Chris Barker

One cliche in the pro-AI community is that you should invest as soon as possible, because: ‘this is the worst the models will ever be.’

As a talking point, that isn’t very persuasive- you could also make the claim that you should invest in a chair company, because chairs are also ‘the worst they’ll ever be’ in terms of materials and construction. Better jump on the chair train now, right?

I would also argue that AI is not ‘the worst it will ever be.’ The underlying models will continue to improve, although potentially at a more incremental pace until a new paradigm is discovered. However, the EXPERIENCE of using AI is probably the best it will ever be.

The situation is analogous to Facebook in 2004-2007: the big AI companies are still trying to build scale, capture the market, and avoid regulation and bad publicity. Their incentive is to burn money offering a good product at a good or reasonable price. When that honeymoon phase is over…that’s when the squeezing begins.

AI is a low-margin business

The interesting thing about the recent release of GPT-5 is the lack of any moat.

If GPT-5 had been released in 2023, it would have blown the field away, but Claude, Grok, and other models are now within spitting distance on the same benchmarks, and open source/China isn’t far behind. OpenAI’s products are good, but don’t stand out against the competition anymore. Every time they seem to get ahead, everyone else catches up within a few months.

When several companies are selling essentially the same product, margins will be low because everyone will need to compete on price. LLMs are likely to become a commodity.

When margins are low, AI companies will need to squeeze their customers for all they have.

How the experience will get worse

This is the speculative part.

The vast majority of ChatGPT users (98%) currently use the free version of the product. They are currently burning money for essentially no return to OpenAI. How do you monetise them?

The easiest way might be to blanket the page with adverts. To be more evil about it, you might direct your LLM to include links to adverts in search results. To really go overboard, you might direct your LLM to advertise to users directly.

You might also tweak things to make service as cheap as possible, reducing performance when you can get away with it, potentially even for the paid users. This seems to be part of the purpose of GPT-5.

You might say that making service worse would remove AI’s main selling point, but look at Google, nowadays.

The least evil option might be to charge a small fee to cover expenses and keep everything the same. At least that way, people would know what they were paying for.

Conclusion

AI is not immune to economics. And AI models are a different thing to AI products.

The fact that AI is getting better doesn’t mean that the experience of using it can’t get much, much worse.