Why Reddit sells data to Google
45sContrasts the narrative of AI stealing data with actual paid partnerships, sparking curiosity.
▶ Play ClipAI Summary
The video explains why AI companies like Google and OpenAI are paying for data from Reddit, WordPress, and Financial Times. It distinguishes between using data for training AI models and using it as a grounded source for retrieval-augmented generation (RAG). The core concept is that RAG improves AI responses by retrieving real data from a database before generating an answer, making outputs more accurate and reliable.
Chapters
[00:00]
Data Deals with AI Companies
Reddit, Automattic (WordPress/Tumblr), and Financial Times are selling data to AI companies like Google and OpenAI, which seems counterintuitive because AI often undermines these sources.
[00:29]
Two Uses for Data
AI companies can use data for training new models or as a grounded source for retrieval-augmented generation (RAG).
[00:49]
How AI Systems Work
A prompt goes into the AI system, which generates a completion. However, the AI may produce plausible-sounding but incorrect answers because it only predicts tokens, not truth.
[01:23]
Grounded Sources
Adding a grounded source solves the accuracy problem: the AI sends the question to a database, retrieves matching information, and combines it with the prompt to produce a grounded response.
[02:10]
Retrieval-Augmented Generation (RAG)
RAG retrieves information from a database and augments the AI's response with it, leading to more accurate answers.
[02:19]
Semantic Cache
A semantic cache stores completions so that future similar prompts can bypass the AI entirely, improving efficiency.
The future of AI lies in grounding models in real data using RAG, rather than relying solely on training data. This approach makes AI responses more reliable and efficient.
Clickbait Check
95% Legit"Title accurately describes the video's focus on explaining RAG, with clear examples and a simple diagram."
Study Flashcards (5)
What are the two main uses of data by AI companies?
easy
Click to reveal answer
What are the two main uses of data by AI companies?
easy
Click to reveal answer
Training new models and using it as a grounded source for retrieval-augmented generation (RAG).
00:29
What is the problem with AI systems generating responses solely from training data?
medium
Click to reveal answer
What is the problem with AI systems generating responses solely from training data?
medium
Click to reveal answer
They may produce plausible-sounding but incorrect answers because they only predict tokens, not truth.
01:00
What does RAG stand for?
easy
Click to reveal answer
What does RAG stand for?
easy
Click to reveal answer
Retrieval-Augmented Generation.
02:10
How does a grounded source improve AI responses?
medium
Click to reveal answer
How does a grounded source improve AI responses?
medium
Click to reveal answer
The AI sends the question to a database, retrieves matching information, and combines it with the prompt to produce a grounded response.
01:23
What is the purpose of a semantic cache in RAG?
hard
Click to reveal answer
What is the purpose of a semantic cache in RAG?
hard
Click to reveal answer
To store completions so that future similar prompts can bypass the AI entirely, improving efficiency.
02:19
💡 Key Takeaways
💡
Two Uses for Data
Clarifies a common confusion about why AI companies pay for data: not just for training, but for grounding.
00:29
🔧
Grounded Sources
Introduces the key concept of grounding AI responses in real data to improve accuracy.
01:23
📊
RAG Definition
Provides a clear definition of retrieval-augmented generation, a core AI technique.
02:10
🔧
Semantic Cache
Explains an advanced optimization that reduces reliance on the AI model for repeated queries.
02:19Full Transcript
[00:00] you've probably heard the story of how
[00:01] Reddit is selling their data to Google
[00:04] for AI scraping how automatic is selling
[00:06] all the WordPress and Tumblr data and
[00:09] now how Financial Times is selling their
[00:11] data to open Ai and this doesn't sound
[00:14] right right because we've been told that
[00:16] these AI companies are taking all this
[00:18] data and basically invalidating the data
[00:21] sources they're coming from by sharing
[00:23] the data so you don't get to the source
[00:25] so what exactly is happening here this
[00:27] is more complicated than it sounds you
[00:29] see there are two different things these
[00:32] AI companies can do with the data one is
[00:34] use it for training you know building
[00:37] new models based on the data and they're
[00:38] definitely doing that in some respect
[00:40] but the other one is to use the data as
[00:43] a grounded source and that's really
[00:46] interesting so let me explain here's a
[00:49] very simplified drawing of what happens
[00:50] when you use a AI system like chat GPT
[00:54] you put in a prompt The Prompt goes to
[00:56] the AI system and the AI system creates
[00:58] a completion this is the response that
[01:00] comes out of the AI system the challenge
[01:02] with this model is if you ask the AI
[01:05] system something that exists within its
[01:08] training data it's likely to put
[01:10] together something that looks like an
[01:11] answer that's correct but it's just
[01:14] putting together tokens to make up
[01:15] something that looks like language it's
[01:17] not actually answering your question so
[01:19] there's no guarantee that the answer
[01:20] will be correct the way we solve this is
[01:23] by adding a grounded source so when you
[01:25] ask the AI system a question the AI
[01:27] system sends off the question to a
[01:30] database with information and then
[01:32] matching information gets sent back into
[01:34] the AI system it combines that with your
[01:37] request and you get a more grounded
[01:40] response that's actually grounded in
[01:42] truth and this is what they're going to
[01:43] be doing with these different media
[01:45] organizations instead of the AI system
[01:48] just straight up trying to answer the
[01:49] question it'll go to a grounded source
[01:52] to get some information first and use
[01:54] that to process an answer in the same
[01:56] way that when you're working with chat
[01:57] gbt if you ask it to write an article
[02:00] it'll write a okay article but if you
[02:03] write the start starting point of an
[02:04] article and ask it to help you make the
[02:06] language better it'll do a much better
[02:07] job this process is called retrieval
[02:10] augmented generation because it
[02:13] retrieves information and then it
[02:15] augments that information that gives it
[02:16] back to you and doing this we can take
[02:19] things one step further and introduce
[02:21] what's called a semantic cache so that
[02:24] when you put in a prompt we can go over
[02:27] to the grounded Source get some
[02:28] information put that back Ai and then
[02:30] when the completion comes out we'll set
[02:32] it into the cache so that next time we
[02:35] can bypass the AI entirely this is the
[02:37] direction AI is currently going and will
[02:39] be going in the foreseeable future
[02:42] instead of using the AI system to try to
[02:44] answer the question directly have the AI
[02:46] system ground itself in real data and
[02:49] use rag retrieval augmented generation
[02:51] to pull real data out and then augment
[02:53] it in its response hope that helps