# Is AI Replacing Google Search? | Marc Andreessen

*Joe Lonsdale Clips* · https://www.youtube.com/watch?v=tZH1NftU0Hw

## Summary

Marc Andreessen discusses whether AI is replacing Google Search, noting that AI models like ChatGPT are trained on similar internet-scraped data to Google's index. He argues Google is already halfway to the new paradigm with its one box and LLM pilot program, and that the future is clearly shifting from lists of links to direct answers. He also speculates that AI may eventually move beyond ask-and-answer to proactively telling you what you need to know before you ask.

## Key points

- Marc Andreessen says AI systems like ChatGPT are trained on data similar to Google's, using things like the common crawl to scrape the internet.
- Google has already been moving away from the 10 blue links toward the one box, answering questions directly before showing links.
- Google has a pilot program threading full LLMs into search to give LLM responses.
- Andreessen believes the defining paradigm will shift from 'here's a bunch of links, good luck' to 'here's the answer'.
- He compares older paradigms to horseback riding: they may survive as hobbies or exceptions, and an LLM can still give you the 10 blue links if you want.
- Ask a question, get an answer is probably better than navigating links for most things and most people.
- LLMs are very good at writing prompts themselves, and GPT-4 already suggests prompts on screen.
- In the future, the AI may know you so well that it tells you what you need to know before you even ask, like an exceptionally well-trained assistant.

## Notes

### Training Data Overlap
- Andreessen notes AI models like ChatGPT are trained similarly to Google: step one is gathering all the data.
- The "common crawl" is described as the open source version of the Google index, scraping all the text on the internet.
- So both systems use essentially the same dataset; the difference is the chat UI versus the 10 blue links UI.

### Google's Existing Shift
- Google has long been pushing away from the 10 blue links toward what they call the "one box".
- Increasingly, Google answers the question directly or offers another UI before showing blue links, which get pushed further down the page.
- Andreessen says Google is already "halfway there" and has a pilot program giving full LLM responses in search.

### The New Paradigm
- The defining paradigm of the era shifts from "here's a bunch of links, good luck" to "here's the answer" — Andreessen says this is very clear.
- His horse analogy: we still ride horses, but as a hobby or exception. Older paradigms may get incorporated into the AI; an LLM can still give you the 10 blue links if you want, but most people won't.

### Beyond Asking Questions
- Ask a question, get an answer is probably better than navigating links for most things and most people.
- But that may not be the optimal interaction model: LLMs are actually really good at asking questions and writing prompts.
- The big version of GPT-4 already suggests prompts for you on screen.
- Andreessen jokes about an earlink showing him prompts he may want to ask.
- The endpoint may be an AI that knows you so well it tells you what you need to know before you ask — like an exceptionally well-trained intelligent assistant.

## Transcript

[00:00] Is the era of search over? Like what what what future look like one of these changes protect some people saying Google's in trouble from this. Is that true? Yeah, it's a it's a little funny cuz there's there's different layers of it. So so so you know how these AI's are trained like how chat GPT is trained is actually not that dissimilar to how Google's trained in a in the sense of like step one is like who brought all the data. Right? So like basically scrape the internet. There's literally this thing called the common crawl that's basically the the the open source version of the Google index and just basically just like go get all the text on the internet and then train the AI in the same way

[00:29] that Google would go get all the text on the internet and train the search engine. Um so it's a little a little bit of it's like same data set and then you know one is the chat UI um and then the other is the the 10 blue links UI. You know Google's been pushing for a long time to get away from the 10 blue links anyway, right? So they've been pushing to what they call the one box and so you know more and more if you do Google queries, it will give you the 10 blue links, but before it does that it will either just like answer the question for you or it will give you like some other UI to get to the thing that you want. Right? And then the blue links are pushed further down the page. So they're they're already like I would say

[01:03] Google's already like halfway there. Um and now of course they're they're starting to thread in full LLM's in the in they have a pilot program now where basically the search engine is starting to give LLM responses to things. So there's probably a lot of shades of gray in here. Um you know look having said that you know does the defining paradigm of the era go from oh here's a bunch of links good luck to here's the answer? Yes on that I think it's it's very very clear that that's what's going to happen. Um and then my little wife's crack at it on this is you know we still you know ride horses. Um you know except we do it as like a hobby. Um you know it's a sort of an exception, right? Um and so one of the things that an LLM can

[01:40] do for you is give you the 10 blue links. Right? And so it it just may be that these older paradigms just kind of get incorporated into the AI and the AI can also do that for you if you want it, but most people won't. And very interesting. But it's yeah it's been Do you guys think it will Do you think in the next 5 years of less people will search then for things than search away? I think it's just like ask a question also a couple couple things. So one is just like is ask a question get an answer better than try to navigate the internet and search on links to find the answer. Probably yes for most things most people. Look there's another question though which is is is the actual model going to be ask a question

[02:11] get an answer, right? Is is that the optimal way to actually interact with these things? And like and that's the default way to do it and that's the way people are doing it out of the gate and so maybe that sticks. But but maybe not, right? Because like one of these things one of the things these things can do they're actually really good at asking questions. Right? So so there's this right? So there's this idea of prompt engineering, right? Which is like how to like build out more prompts to prompt an an LLM. And and one of the funny things is actually the LLM's are very good at writing prompts. Um >> [laughter] >> so I'm actually I'm actually very scared I'm going to have an earlink in my head

[02:42] and it's going to be showing me prompts I may want to ask. Yeah, exactly. Like that's what's going to get to you. Well the big and then you know you probably know this the big the big version of GPT-4 actually does this, right? And not the head not in the head part, but in the on the screen it actually you know suggests prompts for you. Um and so or take it a step further, you know maybe the prompts don't even exist cuz maybe what happens is the AI just knows you so well that it just tells you what you need to know before you even ask. Right? And it basically is just more like oh it's just like your assistant. It's like a it's like an exceptionally well-trained intelligent assistant.