Very interesting takes! Some of them resonate. I've been using (and now working with!) vybe.build for the agent-team-single-player collaboration and it's worked perfectly. I don't think a super app is the solution, and I feel like context bloat is already an issue with specialized agents.
One last thing: I am 100% sure agent chat alone isn't a solution: whatever happens in a chat (eg. a whatsapp conversation) gets lost at some point. Humans still need apps. The winner will have both an agent and apps, and agents can operate apps!
I agree with much of this. The piece I think is still missing is authority. As agents become the default user, we need a way to preserve who retains the final say over real-world consequences. Capability is scaling quickly. Authority shouldn't silently scale with it.
Thanks for the thoughtful article, Peter. There's a lot to think about, though from context & personal perspective it feels like we're shifting/transitioning from Web2 to a complete redesign of browsing, UX & content delivery + processing. Who is to say HTML (visual representation of a webpage) will remain the default standard when parsing content from a server endpoint - overall there might not be a need for 50 % of the CSS embellishments that stylize contents. Where's the future of AI headed? That's not a simple Q that can be replied with the latest released model, in my opinion.
I think its pretty clear that the value of the frontier models is justified when used correctly. Even at pay-per-token rates, the cost is justified. Especially as the gap between open source and frontier models widens. I think the value of AI manifests at the edge of the capability not in the middle. Having Fable 5 draft your emails or navigate your ticket system is probably not a good use of that money, but having it design bespoke simulation software for mechanical engineering problems is absolutely an ROI. I think the companies that get buyer's remorse and react by switching to cheaper/opensource models are just being reactionary (and this reactionary approach probably set them up for failure when they adopted the frontier).
Peter's argument is essentially a UX coherence play saying that one capable coworker beats a fragmented suite. That's intuitive, but I think it misses the more important structural reason why separate product lines might actually make sense for Anthropic: they consume compute in fundamentally different ways.
Claude Code, Claude Design, and a general chat interface don't have the same inference cost profile. Bundling them into one super app doesn't make those differences disappear, it just makes them harder to manage, price, and optimise. We already saw what happens when you ignore that: Sora cost OpenAI roughly $1 million a day to run and generated $2.1 million in revenue over its entire lifetime before they killed it. That wasn't a branding problem. It was a compute dilution problem dressed up as a product strategy.
The Law of Line Extension — which Ries & Trout articulated thirty years ago and the AI industry is currently stress-testing in real time — suggests that focus usually beats expansion. But the corollary is that division often creates more value than consolidation, precisely because it forces clarity about what each product actually costs and what it actually earns.
I wrote about both of these dynamics (with the Sora numbers and the SpaceX S-1 segment breakdown) in my own series applying the 22 Immutable Laws of Marketing to the AI industry, if you want the longer argument: https://thequietedge.substack.com/p/the-22-immutable-laws-of-ai-vol-2
Well, if you look at it purely from an AI perspective (ie: the science, the R&D). Then I suppose it is logical, what you are describing.
However I think the financial underpinnings of AI development are extremely shakey and the next steps are going to be dictated by what is going on in the bond markets.
If rates rise and all that debt becomes a significant liability then the hyperscalers are going to suck pavement trying to make their over-leveraged investments pay off in a market that doesn’t have a big enough user base to justify the current investment.
Personally I think the next step for AI is going to be somewhere between bankruptcy & nationalisation, or extreme constriction of competition in the marketplace to attempt to leverage market share into financial health.
Very interesting takes! Some of them resonate. I've been using (and now working with!) vybe.build for the agent-team-single-player collaboration and it's worked perfectly. I don't think a super app is the solution, and I feel like context bloat is already an issue with specialized agents.
One last thing: I am 100% sure agent chat alone isn't a solution: whatever happens in a chat (eg. a whatsapp conversation) gets lost at some point. Humans still need apps. The winner will have both an agent and apps, and agents can operate apps!
Great read, thank you
sewage in, sewage out,
It’s the law of entropy.
I agree with much of this. The piece I think is still missing is authority. As agents become the default user, we need a way to preserve who retains the final say over real-world consequences. Capability is scaling quickly. Authority shouldn't silently scale with it.
Thanks for the thoughtful article, Peter. There's a lot to think about, though from context & personal perspective it feels like we're shifting/transitioning from Web2 to a complete redesign of browsing, UX & content delivery + processing. Who is to say HTML (visual representation of a webpage) will remain the default standard when parsing content from a server endpoint - overall there might not be a need for 50 % of the CSS embellishments that stylize contents. Where's the future of AI headed? That's not a simple Q that can be replied with the latest released model, in my opinion.
I think its pretty clear that the value of the frontier models is justified when used correctly. Even at pay-per-token rates, the cost is justified. Especially as the gap between open source and frontier models widens. I think the value of AI manifests at the edge of the capability not in the middle. Having Fable 5 draft your emails or navigate your ticket system is probably not a good use of that money, but having it design bespoke simulation software for mechanical engineering problems is absolutely an ROI. I think the companies that get buyer's remorse and react by switching to cheaper/opensource models are just being reactionary (and this reactionary approach probably set them up for failure when they adopted the frontier).
Good read. I'd push back on take #9 though.
Peter's argument is essentially a UX coherence play saying that one capable coworker beats a fragmented suite. That's intuitive, but I think it misses the more important structural reason why separate product lines might actually make sense for Anthropic: they consume compute in fundamentally different ways.
Claude Code, Claude Design, and a general chat interface don't have the same inference cost profile. Bundling them into one super app doesn't make those differences disappear, it just makes them harder to manage, price, and optimise. We already saw what happens when you ignore that: Sora cost OpenAI roughly $1 million a day to run and generated $2.1 million in revenue over its entire lifetime before they killed it. That wasn't a branding problem. It was a compute dilution problem dressed up as a product strategy.
The Law of Line Extension — which Ries & Trout articulated thirty years ago and the AI industry is currently stress-testing in real time — suggests that focus usually beats expansion. But the corollary is that division often creates more value than consolidation, precisely because it forces clarity about what each product actually costs and what it actually earns.
I wrote about both of these dynamics (with the Sora numbers and the SpaceX S-1 segment breakdown) in my own series applying the 22 Immutable Laws of Marketing to the AI industry, if you want the longer argument: https://thequietedge.substack.com/p/the-22-immutable-laws-of-ai-vol-2
Well, if you look at it purely from an AI perspective (ie: the science, the R&D). Then I suppose it is logical, what you are describing.
However I think the financial underpinnings of AI development are extremely shakey and the next steps are going to be dictated by what is going on in the bond markets.
If rates rise and all that debt becomes a significant liability then the hyperscalers are going to suck pavement trying to make their over-leveraged investments pay off in a market that doesn’t have a big enough user base to justify the current investment.
Personally I think the next step for AI is going to be somewhere between bankruptcy & nationalisation, or extreme constriction of competition in the marketplace to attempt to leverage market share into financial health.
I am less than hopeful tbh.