[0:00] you've probably heard the story of how
[0:01] Reddit is selling their data to Google
[0:04] for AI scraping how automatic is selling
[0:06] all the WordPress and Tumblr data and
[0:09] now how Financial Times is selling their
[0:11] data to open Ai and this doesn't sound
[0:14] right right because we've been told that
[0:16] these AI companies are taking all this
[0:18] data and basically invalidating the data
[0:21] sources they're coming from by sharing
[0:23] the data so you don't get to the source
[0:25] so what exactly is happening here this
[0:27] is more complicated than it sounds you
[0:29] see there are two different things these
[0:32] AI companies can do with the data one is
[0:34] use it for training you know building
[0:37] new models based on the data and they're
[0:38] definitely doing that in some respect
[0:40] but the other one is to use the data as
[0:43] a grounded source and that's really
[0:46] interesting so let me explain here's a
[0:49] very simplified drawing of what happens
[0:50] when you use a AI system like chat GPT
[0:54] you put in a prompt The Prompt goes to
[0:56] the AI system and the AI system creates
[0:58] a completion this is the response that
[1:00] comes out of the AI system the challenge
[1:02] with this model is if you ask the AI
[1:05] system something that exists within its
[1:08] training data it's likely to put
[1:10] together something that looks like an
[1:11] answer that's correct but it's just
[1:14] putting together tokens to make up
[1:15] something that looks like language it's
[1:17] not actually answering your question so
[1:19] there's no guarantee that the answer
[1:20] will be correct the way we solve this is
[1:23] by adding a grounded source so when you
[1:25] ask the AI system a question the AI
[1:27] system sends off the question to a
[1:30] database with information and then
[1:32] matching information gets sent back into
[1:34] the AI system it combines that with your
[1:37] request and you get a more grounded
[1:40] response that's actually grounded in
[1:42] truth and this is what they're going to
[1:43] be doing with these different media
[1:45] organizations instead of the AI system
[1:48] just straight up trying to answer the
[1:49] question it'll go to a grounded source
[1:52] to get some information first and use
[1:54] that to process an answer in the same
[1:56] way that when you're working with chat
[1:57] gbt if you ask it to write an article
[2:00] it'll write a okay article but if you
[2:03] write the start starting point of an
[2:04] article and ask it to help you make the
[2:06] language better it'll do a much better
[2:07] job this process is called retrieval
[2:10] augmented generation because it
[2:13] retrieves information and then it
[2:15] augments that information that gives it
[2:16] back to you and doing this we can take
[2:19] things one step further and introduce
[2:21] what's called a semantic cache so that
[2:24] when you put in a prompt we can go over
[2:27] to the grounded Source get some
[2:28] information put that back Ai and then
[2:30] when the completion comes out we'll set
[2:32] it into the cache so that next time we
[2:35] can bypass the AI entirely this is the
[2:37] direction AI is currently going and will
[2:39] be going in the foreseeable future
[2:42] instead of using the AI system to try to
[2:44] answer the question directly have the AI
[2:46] system ground itself in real data and
[2:49] use rag retrieval augmented generation
[2:51] to pull real data out and then augment
[2:53] it in its response hope that helps