back
1 comments
Sure, I was just pointing at the fact that the question "And where does ChatGPT get the data for those answers?" has a two-part answer:

1) training data (common crawl as one example web data source), and

2) live data optionally retrieved at runtime.

My comment was about 2), and that part runs via a search engine (Bing?), at least if you look at how ChatGPT does it.

Indeed, but what is the difference between crawl data and model data but decay rate? Models are trained on previous crawl data, but if an LLM provider engages a search engine to get "live data," that data isn't live but previously crawled as well (and perhaps not yet integrated into models as crawl data).

So, why would you use Google as a tool or search target when you can, in some combination, go direct to the website (or whatever the target data endpoint is) yourself as an LLM provider to retrieve the most recent data or rely on your own "hot cache" of that data that was crawled recently but said data is not stale enough warranting a live web crawl to retrieve and present to the user or AI agent? Is this capability to perform retrieval from a data source in real time not similar to an AI agent?

Broadly speaking, I'm just spitballing on the concept of "You must use a search engine for an LLM to return 'live-ish' results" as I think we're directionally headed to where that isn't the case.

> but if an LLM provider engages a search engine to get "live data," that data isn't live but previously crawled as well

I'm not sure frontier labs do it, but imo fetching live data requires fetching both the search engine (1) and the website/page (2). The search engine gives you search results + content snippets (potentially stale), the website/page gives you the actual content fresh from the source.

The tradeoff I see here is between liveness/staleness and cost (hitting an index if of course cheaper and faster than querying live websites again).

Cloudflare: Give any website a WebMCP interface - https://blog.cloudflare.com/webmcp/