SOCIAL MEDIA –> ALGO MEDIA –> AI MEDIA
Algorithmic feeds turned social networks into money machines; AI media could make them infinite.
Thesis: The evolution of online platforms shows a clear progression: social media connected people through their networks, algorithmic media maximized engagement by surfacing endless content from anywhere, and now AI media promises to collapse the distinction between consumption and creation itself. Algorithmic feeds already transformed user time, ad targeting, and virality into a far superior business model, and with LLMs entering the front-end, the next shift won’t just recommend what we see — it will generate what we see. The opportunity (and risk) lies in how these AI-native tools move from optimizing feeds to personalizing worlds, opening both massive new consumer markets and unsettling questions about authenticity, virality, and human connection.
Social media has become a strange place. I have spent enough time on it to know that I don’t really want to spend that much time on it. And social media isn’t really strictly social anymore – it’s algorithmic media. How often do you log into a social platform and see content from friends and family – people you actually know? It’s less and less every time. That’s because the algorithmic feed is a much better business model. You can make people more engaged, you don’t have to worry about a big enough social graph (people you know on the platform), and produces even better behavioral data. That cocktail makes a user of a social platform much more monetizable.
And the proof is in the pudding – social media revenues have grown significantly since they made the switch. Instagram’s revenue is on track to grow by 24% in 2025, largely driven by Reels ad revenue, one of the ultimate algo feed products.
And while all anyone wants to talk about is AI (and specifically generative AI), this shift didn’t happen as the result of the rise of LLMs either. It has been going on for some time, dating back to when we used to just call it machine learning. So despite the fact that it still feels like we are in the early innings of AI product releases, the reality is that AI has already seeped its way into social media. If the last decade was about algorithmic media optimizing engagement, the next decade is about AI media reshaping creation itself.”
BUT WHY DID ALGO MEDIA HAPPEN?
For one thing, Algorithmic Feeds provide a better pool of content for technology media to draw from. If you are old enough, you remember the days when you could reach the “bottom” of your social media scrolling. You literally ran out of things to look at and consume. This is a problem for the media companies because as soon as those products run out of content, they run the risk of being boring, or worse, having the consumer log out. It’s hard to sell ads if people aren’t on your platform. Social sites, and media in general, are encouraged to capture your time. That has been true since media has existed and the dawn of time.
As previously stated, algo feeds are just way better business model. More ads, better ads, more users, more optimized content, steal underpants = profit.
MORE ADS. The algorithmic feed solves this problem. Instead of needing your own social graph to be very active, you can just rely on the rest of the world being active for you instead. User-generated content allowed for it such that you would never run out of potential content to consume. Importantly, there is virtually no “bottom” of the social feed today. There are too many people posting, and it doesn’t matter if you don’t know them or not – their content will be delivered to you. This allows users to stay on platform for longer.
BETTER ADS. It also, likely, allows for an improved ad product for the ad-buyers within these platforms. If users are conditioned to expect things in their feeds from people they don’t know, it’s a lot easier to accept ads in their feed as well. And, more importantly, because the algorithm is more finely tuned for what a user’s likes and dislikes are, it’s a lot easier for an advertiser to target them based on their personality profile. Engagement-based algorithms optimize for whatever holds attention. That produces incredibly granular behavioral data (scroll speed, likes, replays, watch-throughs) that can better zero in on a user’s interest graph. This supercharges targeting and ad pricing: TikTok’s ad platform is now rivaling Meta’s in efficiency because of this feedback loop.
MORE USERS. There is also a data flywheel that is generated by algorithmic feeds. As more people scroll on platform, the product is getting feedback described above. This then helps reinforce the training for the model on what content needs to be matched with each user. This in turn causes the user to stay on platform for longer, and on and on the flywheel turns.
Not only that, but algo feeds also results in trend acceleration – if people are spending more time on the platform, that means they are way more likely to spot trends. Which allows trends to go more viral faster. Information just moves more quickly. And as such, users are more likely to share trends simply because they are seeing more of them. They can then use social media’s natural virality features to invite new members to the platform.
“I should send this to XYZ person because they will like it” has always been true, but now social platforms just get more shots on goal.
OPTIMIZED CONTENT. Algo feeds are also better for creators, in theory. It does the advertising of the creator’s product (their content) for them. It’s not about whether a user subscribes to their content consciously – if the content they create is good and finds a match among users, it will generate more traction on platform, which can spin the discovery flywheel described above. This allows for more virality, which makes these platforms more effective distribution tools for creators. Which, in turn, allows for more and better content (theoretically), which drives more time spent on platform by users, and on and on the wheel turns.
IS THIS REALLY AI?
And while our central thesis has a lot to do AI being the next frontier, the reality is that Algo Feeds are already somewhat infused with AI, depending your definition. They aren’t exactly native to the current concept of AI– meaning LLMs generating content. But the backends are all largely driven by machine learning tools. Here is a Good explainer on how social media algorithms work from Cal Newport. You don’t just have a singular algorithm at work, but a complex set of algorithms that are at play. And while they may have started as somewhat deterministic, they have become extremely complex and self-learning / reinforcing.
So while GenAI / LLM-AI will generate new opportunities for new versions of social media and just media in general, the reality is that AI has already seeped into the media landscape. From Nvidia Founder & CEO Jensen Huang: “You can’t do TikTok without AI. You can’t do YouTube Short without AI … the amazing things that Meta is doing for customized content, personalized content … you can’t do that without AI.”
What Huang is talking about there is that these tools have all adopted various LLMs to make their algorithms even stronger. It’s not like the current format of social media is as fully realized as possible from an AI perspective.
He even went further and said: “until all recommender engines are AI-based, until all content generation is AI based … content generation consumer-oriented content generation is very largely recommender systems … and all of that’s going to be AI generated.” In other words, AI is only powering the backend of most social media platforms, but there is a good chance it will power the frontend / consumer experience of these platforms as well. Meaning: content generation.
SO WHERE ARE WE GOING NEXT?
Of course, we have already seen pretty significant consumer adoption of general purpose AI tools already. While many of them serve enterprise / business use cases, much of AI is used today in the name of content generation and media.
Currently, ChatGPT has reportedly somewhere between 400 hundred million to 1 billion weekly active users. The average ChatGPT user uses the tool for 16 minutes per day. The total system processes 2.5 billion prompts every day. To put that in perspective, Google processes something close to 14 billion searches per day. And it’s not just kids and techies who use the tool: nearly 45 % of Baby Boomers (ages ~61–79) say they have used AI in the past 6 months; ~11 % say they use it daily.
It has reached broad market adoption.
How it’s being used is interesting as well:
So it’s not just search replacement, but a whole host of interesting use cases that are novel to how the digital world worked 2-3 years ago. Consumer behavior has definitely changed. And all of this data is just from one player in the space. There are other challengers as well nipping at OpenAI’s heels.
So if algorithmic feeds allowed for near-bottomless content, improved ability to identify what people like and don’t like (recommenders), and an ability to drive more virality (good for creators and trends), what will LLM-based technology do? Again, in some cases, it is already impacting these various vectors and there is an argument to be made that LLMs will only enhance all three.
Near-bottomless content to ACTUALLY bottomless content. If you want something new, AI can just generate it for you. You don’t need to wait for a creator to actually produce content.
Better recommendations – this has basically already happened. LLMs are better able to digest complex data sets and spit out results. That’s all (easier said than done of course) they need to do on top of existing AI recommendation engines. It’s also entirely possible that they will create better recommendations beyond just what would be normally parsed through existing data. They can be more predictive in this sense and can test new ideas on their own or with little oversight.
More virality – This is the vector I am the most curious about. Theoretically LLMs should be bad for creators – they are the middleman that the bots are replacing, right? Or do creators shift from people who are good with cameras to anybody who has an idea? It’s an interesting idea that these tools will enable more creators, creating even more niche content. Virality in most social media products has become more about a user sharing content with a non-user and getting them to join, and less about convincing someone to join a platform so they can interact on platform together.
In order to better understand these three shifts, I think it’s worthwhile to talk through a couple of recent AI case studies.
AI Media Tools
Released last week, this is one of the new social features through Meta AI. Its premise is pretty simple: using the Meta AI tool, you can create content and then post it to a scroll-feed for other Meta AI users. The user experience is spiritually the same as Instagram or Reels. The difference is that the platform is focused on actually creating for you the content you post as well.
The launch received some mixed reviews, but this might be the closest we have come to actually bottomless content. It’s hard to tell, but I don’t think Vibes is creating any content “on its own” today, meaning with no prompting from users. It does appear, however, that concept could be on the horizon.
The virality of this product, and most of these products, seems to be focused on sharing content with others and driving them on platform to see more similar content. That is how things work in the algo feed world today for the most part, and doesn’t seem that novel.
This seems like a sustaining innovation: taking an existing business model and using new technologies to enhance it. It even looks like, from the announcement, that they are going to get this product into the Reels / Instagram feed at some point. So it’s not even cannibalizing the existing product, rather enhancing it. If it works – and I am deeply skeptical for now – it could generate a lot of revenue for Meta.
Along with Meta Vibes is Pulse by ChatGPT, which was also released this week. It is a Pro feature for ChatGPT users that automatically generates content for them based on their profile + recent searches and interests. It acts, essentially, as a daily hyper-personalized newsletter based on what you are looking for and your profile. And it doesn’t just find likes, but it provides commentary, addresses questions, and allows for deeper interactions with the content.
Much of ChatGPT is pretty dependent upon a user taking agency – sending out a query or writing a prompt. Pulse is one of the first products of theirs that is inbound, which is a key feature of most media tools these days. Tools like TikTok and even Twitter are mostly consumer by people who take no agency in the product – the lurkers. ChatGPT, on the other hand, does require human agency to get started. Through ChatGPT, Pulse is collecting pretty interesting details and context on a person based on their interactions with ChatGPT’s tools already. Depending on a person’s usage of the tool, that means their recommended content should get better with time. As a person uses ChatGPT more and more, Pulse will be able to provide better and more personalized content.
I am not sure about the virality of this product. One of the tricky things about LLMs is that they are so personalized – which is exciting from a bottomless content perspective – but it’s less compelling from a virality perspective. I do believe there is a limit for personalized content and its shareability. Something eventually becomes so personalized that you can’t share it with anybody else. But there is still a viral path here, it’s just harder to see. It could be collecting data from friends in some unique and compelling way and floating that into the product output, or something like that. Who knows what it will actually be, however?
Sora 2 was also recently launched, it is OpenAI’s most recent video & audio generation tool. If you haven’t seen the drop video, it’s incredible and sort of breathtaking. In some ways, this is just about content creation, and has a lot of similarities to what was shared above about Meta Vibes.
But, they do have the Cameo feature, which allows for you to insert yourself into a video. But it also allows for you to insert your friends into the videos as well. This has some implications around the shareability of this content and the group dynamics of creation. This could allow for some significant virality potential.
There is even potential for broader social use cases, like monetization for creators of their cameos. But more to come there.
AI Companion Apps
These companion apps are the most likely to be a disruptive innovation. It is media that is the least like the existing platforms – you are almost entirely responsible for your own content and output, but with the help of LLMs, of course. Users can build their own worlds and people and personalities. A couple of examples of these tools out in the real world:
Friend: Building a persistent AI “companion” in the form of a wearable necklace that you can talk to and carry with you. It aims to merge the intimacy of AI companionship with always-on, ambient hardware. Backers include founders from Solana, Perplexity, and zFellows.
Character.ai: Lets users create and interact with AI characters that have distinct personalities, voices, and backstories. It’s widely used for roleplay, entertainment, and even pseudo-therapeutic or tutoring interactions, with deep session times and viral creator communities.
Paradot: A mobile app focused on highly emotional AI companions, often framed as “virtual friends” or “partners.” It emphasizes visual avatars, memory, and emotional conversation—targeting users who want a stronger sense of personality and intimacy than a standard chatbot.
I am curious how viral and sticky these can be. We have already heard the horror stories, and as a purely social feature, they creep me the hell out.
However, at a certain level, I do see some of the appeal. I remember playing the first Halo on my Xbox in 2003 and thinking how cool and extremely futuristic Cortana was. Aside from the obvious appeals that Cortana had to an 11 year old boy, I also just thought the idea of having a tool on your person at all times that can help provide context on your surroundings and situation at the same level a computer could, but delivered in a UI that is more palatable / easier to understand than just a computer screen + data. That is, in many ways, what a lot of companion apps are trying to do.
The important point of companions is that their interface should be easier to digest and they should have more personalized context on your current situation. The positioning of Friend, in this sense, is the closest we have to building out a Cortana we carry around everywhere. If the companion doesn’t really have appropriate context, then it’s not a whole lot better than AIM’s SmarterChild. Texting with a chatbot is not a very interesting tool, it’s not very social, and it’s been in market for a long time. If there was going to be a consumer media adoption point for that tool, it would have already happened. (To be fair, Character.ai, has a TON of users, and that’s mostly what they do. So I could be wrong!)
Meta Ray Bans
Along the same lines as these social companions, one thing that is going to likely shift beyond just normal media is the potential for hyper personalization. That is the promise of LLMs in many ways. And there are tools that can deeply understand your context based on the data that you feed them. The one that sticks out to me is Meta Ray Bans. They have already been a relative hit (sold nearly 2 million units from 2023-2024) on the back of some relatively basic features. Of course, the compelling thing about the MRBs is not necessarily features, but they are the first AR tool to really make those features palatable to be worn in public. People will always relatively care about being cool, and we need our tech products to be cool, not just have cool features.
They have a built-in AI tool that could serve as a companion through Meta AI, they have a camera, microphone, and speaker, so they get several layers of data that most products out there don’t have about a person’s context & surroundings. And it is built by one of the most successful consumer social companies of all time. That is the most compelling part of MRBs: the enhanced context.
But the thing about AI companions in general is that there is no reason they need to be only 1:1 – for just one person to interact with just one companion. In theory, these companions could interlink several people and serve features to all of them at once or parsed out at different times. Could you have an AI companion that syncs what’s going on in your world to what’s going on in your wife’s? What about a companion that you and your buddies share to collectively search for tickets to concerts of a band you all like? Or a temporary companion that helps a school team work on a group project together?
The virality potential is high here, especially if we find a way to use these tools to connect each other, instead of just creating new, individualized worlds.
Summary
And that’s the thing about AI Media in general: we are still pretty early in the experience AND we already have pretty massive consumer adoption. Close to a billion people are using LLM tools right now. The ability to build social and media content on top of that user base is out in the world today. It’s not going to be easy, of course, but the right founders will figure it out.