The most intimidating thing about thinking about technology, whether as an academic, a policy maker or just as a human being trying to imagine the next five years without collapsing into a puddle of anxiety, is the fast pace at which it seems to move. Emergent technology makes us feel the future as something over which we have no control, that is difficult to understand, and that is catastrophically disruptive to our society, opening the door to the unfamiliar and the dangerous. This is not false, but is only ever going to lead to emotional paralysis. Particularly for philosophy, the slowest mode of thought, whose owl only flies at dusk, it feels impossible to find secure footing for analysis and critique.1 We stand instead on intellectual quicksand, the more we twist and turn the deeper we sink. Why even bother thinking about this new technology? Whatever we come up with will be obsolete by tomorrow?

And so, the natural reaction is to just ignore it and wait for someone else to do the thinking. The problem here, and one that LLMs ought to teach us well, is that the fewer minds we have thinking about a social problem the less capable we’ll be to address that problem. To use a contemporary term, this is a true doom loop. The more disruptive a piece of technology is, the more thought-worthy it is and the harder it is to think about, which means the demand for thinking is never met and it continues to stomp over everything.

Artificial Intelligence, in particular the Large Language Model (LLM), is a perfect storm for these problems.2 It has the backing of the most powerful corporations in the world, making it omnipresent and easily available. It challenges, at least on the surface, our understanding of human nature on the most fundamental level. It creates distrust between friends and colleagues, teachers and students, managers and staff as we all start to assume everyone else is sending us AI slop. It is also an extremely abstract technology, in the phenomenological sense. Its true nature is impossible to visualise and is nothing like we experience in our day to day lives. It is alien, unfamiliar, extremely technical and dull to try and understand it properly. All of this works together to leave us being paranoid, feeling overwhelmed and obsolete, and utterly powerless against this onslaught of technology that seems to be slipping out of our control.

If that’s how you feel, then I have good news for you. As disruptive as the LLM chatbot has been to our society, it seems now to at least be complete. It’s done, it’s finished. The experience you have with a free account using ChatGPT, Claude or DeepSeek is more or less as good as this particular technology is ever going to get. According to some, it’s even slightly less capable than it was a few months ago. The point is, we’re clearly now within the error margin of what this tech can do, so finally we can take some time to properly think about it.

GPT-5: A Damp Squib

This week, OpenAI finally released what it has called GPT-5. I’ll say more about the details in a moment, but if you caught any of the press releases or social media hype about this launch you might have expected something truly significant. That is not what happened. Even the some of the loudest AI hype influencers are profoundly underwhelmed.

This damp-squib of a release is likely to be a decisive turning point for future historians to track, and perhaps marks at least ’the end of the beginning’ of The Great AI Moral Panic of the 2020s. AI is not going away, and new AI technologies will be produced, but the LLM chatbot, that specific technology that has been most disruptive and sparked this current AI boom, has clearly stalled in its development. For about eighteen months I’ve been saying in private, with more confidence than I have a right to, that chatbots aren’t going to get much better. I no longer doubt this guess.

Gary Marcus, an AI expert of such acumen that AI boosters speak of him like he’s the devil, put it like this:

Ultimately, the idea that scaling [of LLMs] alone might get us to AGI is a hypothesis. No hypothesis has ever been given more benefit of the doubt, nor more funding. After half a trillion dollars in that direction, it is obviously time to move on. The disappointing performance of GPT-5 should make that enormously clear.3

What Marcus says here aligns with my own experience. LLMs as we understand them seem to have plateaued, and not even recently. The improvements to them over the last year or so have been incremental at best and more to do with external features and app design than the underlying technology.4 Naturally, no one can tell the future, and (for the first time in history) the vast majority of users of AI are not engineers or expects. However, as people like Marcus have been arguing for a long time, the underlying tech needs to be developed in creative and, more importantly for the key players, expensive ways to create something new. So far, OpenAI et. al. have just been burning venture capital in the hope that more power and more data will make LLMs better and closer to genuine human intelligence. One does not need to be an engineer to recognise the law of diminishing returns at work here. LLMs are a one trick pony, and it doesn’t matter how much petrol you feed that pony, he’s not going to turn into a Bugati Veyron. If anything, he’ll just become less good at being a pony.5

My suspicion is that it will still take the big players a while to go back to the drawing board. Marcus, for example, argues for a turn towards what’s known as ’neurosymbolic AI’, for example. This would be even more expensive than what the companies have been doing already, not because it needs even more data-centres and fossil fuels, but because it needs something much more expensive and much more unreliable: human beings. Neurosymbolic AI is the attempt to combine machine learning (neural) techniques that have proven so successful recently with good old-fashioned manually-coded logic, known as Symbolic AI. In a word, it means more manual curation of the content in the model and a return to some level of manual, human programming. That endeavour will make the skip fire of venture capital funding that is the current AI boom look cheap.

We can now stop to catch our breath

If there’s one thing we really need to do in response to the stalling of the LLM Hype Train, it’s stopping ourselves from saying and thinking “Whatever we do to address AI, it will all be obsolete by next year”. It’s not true, and it probably was never true. Anything that is a solvable problem today is likely to be a solvable problem for years to come. If, like me, you work in academia and are trying to find ways to address LLM usage in student assignments, your ideas and interventions, if they’re any good, should now last as long as they will need to. Moore’s Law does not apply to AI, and GPT-6, if it is ever released, will not be meaningfully better than GPT-5 at writing an academic essay, an email or anything else.6 There may be a technology in our future that offers another ChatGPT style disruption to our practices, but there is no reason to assume that it is imminent and no reason to assume it will necessarily be anything to do with AI. In truth, you’d probably be better off worrying about metaverse adjacent technologies that will crop up should you make the ill-advised decision to revert to traditional exams or in-class tests to stave off cheating.

Now is the time, if you’ve been avoiding it, to sit down with a chatbot and learn how to use it, learn what it’s good at and learn what its limitations are. Learn how it might help you, how it won’t, and how to educate others in that use. Do it now before they have to hike the price once the bubble bursts.

  1. ChatGPT thinks the reference about owls is a bit gatekeepy. It’s probably right. This is a reference to Hegel’s claim that philosophy is a science that makes sense of things after they’ve already happened, saying that the owl of Minerva raises its wings only at dusk. See also Oxford Reference article ↩︎

  2. Naturally, LLM chatbots are not the only angst-inducing disruptive AI technology. They’re closely followed by image and video generation technologies, in particular for the prospect they have for the spread of misinformation and enabling new forms of harrassment and harm. These also show some signs of stalling, but less so than their text-generating cousins. The chatbots are the current focus of my research and therewith the focus of this post. ↩︎

  3. GPT-5: Overdue, overhyped and underwhelming ↩︎

  4. Naturally I don’t expect you to take my word for that. If you want to look into this yourself, start with Marcus’ article, which refers to several benchmarks. I personally don’t particularly trust the benchmarks anyway, which is why I’ve not said much about them. I mainly speak from my own experience. For a contrasting view, making the case that GPT-5 is genuinely a meaningful step forward, try the Latent Space review. ↩︎

  5. Please do not feed petrol to ponies. ↩︎

  6. Obviously that’s stating my case a little bit too strongly. If you want a more academic and conservative assessment, then I’d put it that any major advance in the chatbot will be on the basis of new breakthroughs in engineering distinct technologies, rather than iterating the current methods. Those new technologies may incorporate LLMs in part, but will clearly need significant development in new directions using new techniques. We should not be holding our breath and delaying action over what to do with the tech that actually exists because of concerns over hypothetical techniques. ↩︎