⚠ Pretraining Warms the StartModerate threat

The cold-start moat rests on the assumption that a newcomer begins blind, without the behavioral history that makes an incumbent's answers feel prescient — and the danger is that foundation models undermine that assumption. A large model pretrained on the whole of the public web arrives already knowing an enormous amount about the world and how to answer questions, so a challenger can now start warm, offering competent results from day one without ever having watched billions of users click — as ChatGPT proved by reaching 100 million users within two months of launch¹.

This is dangerous because the cold start was the specific mechanism that made Google's data lead un-catchable. If pretraining supplies much of what usage-history once had to, then the circular trap — needing users to get good and needing to be good to get users — loosens, and a well-funded newcomer with a strong model can break in where the cold start previously kept everyone out.

Google's reply is that pretraining supplies general knowledge, not the live, behavioral, up-to-the-minute signal that ranking the changing web and reading fresh intent still require — and Google possesses both the models and that behavioral data, while a challenger has only the model. The warm start narrows the gap on general questions while leaving the incumbent ahead on everything that depends on knowing what people want now.

A moderate concern. Pretraining genuinely softens the cold-start wall that made Google's data lead so protective, and it lets capable newcomers begin far warmer than before — but the real-time behavioral advantage Google alone holds still matters for the freshest and most commercial queries, and Google brings model and data both. The wall is lower, not gone.