Hacker News Reader

Comment by ren_engineer | Hacker News Reader

I'd imagine Google also has a far superior internal dataset for training considering their core competency literally involves indexing the entire internet.

All Google needs to do is launch a superior LLM API on Google Cloud and essentially hedge their bets that any replacement for Google Search will be built on their API anyway. Or just spinoff a shell company so Google doesn't get bad PR for any bad results returned by their ChatGPT equivalent. Microsoft has avoided any real flak for the stuff that OpenAI releases

The other question is how does something like ChatGPT monetize to pay for the massive costs to serve queries? Google engineers have already tested this and say that LLM queries are orders of magnitude more expensive to serve than current results and I don't seen any paid service replacing a free search engine

View on HN

> Microsoft has avoided any real flak for the stuff that OpenAI releases

Microsoft clearly learned the lesson. Remember Tay from Microsoft?

> All Google needs to do is launch a superior LLM API on Google Cloud and essentially hedge their bets that any replacement for Google Search will be built on their API anyway.

That would still mean a huge disruption to their core business.

> The other question is how does something like ChatGPT monetize to pay for the massive costs to serve queries? Google engineers have already tested this and say that LLM queries are orders of magnitude more expensive to serve than current results and I don't seen any paid service replacing a free search engine

GPT costs cents, I'd happily pay that. I'm looking at how best to use the newer models personally as it's so cheap.

I have not been able to make sense of their API for doing bulk indexing / analysis though.

I work at a place that is pretty reliant on knowledge bases (KBs) and they're the ideal target for this sort of thing, where you could prompt a situation into a chat mechanism and get a response giving some variation of the following:

- historical information on the issue (we've seen this before, logged here etc) - surface a solution quickly - If all else fails, auto redirects the experience to capturing useful information via branched prompts (e.g., ask the customer this, do this, record this, tell me what happened, and at the end, will record the response etc back into the KB automatically)

I can't figure out how to get the model to do this though.

Disclaimer: I'm not an AI / ML engineer, my background is Web Development, but this is too tasty of a situation to pass up, and I really want to build it on top of something like ChatGPT

Same here :) zero AI / ML experience but think the potential is huge.

I’m even thinking starting much smaller inside our helpdesk system: surface best matching previously-answered conversations based on the customer query (and perhaps one step further: a suggested response or at least some useful snippets).

Astute point. The greatest risk is probably reputational from weird/gamed responses. DeepMind might be a great brand to trial this with.

Reply on news.ycombinator.com