Hacker News Reader

The first of the two steps is a web search. The model didn’t answer from its trained weights.

The model just asked the crawler to get a recent copy of relevant pages and used that to give the answer. That is why AI companies experimenting with browsers or at least agentic extensions to browsers as that allows to fetch on behalf of the user directly from the device with a residential IP and ditch expensive to maintain crawling infrastructure.

> relevant pages

If you want to know what the relevant pages are, you need a search index.

Index is just a part of the model. The nice property is that the index for LLM does not need to be updated often so LLM plus index can be static. This even works for the latest news as LLM can fetch major news sites to learn the latest stories and then fetch individual articles for the final answer.

> Index is just a part of the model

I can assure you that it is not. If you download any model off of huggingface, it does not also include an index of the internet.

That screenshot shows the model making a tool call to an external search index.

Google search may survive as a backend service, sure.

Reply on news.ycombinator.com