through consumption of human brains. (Since prions are resistant to heat, cooking is not sufficient to destroy them, and there are currently no effective treatments.)
Which brings us to AI enshittification!
As Large Language Models have evolved, the general approach has been to feed them more. More parameters, more RAM, more GPUs, more data. And, while the increase in power of these models has been explosive over the past few years, I believe that we’re approaching the point of diminishing returns, where incremental improvements are only gained by an exponential increase in resources.
And, while there are certainly options available for future evolution of these models, most of the large AI companies appear to assume that the exponential growth will go on forever. This, obviously, is not sustainable, and we’re actually running out of human-generated content for future training.
One solution is to create synthetic data, which is then used to train other models. Or, put another way, one model eats the brain of another.
Which means that our future AI risks now include prion disease.
(NOTE: I wish I could take credit for this one, but I actually stole it from Michael Marshall, via Skeptics with a K: Episode #479)
Developing language models to generate vast amounts of content for the training of future LLMs is problematic for a number of reasons. The quality of the human-generated content already used has much to be desired, so we’d be using that as a model to generate more of the same, possibly resulting in more and more data of lower and lower quality.
I don’t think this is a viable approach, and I believe we’re nearing the end of what I will refer to as First Generation LLMs.
This leads to the question of what future LLM generations might be like.
One feature is more extensive use of model distillation, by which a large model can be used to train a smaller one. This technique is already being used, but is still limited by the size of the large model.
Another option may be to focus on limited models which are carefully curated and tuned to ensure quality, then develop approaches to make their use transparent to the user. So, rather than using a vast, generic model to try and answer your question about prion diseases, the generic model would forward the question to a model that is specifically-optimized for medical questions.
It may even be viable to create synthetic content, so long as enough thought is given to ensuring accuracy, quality, and minimizing bias.
I think we’ll get there – hopefully before we have an AI zombie apocalypse.
Happy Halloween!
Cheers!
Comments