DO AGENTS HAVE NETWORK EFFECTS? We are seeing a new wave of general purpose consumer-friendly agents, and it’s causing many thousands of AI startups to re-evaluate their place in the market. Will these new horizontal players take over completely? Will it be winner take all? Can products that focus vertical use cases survive? The new wave of products - Muse, Town, Instinct, Grokbot, etc - come from taking the power and magic of Openclaw/Hermes but wrapping it in polished, secure, consumer-friendly experience. This is only the beginning, and soon every major tech company will throw their hat in the ring and thousands of startups that have been working on vertical agents in X (travel, shopping, family, work, SMB, etc etc) have big existential questions to answer. The simple case against winner-take-all dynamics: These agents are tools, not networks. They inherently provide a solo experience, and although packaged as chat experiences, they don’t have network effects the way that Whatsapp or phone networks do. They’re more like email clients - I can use Superhuman and someone else uses Outlook and someone else is on Gmail, but in the end we just send email to each other. In that metaphoi, if a Muse agent encounters a Instinct agent, they will just figure out how to talk to each other by building APIs, or using email/messaging/etc., so that there’s no advantage to everyone running the same thing. Further, you can swap out the underlying LLM - as we saw in the Claude vs Clawdbot saga - things will still work. The memory and underlying context are just text files, and AI are great at porting things from one system to another. So where’s the moat? Just because they are tools today doesn’t mean they can’t evolve into networks. As I’ve written about in the past, instead of discussing “network effects” as a vague/amorphous concept, let’s instead overlay it against core KPIs each product has to build against: - Acquisition network effects: More users** = lower CAC, higher virality, more signups - Engagement effects: More users = higher retention, deeper usage, more frequency - Monetization effects: More users = higher ARPU, higher conversion, more share of wallet (**PS. and by “more users” of course I actually mean, of course, higher network density with more interconnection within/across networks. Not scale effects) So looking at these sub-categories of network effects, you might ask: What are the features that you’d build to actually create network effects in agents? How does this category become winner-take-all? Acquisition: In the pre-AI world, network effects in customer acquisition were driven by users taking actions that then bring in even more users - whether that’s sending an invite, sharing a piece of content, or getting mentioned in a comment. Those users would then join the network, repeat the same actions, creating more viral loops that would throw off more users. In the post-AI world, both users and agents can initiate these loops. An agent might suggest, “hey, do you want me to help organize a New Years celebration with family” and then might pull together a group chat with your siblings, in-laws, and parents. It might build a microsite for the event and ask people to sign up and RSVP, creating an opportunity to generate a signup. Or for work, you might finish a slide deck and your agent asks if you want to pull together a meeting to present it to your team. When it does that, it might create a prep doc, take notes, and host it all on a microsite that engaged others. These types of agentic viral loops need a few things to work: They need to be able to plug into your contacts, and to have enough context to know when to loop which people into which projects/events/activities. They need access to communication channels like email/messaging/otherwise, specifically where they can reach potential users, as this serves as the viral substrate for propagation. The magic will be how to get all of these things without being creepy, and to create enough trust to reach out to friends/colleagues/etc on your behalf, without being too push. No one wants an agent that says, “hey Andrew would like to get lunch next week” followed by “and sign up for this agent to agree!” - it has to be better and more subtle than that. Many viral loops in the pre-AI era, particularly in the workplace, were built on a content creation loop - you use Google Slides to make a thing, you share a link, and then people would view it. Eventually they might make their own. Many products like Figma, spreadsheets, Notion, Canvas, are all built on this loop, and so are YouTube, Instagram, etc. In the agentic world, you might start by talking to your agent about interior decoration ideas, and it might ultimately build you a Pinterest-like board and make it easily shareable with others. It might build you a spreadsheet mini-app with a budget, a schedule, etc., and again, make it easily shareable. It might volunteer to do this inside a group chat with your friends, and if they want to view it, they should sign up and get their own agent too. If the content creation loop works for agentic virality, I think it means there’s a big incentive for every agent to volunteer to make visually-appealing, shareable content, do this often. AI is so good at codegen that creating a one-off disposable artifact will probably be more useful than not. And there’s a big incentive to drive virality by building/owning the content yourself as the agent. In the interior decoration example, it would be more viral to create a standalone version hosted by the agent, rather than building an actual Pinterest board, or to build a self-contained mini-spreadsheet rather than creating a GSheet. A properly tuned agentic viral loop will do the former because it helps spread the agent, rather than helping spread Pinterest/GSheets. This might all sound like horrific spamminess to you, but I think it might be done tastefully. If an agent is high-retention and high-usage, you’ll have many shots on goal to gradually invite your friends onto the same platform as you. The first shot might come from an invite loop that says your colleague/friend wants to set up a place to share photos/coordinate calendars/etc., and even if you say no, over time, your friend might share very useful microsites for projects or events or research that’s relevant, and you might decide that you want to sign up to leave a comment, but then eventually you might try the other functionality as well. Engagement: It’s easy to imagine that in a world where everyone by default has super intelligent agents that product engagement in agents would go up. The simple argument is that the ubiquity of agents would lead to more successful task completion, making agents more useful and applicable to more problems. Routine tasks like scheduling something or summarizing/sharing notes from a meeting, would happen instantly if it ran through agents and you removed humans completely. As agents become more successful over time, not only with your own tasks but in coordinating larger and more complex multi-user decisions, you’d end up using them more. However, this does not answer the question of network effects. The question there is, does your agent get better when other people are all using the SAME agent? As I said earlier, the argument for NO is if each type of agent runs on a frontier AI model, can talk to other agents (all different types) in real-time via email/messaging/APIs/connectors/etc. Then there’s not much leverage. If you’re just organizing a meeting or a dinner party or meeting the agents just ask each other in real-time - and even if they are heterogeneous - they can use email and calendars and figure it out. But some tasks do benefit from everyone running the same agent, because you get the benefits of centralization: Trust, speed, identity, discovery, etc. Some examples: - “Who are 10 people in San Francisco that I should meet?” - agents can’t just ask each other to find an optimal solution - “Find me a VP of Engineering who isn’t publicly looking for a new job, but would leave for a fast-growing startup that pays X in stock” - there’s a mix of private and public info here, and an agent of a passive jobseeker might lie if the asking agent is not trustworthy - “Who’s a trustworthy seller of X that I should buy from?” - reputation might be gamed, and it might be better to have a global source of truth - “Help me pick out the best dinner in the next 5 minutes that my friends and all their +1s would like, incorporating allergies/schedules/locations” - real-time solution incorporating a bunch of shared context - … and even more so, if you imagine that there may be agents that lie, exaggerate, or otherwise to benefit their humans For the above examples, you could imagine Muse/Town/Instinct/whatever building a shared network+context layer that understands the network, its relation to public/private data, acts as a system of record for identity/trust, does multi-sided marketplace matching, etc. Then it would allow only its agents access to this platform, and agents outside of the network wouldn’t have access. If you’re part of this network, and are contributing data, your agent gets more useful over time. The network gets more powerful over time. The counterpoint to this counterpoint is that perhaps new startups will fill the void as a series of agent-facing point solutions, making it possible for people running agents of all times to interface with each other. Perhaps Muse creates a proprietary ecommerce marketplace that only Muse agents can use, but perhaps there will just be a next-gen eBay for any agent to sell to any other agent, and this service maintains central reputation/identity at that layer. In other words, perhaps the centralization will still happen, but just at a layer above agents. Most likely both will happen - every large tech company adds chat, commerce, search, etc, but also there are pure play versions of each of these. But this feels like more fragmentation than a true winner-take-all. Some parting thoughts: My argument in this writeup is simple: Agents don’t inherently have network effects. If all we build are super-intelligent assistants that can use tools and talk to each other, then the category might look surprisingly fragmented. Your agent can talk to my agent, just as Gmail can send email to Outlook, and there’s no particular reason we need to use the same one. Better models, UX, integrations, memory, and distribution might create very large companies, but those are not network effects. But the next gen of agents have a big opportunity design network effects in their interactions. On acquisition, they can turn the things they create—documents, events, websites, research, group chats, plans—into viral objects that naturally pull other people in. And on engagement, they can create shared networks of identity, reputation, relationships, private context, and intent that make the agent increasingly useful as more of the people around you join. There’s many ways for these network effects to instantiate. If we’re lucky, things will look more like the open web - mostly decentralized, with bits of centralization to make certain tasks smoother. Identity could become an open protocol. Reputation could live in marketplaces. Commerce could happen through an agent-native eBay. Professional discovery could happen through an agent-native LinkedIn. Agents could remain interchangeable clients sitting on top of a constellation of networks and marketplaces, much as browsers and email clients do today. In that world, enormous network effects emerge around agents without producing a winner-take-all market for the agents themselves. My guess is that we’ll see a fight over exactly this boundary. Every major horizontal agent will try to pull valuable network functionality inside its walls: host the artifact instead of creating a Google Doc, execute the transaction instead of sending you to Amazon, make the introduction instead of searching LinkedIn, create the event instead of using Partiful. Every time it does this successfully, it converts somebody else’s network effect into its own. Meanwhile, every incumbent network and a new generation of startups will have the opposite incentive: make their networks universally accessible to every agent so that no horizontal agent can recreate and capture them. This is why I don’t think the interesting question is simply, “Will agents have network effects?” Of course some will. The much more consequential question is: Where will the network effects live? If they live primarily in open protocols and specialized networks, we may end up with thousands of agents competing on intelligence, personality, UX, specialization, and price, all interoperating with the same underlying ecosystem. Vertical agents can thrive in this world. But if a few horizontal agents successfully accumulate identity, relationships, private context, reputation, latent intent, and distribution - and keep those assets proprietary - then switching agents starts to mean leaving your network behind. Next couple years will be very fun to see how this plays out.

Physicists may have found the origin of gravity. And that would finally explain why gravity is different from the other forces of nature.

For 9+ years we’ve been building simulators, data engines, data pipelines, scenario editors, RL environments, what's now world models... and dozens of other products. Now we’ve connected them all under one agentic layer. It is not just another harness but also the broader system to safely design, develop, and deploy Physical AI. I continue to believe a modern, AI-first development environment is one of the biggest bottlenecks to putting intelligence on a billion machines. Unsurprisingly, we’ve been using it internally for a while now and the impact has been profound in almost every measurable way. As you might know, our old office was on the corner of Pioneer and Dana – and calling it Pioneer would be too cliche. Introducing Dana :) applied.co/dana-blog

I love this idea from Tobi Lutke: “I take the view that I'm a corporate raider. That Shopify went bankrupt and I bought it on a fire sale and I'm marching in on day one, and that previous management was crazy and we need to turn this place around.”

Advanced Machine Intelligence (AMI) is building a new breed of AI systems that understand the world, have persistent memory, can reason and plan, and are controllable and safe. We’ve raised a $1.03B (~€890M) round from global investors who believe in our vision of universally intelligent systems centered on world models. This round is co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, along with other investors and angels across the world. We are a growing team of researchers and builders, operating in Paris, New York, Montreal and Singapore from day one. Read more: amilabs.xyz/ AMI - Real world. Real intelligence.

New paper out! We present a training method for multimodal generative models, called Self-Flow, which combines classic flow matching and representation learning. Why? Unlike most representation alignment methods, our new approach does not require external, pretrained models and thus scales gracefully to joint multimodal training on images, videos and audio. How? It combines per-timestep flow matching with dual-timestep representation learning, improving the models' internal representations. This approach outperforms prior methods and shows promising scaling behavior in multimodal pretraining. It also enables downstream applications such as action prediction for embodied AI. webpage+paper: bfl.ai/research/self-flow code: github.com/black-forest-labs… Credit to @hila_chefer, @pess_r, Dominik, @dustin_podell, Vikash, @Vinh_Suhi and Antonio. If you enjoy doing open research like this, come and join BFL! We are actively hiring🌲

Finally had a chance to listen through this pod with Sutton, which was interesting and amusing. As background, Sutton's "The Bitter Lesson" has become a bit of biblical text in frontier LLM circles. Researchers routinely talk about and ask whether this or that approach or idea is sufficiently "bitter lesson pilled" (meaning arranged so that it benefits from added computation for free) as a proxy for whether it's going to work or worth even pursuing. The underlying assumption being that LLMs are of course highly "bitter lesson pilled" indeed, just look at LLM scaling laws where if you put compute on the x-axis, number go up and to the right. So it's amusing to see that Sutton, the author of the post, is not so sure that LLMs are "bitter lesson pilled" at all. They are trained on giant datasets of fundamentally human data, which is both 1) human generated and 2) finite. What do you do when you run out? How do you prevent a human bias? So there you have it, bitter lesson pilled LLM researchers taken down by the author of the bitter lesson - rough! In some sense, Dwarkesh (who represents the LLM researchers viewpoint in the pod) and Sutton are slightly speaking past each other because Sutton has a very different architecture in mind and LLMs break a lot of its principles. He calls himself a "classicist" and evokes the original concept of Alan Turing of building a "child machine" - a system capable of learning through experience by dynamically interacting with the world. There's no giant pretraining stage of imitating internet webpages. There's also no supervised finetuning, which he points out is absent in the animal kingdom (it's a subtle point but Sutton is right in the strong sense: animals may of course observe demonstrations, but their actions are not directly forced/"teleoperated" by other animals). Another important note he makes is that even if you just treat pretraining as an initialization of a prior before you finetune with reinforcement learning, Sutton sees the approach as tainted with human bias and fundamentally off course, a bit like when AlphaZero (which has never seen human games of Go) beats AlphaGo (which initializes from them). In Sutton's world view, all there is is an interaction with a world via reinforcement learning, where the reward functions are partially environment specific, but also intrinsically motivated, e.g. "fun", "curiosity", and related to the quality of the prediction in your world model. And the agent is always learning at test time by default, it's not trained once and then deployed thereafter. Overall, Sutton is a lot more interested in what we have common with the animal kingdom instead of what differentiates us. "If we understood a squirrel, we'd be almost done". As for my take... First, I should say that I think Sutton was a great guest for the pod and I like that the AI field maintains entropy of thought and that not everyone is exploiting the next local iteration LLMs. AI has gone through too many discrete transitions of the dominant approach to lose that. And I also think that his criticism of LLMs as not bitter lesson pilled is not inadequate. Frontier LLMs are now highly complex artifacts with a lot of humanness involved at all the stages - the foundation (the pretraining data) is all human text, the finetuning data is human and curated, the reinforcement learning environment mixture is tuned by human engineers. We do not in fact have an actual, single, clean, actually bitter lesson pilled, "turn the crank" algorithm that you could unleash upon the world and see it learn automatically from experience alone. Does such an algorithm even exist? Finding it would of course be a huge AI breakthrough. Two "example proofs" are commonly offered to argue that such a thing is possible. The first example is the success of AlphaZero learning to play Go completely from scratch with no human supervision whatsoever. But the game of Go is clearly such a simple, closed, environment that it's difficult to see the analogous formulation in the messiness of reality. I love Go, but algorithmically and categorically, it is essentially a harder version of tic tac toe. The second example is that of animals, like squirrels. And here, personally, I am also quite hesitant whether it's appropriate because animals arise by a very different computational process and via different constraints than what we have practically available to us in the industry. Animal brains are nowhere near the blank slate they appear to be at birth. First, a lot of what is commonly attributed to "learning" is imo a lot more "maturation". And second, even that which clearly is "learning" and not maturation is a lot more "finetuning" on top of something clearly powerful and preexisting. Example. A baby zebra is born and within a few dozen minutes it can run around the savannah and follow its mother. This is a highly complex sensory-motor task and there is no way in my mind that this is achieved from scratch, tabula rasa. The brains of animals and the billions of parameters within have a powerful initialization encoded in the ATCGs of their DNA, trained via the "outer loop" optimization in the course of evolution. If the baby zebra spasmed its muscles around at random as a reinforcement learning policy would have you do at initialization, it wouldn't get very far at all. Similarly, our AIs now also have neural networks with billions of parameters. These parameters need their own rich, high information density supervision signal. We are not going to re-run evolution. But we do have mountains of internet documents. Yes it is basically supervised learning that is ~absent in the animal kingdom. But it is a way to practically gather enough soft constraints over billions of parameters, to try to get to a point where you're not starting from scratch. TLDR: Pretraining is our crappy evolution. It is one candidate solution to the cold start problem, to be followed later by finetuning on tasks that look more correct, e.g. within the reinforcement learning framework, as state of the art frontier LLM labs now do pervasively. I still think it is worth to be inspired by animals. I think there are multiple powerful ideas that LLM agents are algorithmically missing that can still be adapted from animal intelligence. And I still think the bitter lesson is correct, but I see it more as something platonic to pursue, not necessarily to reach, in our real world and practically speaking. And I say both of these with double digit percent uncertainty and cheer the work of those who disagree, especially those a lot more ambitious bitter lesson wise. So that brings us to where we are. Stated plainly, today's frontier LLM research is not about building animals. It is about summoning ghosts. You can think of ghosts as a fundamentally different kind of point in the space of possible intelligences. They are muddled by humanity. Thoroughly engineered by it. They are these imperfect replicas, a kind of statistical distillation of humanity's documents with some sprinkle on top. They are not platonically bitter lesson pilled, but they are perhaps "practically" bitter lesson pilled, at least compared to a lot of what came before. It seems possibly to me that over time, we can further finetune our ghosts more and more in the direction of animals; That it's not so much a fundamental incompatibility but a matter of initialization in the intelligence space. But it's also quite possible that they diverge even further and end up permanently different, un-animal-like, but still incredibly helpful and properly world-altering. It's possible that ghosts:animals :: planes:birds. Anyway, in summary, overall and actionably, I think this pod is solid "real talk" from Sutton to the frontier LLM researchers, who might be gear shifted a little too much in the exploit mode. Probably we are still not sufficiently bitter lesson pilled and there is a very good chance of more powerful ideas and paradigms, other than exhaustive benchbuilding and benchmaxxing. And animals might be a good source of inspiration. Intrinsic motivation, fun, curiosity, empowerment, multi-agent self-play, culture. Use your imagination.

FLUX is now also in Meta AI. What a day. 🥵

Nice job! Open research / open source accelerates progress.

Here's my conversation with @jeffbezos, founder of Amazon and Blue Origin. This is his first time doing a long-form conversation of this kind, and it was an epic one. It's here on X in full & is up on YouTube, Spotify, and everywhere else. Timestamps: 0:00 - Introduction 0:24 - Texas ranch and childhood 4:02 - Space exploration and rocket engineering 16:36 - Physics 26:10 - New Glenn rocket 1:08:59 - Lunar program 1:18:55 - Amazon 1:36:16 - Principles 1:54:56 - Productivity 2:05:34 - Future of humanity

Running and swimming are 100% where I get my best ideas, I just run a bit slower to make sure I keep some blood for the brain. You just need to find what’s best for you.

"The best thing founders can do is subtraction. It's much much easier to add things than it is to remove." Tobi is acutely brilliant & focused; it's why working with him the last 12yrs has been the privilege of my life. Great episode @tobi & @shaneparrish bit.ly/3TyLXAk

My Ted talk is on YouTube. I did not watch it so I’m not sure if any good. But I tried to envision what the future of AI-powered software creation will look like in the next few years, and what it means for the world. Of course, the best way to predict the future is to build it.

5-6 years ago I was working on music generation at DeepMind, but let me tell you, this is... something else. Incredibly excited to be able to finally share what our team has been working on!

Italian singer Adriano Celentano released a song in the 70s with nonsense lyrics meant to sound like American English, apparently to prove Italians would like any English song. It was a hit

A reminder that people can disagree about important things but still be good friends.