Why GPT-5 was never going to impress you

GPT-5 was the most advanced AI when it was released, but most people were disappointed. Why? In this episode, I unpack the two key paradoxes that shape how we judge new technology: shifting goalposts and negative space.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Azeem Azhar?

  Today I want to talk about GPT-5. This model was met with mixed emotions. I called it evolutionary rather than revolutionary. Many people were a bit underwhelmed by it. Some were positively cross. Here's the thing, GPT-5 could never have impressed us. It's because it falls between two paradoxes of progress. Paradoxes that have played out time and time again through history. And they do give us a clue to how we're going to react to ever improving artificial intelligence.

• 0:38Azeem Azhar?

  To understand this first paradox, we have to go back to the early days of computing. Actually, back to 1950, before the term artificial intelligence had been coined. We'll go to Alan Turing. He was doing all that breakthrough work in cryptography and the theory of computation. And he came up with this test for machine intelligence that later on got known as the Turing Test. The test was reasonably simple, right? If the output of a computer system, a machine, was indistinguishable to other humans from the outputs from other humans, you've got a machine that is exhibiting some type of thinking. And the Turing Test became the thing people measured towards artificial intelligences.

• 1:22Azeem Azhar?

  So we have this test and it's pretty explicit. But here's the thing, back in 2014, Eugene Guzman, a computer program, won the Turing Test. It persuaded judges at Britain's Royal Institution that it was human years before ChatGPT. And so we end up in this world where we say, well, it used deception. It's a parlor trick. This isn't a good test. Today's LLMs easily pass the Turing Test. And we've already started to see some media outlets having to retract stories that they now realize weren't written by freelancers, but were actually written by people using AI systems end to end. So today, we don't use the Turing Test as a test for machine intelligence.

• 2:07Azeem Azhar?

  We have shifted the goalposts. We measure AI's performance against a series of increasingly complex benchmarks. This effect of shifting goalposts has been noticed since the 1970s. Rodney Brooks, who is a professor of computer science and robotics at MIT, puts it very, very pithily. He says, every time we figure out a piece of this artificial intelligence, it stops being magical. People say, hey, that's just computation. And that's what's happening, you know, we're moving these goalposts. I think that if you took the capabilities of GPT-5 and dropped them into that 2014 Turing Test challenge at the Royal Institution, people would have had their minds absolutely blown.

• 2:50Azeem Azhar?

  But now it's just seen as a small improvement from something like O3. Now, the second paradox is the negative space paradox. Now, that sounds all fancy, but it is more subtle. You've probably lived through it, experienced it yourself. I mean, consider flying. When people first got access to transatlantic flights in the 1930s, it was really remarkable. No more five-day steaming across the Atlantic. But within a few years, passengers were complaining about the time it took, the comforts onboard and the food that they were being served. And so the paradox of negative space is that progress makes the gaps stand out much more. And I think for many people who are using large language models today through these chatbots, there are these concrete contrasts.

• 3:35Azeem Azhar?

  Models have got faster, they've become more reliable, better at using tools. They are hallucinating less, you can more reliably get them to search the web and extract information for you. But they still lack a whole range of capabilities, whether it is long-term memory about you, whether it's actually actively learning from your experiences. And you also get a sense that maybe they don't generalize as well as a real intelligence would. And so you end up in this quite odd space. And let me give you that example. If you've got an AI system that is unreliable, so say 10% of the time it makes errors, you're unlikely to put it into any kind of automated workflow. You'll want to sit on top of it because.

• 4:20Azeem Azhar?

  One time in ten, it's going to make a muck up. Now, when that error rate drops to one percent, you'll feel much more confident about putting it into some kind of automated system, automated workflow, hundreds of times a minute, thousands of times an hour, tens of thousands of times a day. But that one percent hallucination rate will show up time and again. Or consider a series of individual steps chained one to another. Imagine you've got a process with twenty five steps. Well, a one percent hallucination rate means that each step succeeds ninety nine times out of a hundred. But across a chain of twenty five, it will mean one in five times that chain will fail. And so, wow, we've got the hallucination rate down as OpenAI has with GPT five.

• 5:05Azeem Azhar?

  Enough to say we don't need to attend to it all the time, but not so low that you can't let it run thousands of times without there being lots of problems or power a twenty five step workflow through it. And so you end up with this, this funny space, which is of course the technology is better than it was. Of course, it's giving us more value for the twenty dollars or two hundred dollars a month, but we really feel that gap. So why does this all matter? Shifting goalposts mean we redefine success as soon as it's achieved. Negative space means every improvement makes what is still missing even more obvious. Together, they guarantee that GPT five or GPT six or any new model from Anthropic or from Google or Deep Seat

• 5:50Azeem Azhar?

  will probably feel less revolutionary than perhaps it really is. GPT five didn't fail to impress us because it wasn't impressive. It failed to impress us because we were never going to be impressed. This probably means that we're not going to have a before and after artificial general intelligence moment. Rather, improvements will get delivered on a smoothish curve. To have a moment of awe, a sense of before and after, I think we will need a paradigm shift in the way we build and deliver AI models and in what they can actually. Do. We have some experience of that because that's exactly what ChatGPT did.

• 6:34Azeem Azhar?

  There is a before ChatGPT and after ChatGPT moment. But that moment came about by something that was rather surprising and surprising to OpenAI itself, which was ChatGPT and how effective it was and how it delivered a new paradigm. So until we actually get AI systems that are built on that new paradigm, that have such a distinct flavour to them, to the systems before, we'll continue to be unimpressed. And that would be my median expectation of the coming months and couple of years at least as the AI companies roll out better and better systems.