Google may be only a year or two away from total disruption(twitter.com)
twitter.com
Google may be only a year or two away from total disruption
https://twitter.com/paultoo/status/1598434161332981760
8 comments
> What may be possible however is synthesizing the top N results into one coherent answer.
And there's an organization that excels at selecting the top N results: Google.
And there's an organization that excels at selecting the top N results: Google.
Lately they excel at selecting the top N SEO blog spammers, which IMO is unlikely to result in the synthesis of coherent, or at least truthful, answers.
Thinking out loud here, is it possible that Google is doing the same thing as always, but the results have gotten more spammy? Were there ever good results for “best headphones for running” or “how to make salsa”?
I checked Neeva and it’s not much better. Ditto on Bing.
Maybe Google has gotten worse because the internet has gotten worse.
I checked Neeva and it’s not much better. Ditto on Bing.
Maybe Google has gotten worse because the internet has gotten worse.
I don’t think “how do I write differential equations in LaTeX” is a high-ad-revenue search query? Whereas a query like “what’s the best car insurance” or “where can I buy air Jordans” is.
Does GPT do a good job on those types of queries?
Does GPT do a good job on those types of queries?
Isn't this tweet like the second of its kind. This was already discussed yesterday and people listed quite a few caveats.
Do you have a link?
How are you going to train this AI? With a corpus. And how are you going to choose the corpus? You can't use the whole internet; that would train it on lies, propaganda, advertising, ignorance, rants, and trolls (along with some truth). That won't get you good results.
You need something outside the AI deciding what material to use to train the AI. How do you decide what the best stuff is? You're probably going to need a large organization to do that, before you get results that lead to money. How are you going to pull that off?
You need something outside the AI deciding what material to use to train the AI. How do you decide what the best stuff is? You're probably going to need a large organization to do that, before you get results that lead to money. How are you going to pull that off?
And if a chatbot merely regurgitates the content I author without providng any credit, views, or ad revenue, how do you incentivize authors to keep writing to feed your chatbot?
It’s a thought that AI ( not ML) can provide better search result than the current algorithms
How a startup can get funding for crawling and storing the internet, running ML (not AI) training across the unnumbered TB and running the trained models to reply in the same latency. Is the hard part.
How such a startup would out compete google for the same idea is another hard part.
The idea of AI returning search results is the easy part
I really don’t mean to be snarky, I see this post as similar to saying Tyre companies will be disrupted by flying cars
I suppose Tesla disrupted traditional car makers so it possible
How a startup can get funding for crawling and storing the internet, running ML (not AI) training across the unnumbered TB and running the trained models to reply in the same latency. Is the hard part.
How such a startup would out compete google for the same idea is another hard part.
The idea of AI returning search results is the easy part
I really don’t mean to be snarky, I see this post as similar to saying Tyre companies will be disrupted by flying cars
I suppose Tesla disrupted traditional car makers so it possible
I have a lot of respect for Paul but this is very sensationalist.
Doubt.
Oh yah total disruption
1. Money. How cheaply can this be run per query, and how much value add does this have over the current product?
2. Modalities. GPT is only just ok at factual responses. It hallucinates a lot but let’s assume it didn’t. It’s damn near useless for location based and recency based queries. The day after an election, you’re gonna want to update it on who the president is. It’s not feasible to retrain. And don’t even think about answering “best restaurants near me” based on model parameters alone.
What may be possible however is synthesizing the top N results into one coherent answer.