The AI Boom's Own Analogy Says It Ends With a Few Models

The Trigger We Can Name

A short post on Readables puts the current moment next to the Cambrian explosion — the period beginning about 539 million years ago when animal life diversified rapidly in Earth's oceans — and notes one difference. No one knows exactly what set off the original. The author says we know precisely what set off this one: GPT.

That is the easy half of the argument. The harder half arrives at the end of the same post, where the author extrapolates from the analogy and predicts that power will concentrate among a handful of models. An explosion of variety that terminates in a short list of winners is not a contradiction. It is the part of the story most people building on top of this technology have not said out loud.

What Actually Changed Underneath

The post names four ingredients: more computing power, more training data, better architectures, and pretrained models that handle general-purpose tasks. Those are not four versions of the same thing, and the fourth is the one that matters most.

Architecture means the arrangement of the mathematics — how information moves through the system as it learns. Pretrained means the expensive part of the learning already happened, at someone else's cost, on someone else's hardware. General-purpose means the same model handles tasks it was never specifically trained on.

Stack those and the economics invert. Before, teaching software to do something new meant assembling a dataset for that specific thing and paying for a training run. A single narrow capability was a project. Now the capability is rented, and the work shifts to the layer above it: what you ask, what tools you connect, how you check the output. The post describes exactly this — people building more and more software on top of AI, which makes the AI itself more powerful — and lists where it is landing: coding, tutoring, design and research.

The biological half of the analogy is sharper than it first appears. The source attributes the original burst of variety partly to changes in how genes were controlled — not new raw material, but new control over existing material. That is a fair description of the last few years in software. The substrate did not multiply. The control over it did.

New, Or Repackaged?

What is genuinely new is the cost of attempting something. What is repackaged is most of the variety that cost buys. Thousands of visibly different products can sit on a small number of pretrained models, which is why the author's two claims fit together: diversification at the top, consolidation at the bottom.

Treat the prediction as what it is. The post offers no data for it, and does not pretend to — the language is "we could theorize" and "I think". The competition it describes is real and specific: researchers and companies pushing on usefulness, reliability, speed and cost. Whether that race ends in a handful of survivors is an inference from a 539-million-year-old precedent, not a finding. The author's other forecast is softer and probably safer: the world gets more complex and less easily comprehensible, and we eventually absorb it the way we absorbed computers and the internet.

Questions You Should Be Asking

• If your product's advantage lives in the layer above the model, what is left of it the day a base model absorbs that layer?
• Ask your vendor to name the pretrained model underneath their product. If they will not, what exactly are you buying?
• Of the four things being competed on — usefulness, reliability, speed, cost — which does your use case actually require, and are you paying a premium for the other three?
• If concentration is the endpoint, what leverage do you still hold? Your data, your distribution, your customers' switching costs — or nothing?
• What is your fallback if the model you depend on changes its behaviour, its terms or its price with no notice?

What To Watch Next

The signal is absorption. Watch whether the software built on top of these models keeps producing capabilities the next base model release does not simply swallow. If each upgrade quietly deletes a category of startup, the concentration thesis is playing out on schedule. Watch pricing too — concentration shows up as terms set unilaterally long before it shows up in anyone's market share chart.