In

some other news, a new AI startup with their own ambitious vision for superintelligence has emerged from stealth mode this week. So this is called Reflection AI, and they have raised $130 million in funding and a $555 million valuation to build what they call autonomous coding agents. So they believe this represents a crucial step towards achieving superintelligence. Now, this company was founded by Misha Laskin and Ionis Antonoglu, who are two elite researchers from Google DeepMind. Antonoglu was a founding engineer at DeepMind who helped create AlphaGo, the breakthrough AI system that defeated world champion Lee Sedol at the board game Go in 2016, which is the moment many people consider a watershed in AI history. Now, unlike the coding assistance tools out there that just help you write code more efficiently, Reflection AI aims to create fully autonomous agents that can handle entire programming tasks from start to finish. Now, they believe that by combining reinforcement learning with large language models, they can tackle the essential complexities of software development. And early results suggest models outperform traditional code generation approaches by a wide margin. Now, as they develop this technology, they plan to expand the capabilities of their coding agents. The vision is that eventually, developers become directors of autonomous coding agents, and in the long term, this could extend to all knowledge work, not just coding. Laskin actually said, quote, our team pioneered reinforcement learning and large language models. And we decided that now is the time to bring both of these advancements together and build out a practical super intelligence that will do work on a computer. Now, Paul, we're seeing a lot of like agent startups out there, a lot of autonomous coding agents. Seems like with the background of these guys, this one might be a bit special. Yeah.

And I, you know, I think this is a lesson we've mentioned many times on this show, which is you follow the top researchers from the top labs. it's uh you know noam shazir you know when he launched character.ai i think we talked about that on the show um and then he eventually goes back to google so noam was at google multiple times goes that builds character.ai google acquires that the uh technology that they didn't buy the company they kind of think they could but they basically acquihire him and the team back for a few billion dollars. Like the top researchers are fundamental to understanding the research direction and to following along kind of what develops in this space. So yeah, like will this one work out? I don't know. Will they eventually get pulled back to DeepMind for a couple billion dollars in two years? Maybe, but it's always noteworthy. Now, the question here is why pursue autonomous coding agents? You hear us talk about that a lot. What was the thing called? Manus? Manus, yeah. Cursor. You hear about all these things. Here's why. AI researchers, there's tens of thousands of AI researchers. There's probably a few hundred, maybe up to a thousand who would be like top tier AI researchers that everyone would compete for and would pay million dollar plus bonuses to get them to come to their labs. I'm not an AI researcher, but my understanding of the space, one of the key values or traits of an AI researcher is their taste, their knowledge of which direction to pursue. So all of these labs are trying to get to AGI and beyond. The reason Ilya is so valued is because he has a history of very high taste, meaning he tends to know which research direction to go in. That leads to the greatest valued output. So if you're sitting in a major lab today, you all kind of have the same idea of how these models are improving, but you got to pick where your NVIDIA chips are going to get used and which things your top researchers are going to work on. So is it multimodal? Is it improving memory? Is it planning capabilities? Is it improving context window? Is it computer use? Is it reasoning? Is it agents? Is it reinforcement learning? Is it understanding world models? You have to make bets as to where to put your energy. So what does an autonomous coding agent do? It gives you almost infinite shots on goal. can now be running these things, pursuing all of these paths through low compute experiments that then when you hit on something, you go. And so that's what these labs do. They take all of these experiments. They fight over compute access within the companies every day. It happens at Google. It happens at Meta. It happens at OpenAI. They fight over access to compute to run their experiments to prove their hypotheses. And once you approve a hypothesis, you go. So like reasoning models were that. That's what Ilya did with Strawberries. He proved test time compute scaling law was likely going to hold. And that enabled OpenAI to double down on reasoning. So that's why this matters. It's why we, it's why we keep talking about these like coding agents. You may be like a VP of marketing or a CEO being like, what do I care about coding? You care about coding agents. Like they drive everything once they solve how to do this. And

it's probably likely at least an element of a lot of the talk around AGI and ASI. Even if this stuff feels far away, the moment you start cracking some of these autonomous coding agents is the moment we have kind of a fast takeoff, right? Yeah, because you

can run millions of experiments instead of dozens. Right. All

right, next up, we just wrapped up last week our AI for Writers Summit, which is a half-day virtual event that had 4,500 plus registrants from 90 plus countries. And this entire event was about how writers can begin to reimagine their work and careers in the age of AI. So Paul, to kick that event off, you gave a keynote on the state of AI for writers and creators, which was an overview of how the latest AI models are reinventing the future of creativity. And as part of the keynote, you did debut something called the human to machine scale for writers, which is a framework anyone can use to better understand their way forward with AI. Could you walk us through that scale and what inspired it?

Yeah, so we'll put the link to a LinkedIn post that I shared at the end of last week that actually has the 12-slide excerpt from the full presentation that plays out this whole human-to scale for writers. In essence, what I did is I iterated on a framework I had developed a few years back called the human-to scale that actually looked at levels of autonomy, like what is the human's role at a use case level when AI is a Could to their job or to the tasks within their job. And so as I was trying to like answer this question, when should we use AI to write, I realized I could probably actually adapt that human to machine scale to this. And so that's basically what we did is I hear from professionals all the time, specifically creative professionals who struggle with this question of like, when do I let the AI help? What do I let it actually do the writing for me? Because I'm a writer. It's like my art, my passion. It's the thing that gives me fulfillment. Like if people aren't familiar with me, like that's my background. I actually came out of journalism school. I've authored three books. We do the podcast. Like I consider myself a writer and storyteller by trade. And for me, writing is a very important part of my process. Like it's how I think, it's how I learn topics. It's how I develop an understanding. Um, it's, I can't just take an article, have AI spit out a summary for me and then talk to you all about the key points in it. It doesn't work for me. I don't develop a true comprehension of the topic. And so like the litmus test I gave, I think during the talk, because again, like I didn't script the talk, so I'm not actually sure exactly what I said, but I think I said something to the effect of anybody can use deep research or like chat GPT to write a summary of a topic. But to actually understand that topic in a deep way to the point where you could be a thought leader on it, imagine throwing all that aside and sitting there for 30 minutes and answering questions about the topic. That's my goal with everything we do with this show is like, I want to be so deeply ingrained in the things we analyze, the things I read and watch and listen to that I can throw away any script and just talk about the topic. Right. And so that's like kind of one of the fundamental things I shared with this idea is you have like level zero is all human. It's the human is the sole creator. Your voice matters tremendously. The audience expects authenticity. They expect you to just be sharing your knowledge. So that's all you. Level one is mostly human. That's where the author is still leading, the human author. But you're using AI for things like research or refining your work or brainstorming. Level two gets into half and half. It's like a co-writer situation where the author and the AI truly start to work together. There's an increasing focus on efficiency of rewriting, the voice and the human touch still matter. Level three gets into mostly machine. That's where it's largely AI driven. The AI is probably writing the first draft. The human maybe tweaks it, refines it, approves it. So efficiency starts to take on far greater meaning. And then level four is all machine, where the human's basically removed from the loop. It's an AI writer purely. It autonomously writes the stuff with little or no human oversight. And so in the, again, I would encourage people to go download the PDF from my LinkedIn post because it goes into like examples and characteristics at each level. And then it gives some tips at the end, like when does more human writing matter and when is it more okay to work with machines? But the big point I made is it's not a binary decision. Do I or do I not use AI? It sort of like exists on this spectrum. And that spectrum, the level zero to level four, in this instance, is very subjective and personal. The thing I didn't really address during the talk that's important is like some people aren't very good writers and like they want to express themselves, but they don't have the ability to. And so like level two in the co-writer situation may be the sweet spot for you because you're not a writer by trade. Whereas for me, I would say probably like 80 to 90% of mine is level zero podcast stuff, my keynotes, my LinkedIn posts. I have zero use for AI for that stuff. Like I want that to come from me. And the process is the purpose is what I said on LinkedIn, like going through the process is why I do it. But there's a lot more that's become level one where it's still mostly me, but I'm increasingly using AI on the research front, outlining, refining, brainstorming. And that's okay. As long as it's clear with the people reading it or hearing it. Um, and so, yeah, I mean, there's a thank you to everyone who's commented on that LinkedIn post. There's like, I don't know, maybe a hundred comments by now. Um, and I, it sounds like it was helpful framework for people. So, you know, definitely go check it out. It was honestly, it was one of those things I finished at 11 PM the night before the talk. So no one had seen it except my daughter. I was laying in bed with my daughter. Like, I was like, can I just show you this? Because I got to make sure this makes sense. And so she, yeah, she's the only one that had even seen the framework before I did

the talk the next morning. That's awesome. So Google is doubling down on its incorporation of AI into search. The company announced last week it'll show AI overviews for even more queries and add Gemini 2.0 to AI overviews to make those results more useful. It's also getting closer to debuting AI mode. AI mode is a new feature that will generate the answer to a search query based on everything in Google's search index. Basically, just like you'd expect from Perplexity or ChatGPT search. Currently, this is only available if you pay for Google One AI premium. They're like paid tier service. But it will be rolling out a bit to users in the future. Now, with all these updates, the official line here is kind of that more AI overviews, more AI in search. none of this will really cannibalize people going to websites via links, which is the behavior, of course, that powers today's SEO and search ad ecosystem. Google claims that people are still clicking in and going to websites through AI overviews and that AI overviews and AI mode will bring new people to Google for new things, according to The Verge. But there is some other data that seems to tell a different story. So Forbes actually reported that some new research from a content licensing platform called Toolbit, which was shared exclusively with Forbes, says the AI search engine said 96% less referral traffic to new sites and blogs compared to traditional search. So the report actually analyzed 160 websites that included some new sites, consumer blogs over the last three months of 2024 to kind of understand how this was all working. So Paul, like we keep hearing that seo isn't necessarily dead it's just gonna change like do you believe that i mean we're gonna need to make sure of course we show up in llms but beyond that it just seems like this whole traditional model is on its way out yeah

i mean all i can say is like from my personal experience i certainly go to fewer links like i mean if i go to google and i'm doing research i'm clicking on every link and i'm curating it and you know if i think about research for the show or research for you know writing a book or research for planning a trip like if i go into google and i type links and i get 10 links or however many it is i'm gonna click them if I go into Google and I get an AI overview that answers my question directly, even if the links are prominently shown, I generally look at the links to make sure they're pulling from legitimate sources that I would trust. And if they are, I'm kind of assuming it gave me the answer I needed. Or if I use deep research, the better it gets, like the less I need to go into the citations. I just look and make sure they're legitimate. So I'm not saying my personal use is representative of the market. Right. But those seem like really logical assumptions. Like my hypothesis would be, sure, you'd have less traffic coming from it. So regardless of what Google and others say, I just have to believe that how people consume information is dramatically going to change for sure. So yeah, what it does to SEO, all I will say is on our intro to AI class I teach every month, we're getting way more questions about how do I show up in large language models like ChatGPT than I used to get. So I think people are starting to catch on to the fact that maybe that's the new SEO is like, how do we show up in ChatGPT and AI overviews? Is it different or the same than past search and how the algorithms work? Next

up, Andreessen Horowitz has come out with their latest Top 100 Gen AI Consumer Apps Report. So this report, which comes out every six months, ranks the top 50 AI-first web products by unique monthly visits per similar web, and the top 50 AI-first mobile apps by monthly active users per Sensor Tower. Some highlights from this latest report, ChatGPT's explosive resurgence. We talked about how it has reached 400 million weekly active users as of last month. And the mobile story is equally impressive. ChatGPT is consistently growing its active user base by 5% to 15% every month over the past year. And approximately 175 million of those 400 million weekly active users now access it through the mobile app. Second is the meteoric rise of DeepSeq. So they launched their public chatbot on January 20th, 2025, and they accumulated enough traffic in just 10 days to rank as the second most popular AI product globally in January. The Chinese hedge fund-backed AI tool reached 1 million users in 14 days, which was slower than ChatGPT's five-day mark, but then surged to 10 million users in just 20 days, which according to Andreessen, outpaces ChatGPT's 40-day timeline. By February, they had claimed the number two spot on mobile as well, capturing 15% of ChatGPT's mobile user base, with engagement levels slightly higher than competitors like Perplexity and Clot. Now, in total on this list, there were 17 new companies that entered the rankings. AI video apps are on the rise. They are, quote, bringing the true usability with reliable outputs, according to A16Z. There are three new entries on the list of those video apps. at number 23. AI coding tools are also really taking off. These include agentic integrated development environments or IDEs and text-to app platforms for non-technical users. So tools here include things like cursor, which we've talked about at number 41, and Bolt at number 48. So Paul, this certainly seems to be a solid barometer of some recent trends we've seen in AI largely. Like, did anything jump out to you here?

Top web products. I don't see Meta.ai anywhere, and I don't see Gemini anywhere. That's probably not a good sign. Yeah. Top Gen.ai mobile apps. Gemini is coming in at 22. Yeah, I mean, tops are interesting but i also think the uh middle to back of the top 50 or non-existent yeah it's the omission is two of where the market is so yeah it's uh it's fascinating i do think the deep sea thing is just it's gotta just burn anthropic and meta in particular i would imagine google to a degree too with gemini um that they just sort of show up out of nowhere and skyrocket up there with none of the marketing that these other ones have had yeah

well like we talked about last week also likely a reason meta is spinning out its own meta ai app right getting left off all these lists and all this attention yep all right so in some other news, according to Y Combinator, managing partner Jared Friedman, a quarter of startups in YC's current W25 batch now have 95% of their code bases generated by AI. Now, what's really interesting here is these aren't non-technical founders building businesses by leveraging AI as a crutch. Friedman emphasized, quote, every one of these people is highly technical, completely capable of building their own products from scratch. A year ago, they would have built their product from scratch, but now 95% of it is built by an AI. So essentially, developers are starting to become directors of AI systems rather than hands-on coders, describing what they want built and letting AI handle the implementation details, which Y Combinator says has some big implications. So for one, it dramatically accelerates development cycles. It also lowers the barrier potentially for creating software, allowing people with good ideas but limited coding experience to bring their visions to life. However, there are some new challenges here. YC general partner Diana Hu noted during a discussion that even when relying heavily on AI, founders still need the skill to evaluate the quality of the generated code. And YC CEO Gary Tan emphasized the point further, raising a crucial question about the long-term sustainability here of this approach. He said, quote, let's say a startup with 95% AI-generated code goes out and a year or two out, they have 100 million users on that product. Does the code fall over or not? Paul, what can we learn here about the bigger picture? This isn't just about coding or Y Combinator. It just seems like the barriers to building are falling so fast thanks to AI.

Yeah, it's one of my hopes actually for what I think will be significant job displacement in the coming years is that I think we're going to go through like a renaissance of entrepreneurship, this entirely new age of entrepreneurship, where everyone can be an entrepreneur, where, you know, if you don't have a job or you're coming out of college or, um, you know, you're looking for a transition that you, you can build something because one, two years out, you know, you're going to be able to just use words to build apps. You can do it now in some early demos and stuff. But I think that it's a chance to offset the disruption is through growth of startups. Very

cool.