Rob Wiblin

But out in broader society, you can find a whole separate narrative about AI that feels like it barely interacts with anything that I've been discussing.

It's a narrative that I think particularly circulates among media and academic types.

And

The story that it tells is that GPT-5 was a massive failure that showed AI progress has more or less stopped.

In any case, AI doesn't make people more productive.

It's basically useless.

And on top of that, it costs more to deliver than people are willing to pay for it or are paying for it now.

And perhaps the AI companies might be on the verge of bankruptcy any day now, so bad is their business model.

It's a perspective that I think is clearly completely misguided, and I'm going to give you five reasons why I think that.

Firstly, the EPOC Capabilities Index.

EPOC is this great think tank.

They collect lots of data about AI capabilities, inputs, all of that.

You should definitely check them out if you're interested in this topic.

The EPOC Capabilities Index tries to look at as many different AI models as possible on as many different capabilities benchmarks as possible, going back as far as we can measure, rather than cherry-picking individual results.

And it finds that if anything, AI progress has sped up in recent years.

They report that the performance of the best AI model available at any point in time, that that increased twice as fast after April, 2024, as it was increasing before.

I think we could maybe look at the longer-term graph and say, perhaps it's just a linear increase.

There weren't so many data points back in 2022 and 2023, so it's a little bit hard to say.

And if there really was a doubling in rates of progress, probably much of that was down to the inference scaling that we talked about earlier, something that we probably can't keep up at the same pace.

I think this really does basically rule out the idea that AI progress has slowed down, let alone stopped completely.