the most advanced generalist AI models often outperform specialized models, even in the specific domains those specialized models were designed for. That means that if you want a model that can do a lot - reason over massive amounts of text, help you generate ideas, write in a non-robotic way — you want to use one of the three frontier models: GPT-4o, Gemini 1.5, or Claude 3 Opus.

Working with advanced models is more like working with a human being, a smart one that makes mistakes and has weird moods sometimes. Frontier models are more likely to do extraordinary things but are also more frustrating and often unnerving to use. Contrast this with Apple’s narrow focus on making AI get stuff done for you.

Every major AI company argues the technology will evolve further and has teased mysterious future additions to their systems. In contrast, what we are seeing from Apple is a clear and practical vision of how AI can help most users, without a lot of effort, today. In doing so, they are hiding much of the power, and quirks, of LLMs from their users. Having companies take many approaches to AI is likely to lead to faster adoption in the long term. And, as companies experiment, we will learn more about which sets of models are correct.

Muse as a brand and a product represented something aspirational. People want to be deeper thinkers, to be more strategic, and to use cool, status-quo challenging software made by small passionate teams. These kinds of aspirations are easier to indulge in times of plenty. But once you're getting laid off from your high-paying tech job, or struggling to raise your next financing round, or scrambling to protect your kids' college fund from runaway inflation and uncertain markets... I guess you don't have time to be excited about cool demos on Twitter and thoughtful podcasts on product design.

I’d speculate that another factor is the half-life of cool new productivity software. Evernote, Slack, Notion, Roam, Craft, and many others seem to get pretty far on community excitement for their first few years. After that, I think you have to be left with software that serves a deep and hard-to-replace purpose in people’s lives. Muse got there for a few thousand people, but the economics of prosumer software means that just isn’t enough. You need tens of thousands, hundreds of thousands, to make the cost of development sustainable.

We envisioned Muse as the perfect combination of the freeform elements of a whiteboard, the structured text-heavy style of Notion or Google Docs, and the sense of place you get from a “virtual office” ala group chat. As a way to asynchronously trade ideas and inspiration, sketch out project ideas, and explore possibilities, the multiplayer Muse experience is, in my honest opinion, unparalleled for small creative teams working remotely.

But friction began almost immediately. The team lead or organizer was usually the one bringing Muse to the team, and they were already a fan of its approach. But the other team members are generally a little annoyed to have to learn any new tool, and Muse’s steeper learning curve only made that worse. Those team members would push the problem back to the team lead, treating them as customer support (rather than contacting us directly for help). The team lead often felt like too much of the burden of pushing Muse adoption was on their shoulders. This was in addition to the obvious product gaps, like: no support for the web or Windows; minimal or no integration with other key tools like Notion and Google Docs; and no permissions or support for multiple workspaces. Had we raised $10M back during the cash party of 2020–2021, we could have hired the 15+ person team that would have been necessary to build all of that. But with only seven people (we had added two more people to the team in 2021–2022), it just wasn’t feasible.

We neither focused on a particular vertical (academics, designers, authors...) or a narrow use case (PDF reading/annotation, collaborative whiteboarding, design sketching...). That meant we were always spread pretty thin in terms of feature development, and marketing was difficult even over and above the problem of explaining canvas software and digital thinking tools.

being general-purpose was in its blood from birth. Part of it was maker's hubris: don't we always dream of general-purpose tools that will be everything to everyone? And part of it was that it's truly the case that Muse excels at the ability to combine together so many different related knowledge tasks and media types into a single, minimal, powerful canvas. Not sure what I would do differently here, even with the benefit of hindsight.

Muse built a lot of its reputation on being principled, but we were maybe too cautious to do the mercenary things that help you succeed. A good example here is asking users for ratings; I felt like this was not to user benefit and distracting when the user is trying to use your app. Our App Store rating was on the low side (~3.9 stars) for most of our existence. When we finally added the standard prompt-for-rating dialog, it instantly shot up to ~4.7 stars. This was a small example of being too principled about doing good for the user, and not thinking about what would benefit our business.

Growing the team slowly was a delight. At several previous ventures, I've onboard people in the hiring-is-job-one environment of a growth startup. At Muse, we started with three founders and then hired roughly one person per year. This was absolutely fantastic for being able to really take our time to find the perfect person for the role, and then for that person to have tons of time to onboard and find their footing on the team before anyone new showed up. The resulting team was the best I've ever worked on, with minimal deadweight or emotional baggage.

ultimately your product does have to have some web presence. My biggest regret is not building a simple share-to-web function early on, which could have created some virality and a great deal of utility for users as well.

In terms of development speed, quality of the resulting product, hardware integration, and a million other things: native app development wins.

After decades working in product development, being on the marketing/brand/growth/storytelling side was a huge personal challenge for me. But I feel like I managed to grow into the role and find my own approach (podcasting, demo videos, etc) to create a beacon to attract potential customers to our product.

when it comes time for an individual or a team to sit down and sketch out the beginnings of a new business, a new book, a new piece of art—this almost never happens at a computer. Or if it does, it’s a cobbled-together collection of tools like Google Docs and Zoom which aren’t really made for this critical part of the creative lifecycle.

any given business will find a small number of highly-effective channels, and the rest don't matter. For Heroku, that was attending developer conferences and getting blog posts on Hacker News. For another business it might be YouTube influencer sponsorships and print ads in a niche magazine. So I set about systematically testing many channels.

We still believe that there will be a separation between the “application layer” companies and foundation model providers, with model companies specializing in scale and research and application layer companies specializing in product and UI. In reality, that separation hasn’t cleanly happened yet. In fact, the most successful user-facing applications out of the gate have been vertically integrated.

We predicted that the best generative AI companies could generate a sustainable competitive advantage through a data flywheel: more usage → more data → better model → more usage. While this is still somewhat true, especially in domains with very specialized and hard-to-get data, the “data moats” are on shaky ground: the data that application companies generate does not create an insurmountable moat, and the next generations of foundation models may very well obliterate any data moats that startups generate. Rather, workflows and user networks seem to be creating more durable sources of competitive advantage.

Some of the best consumer companies have 60-65% DAU/MAU; WhatsApp’s is 85%. By contrast, generative AI apps have a median of 14% (with the notable exception of Character and the “AI companionship” category). This means that users are not finding enough value in Generative AI products to use them every day yet.

generative AI’s biggest problem is not finding use cases or demand or distribution, it is proving value. As our colleague David Cahn writes, “the $200B question is: What are you going to use all this infrastructure to do? How is it going to change people’s lives?”

People forget that when Bezos introduced Amazon Prime, Wall St. flipped out, because they insisted that it would cost way too much for too little benefit. But, through it all Amazon survived (and thrived) because Bezos just kept telling investors exactly what his plan was, and never backed down, no matter what Wall St. kept saying to him.

This is too easily forgotten, but your users are everything if you run an internet business. They’re not “the product.” They’re what makes your site useful and valuable, and often provide the best marketing you could never buy by convincing others to join and providing you with all of the best ideas on how to improve things and make your service even better for the users. The moment you’re undermining your own community, you’re beginning to spiral downward.

As you’re developing a business model, the best way to make sure that you’re serving your users best, and not enshittifying everything, is to constantly make sure that you’re only capturing some of the value you’re creating, and are instead putting much more out into the world, especially for your community.

Push the power to make your service better out from the service to the users themselves and watch what they do. Let them build. Let them improve your service. Let them make it work better for you. But, you have to have some trust here. If you’re focused on “Rule 3” you have to recognize that sometimes your users will create value that you don’t capture. Or even that someone else captures. But in the long run, it still flows back to you, as it makes your service that much more valuable.

If you’re charging for something that was once free, you’re taking away value from your community. You’re changing the nature of the bargain, and ripping away the trust that your community put in you. Instead, always look for something new that is worth paying for above and beyond what you already offered.

There are ways to monetize that don’t need to overwhelm, that don’t need to suck up every bit of data, that don’t need to rely on taking away features users relied on. Focus on adding more scarce value, and figuring out ways to charge for those new things which can’t be easily replicated.

You start learning acronyms like “ARPU” (average revenue per user) and such. And then you’re being measured on how much you’re increasing those metrics, which means you need to squeeze more out of each individual user, and you’re now deep within the enshittification stage, in which you’re trying to squeeze your users for more money each quarter (because now everything is judged in how well you did in the last 3 months to improve that number).

What can you do with this thing? There’s a good chance that, whatever killer apps may emerge, they don’t need the entire complement of sensors and widgets to deliver a great experience. As that’s discovered, Apple will be able to open a second tier in this category and sell you a simplified model at a lower cost. Meanwhile, the more they manufacture the essentials—high density displays, for example—the higher their yields will become, the more their margins will increase. It takes time to perfect manufacturing processes and build up capacity. Vision Pro isn’t just about 2024’s model. It’s setting up the conditions for Apple to build the next five years of augmented reality wearable technology.

VR/AR doesn’t have to suck ass. It doesn’t have to give you motion sickness. It doesn’t have to use these awkward, stupid controllers you accidentally swap into the wrong hand. It doesn’t have to be fundamentally isolating. If this paradigm shift could have been kicked off by cheap shit, we’d be there already. May as well pursue the other end of the market.

what starts as clunky needn’t remain so. As the technology for augmented reality becomes more affordable, more lightweight, more energy efficient, more stylish, it will be more feasible for more people to use. In the bargain, we’ll get a display technology entirely unshackled from the constraints of a monitor stand. We’ll have much broader canvases subject to the flexibility of digital creativity, collaboration and expression. What this unlocks, we can’t say.