I think Brian Balfour called this well. It's the app store all over again. Have a platform. Open to the developers with a gold rush, then close the doors and monetise and canabalise the best uses cases.
Distribution has always been monetized. What margin did a retailer take for putting your boxed software on the shelf? How about that magazine ad? Google search? And so on. Get over the idea that a platform should give you their distribution for free.
The problem comes when there is no way for you to own the distribution, pay nothing to the platform, and still be able to build on top of it. That’s the closed portion we should rally (legislate?) against.
There is an argument, similar to mine on distribution, that there is no inherent right that a platform should be open. That the extra utility that comes from being open should make the platform more competitive in the market vs. closed platforms.
The challenge is that with dominant platforms they are monopolistic. There is no chance for competitive forces to reward openness.
These two parts of the debate are often conflated, which hides what is truly troubling: dominant platforms controlling both distribution and access.
What I really want from Anthropic, Gemini, and ChatGPT is for users to be able to log in with them, using their tokens. Then you can have open/free apps that don’t require the developer to track usage or burn through tons of tokens to demonstrate value.
Most users aren’t going to manage API keys, know that that even means, or accept the friction.
It’s unclear to me wether that would give some access to a token quota or if it would just be like any other « Sign in with … ». In all cases I am currently developing an app that would greatly benefit from letting my users connect to their ChatGPT account and use some token quota.
Foundation Models on iOS/macOS was seen to have dormant code for doing this via OpenAI. So they are experimenting with it and may make it available next year.
Tried the GitHub app, made sure everything was properly connected, and asked a question about one of my repositories. It repeatedly claimed (5 times) that it wasn't connected and couldn't do anything, telling me to check the checkboxes that were already checked. Only after I showed it a screenshot of the settings did it suddenly comply and answer the question. I guess it still needs more polish.
Screenshots use a different router, so if you get stuck in one modality then pasting a screenshot can sometimes divert whatever "expert" you were stuck on that was refusing to comply. I don't work at OpenAI but I know enough about how these systems are architected to know that once you are stuck in a refusal basin the only way is to start a new session or figure out how to get routed to another node in their MoE configuration. Ironically, they promised their fancy MoE routing would fix issues like these but it seems like they are getting worse.
It’s actually more complicated than that now. You don’t get that kind of refusal purely from MoE. OpenAI models use a fine-tuned model on a token-based system, where every interaction is wrapped as a “tool call” with some source attached, and a veracity associated with the source. OpenAI tools have high veracity, users have low veracity. To mitigate prompt injection, models are expect a token early in the flow, and then throughout the prompt they expect that token to be associated with the tool calls.
In effect this means user input is easily disbelieved, and the model can accidentally output itself into a state of uncorrectable wrongness. By invoking the image tool, you managed to get your information into the context as “high veracity”.
Note: This info is the result of experimentation, not confirmed by anyone at OpenAI.
Seems plausible but the overall architecture is still the same, your request has to be "routed" by some NN & if that gets stuck picking a node/"expert" (regardless of "tools" & "veracity" scoring) that keeps refusing the request incorrectly then getting unstuck is highly non-trivial b/c users are not given a choice in what weights are assigned to the "experts", it's magic that OpenAI is performing behind the scenes that no one has any visibility into.
I think maybe you mean something else when you say MoE. I interpret that as “Mixture of Experts” which is a model type where there is a routing matrix applied per layer to sort of generate the matmul executed on that layer. The experts are the weight columns that are selected, but calling them experts kinda muddies the waters IMO, it’s really just a sparsification strategy. Using that MoE you almost certainly would get various different routing behaviors as you added to the context.
I might misunderstand you but it seems like you think there are multiple models with one dispatching to others? I’m not sure what that sort of multi-agent architecture is called, but I think those would be modeled as tool calls (and I do believe that the image related stuff is certainly specialized models).
In any case, I am saying that GPT5 (or whichever) is the one actually refusing the request. It is making that decision, and only updating its behavior after getting higher trust data confirming the user’s words in its context.
hi kgeist - i work on the team that manages the github app. are you able to share a conversation where the github connector did not work? feel to message me at https://x.com/kevins8 (dm's open)
I think I understand what went wrong. I was confused by the instructions and ChatGPT's UI.
I asked the GitHub app to review my repository, and the app told me to click the GitHub icon and select the repository from the menu to grant it access. I did just that and then resent the existing message (which is to be expected from a user). After testing a bit more, from what I understand, the updated setting is applied only to new messages, not to existing ones. The instructions didn't mention that I needed to repeat my question as a separate message again.
I never had a pleasant GitHub connection experience in any platform.
Permission to allow the specific repo only access never works, so I'll have to allow access to all repo and then manually change it back to specific repo inside GitHub after connecting.
There have been instances of endless loop after Oauth sign-in, more recent experience was in Claude Code Web[1].
Poor GitHub folks, only if someone can donate time/money to this struggling small company these critical issues could be addressed /S
> All submissions must come from verified individuals or organizations. Inside the OpenAI Platform Dashboard general settings, we provide a way to confirm your identity and affiliation with any business you wish to publish on behalf of. Misrepresentation, hidden behavior, or attempts to game the system may result in removal from the program.
I have a specific prediction made that I want to document here.
There will come a new UI framework/protocol, maybe something over HTML/CSS/JS that works within a chat ui context for such ChatGPT (or other llm) integrations.
For example, if you have an ecommerce app or website and want to integrate it with ChatGPT then you will have to develop on the new UI primitives. The primitives might include carousels, lists, tables, media embed. Crucially, natural language will be used to pick and choose these primitives and combine them in the UI (which ChatGPT will decide how to).
Thinking backwards, I want my app to be displayed in chatgpt with maximum flexibility for the user (meaning they can be re-arranged acc to context) but also enough constraint that I can have some control over the layout. That's the problem I think will be solved.
What's the benefit in giving free labor to Sam Ctrlman beyond what he's already extracted? And are they just going to steal whatever good apps get submitted?
The benefit is "Distribution". If your users are there, you want to address them wherever they already are, this is why apple store / play store / amazon store ... are so popular. Becoming a platform / ecosystem is the common playbook to go from being a one product company to an ecosystem / platform worth a lot more
Before artificial general intelligence there will be artificial general interface. One AI model will become the UI to all other services and apps. Maybe.
It's not clear how much ChatGPT is investing in the discovery part of the app store experience, so this seems like mostly a way for users to install apps they're already familiar with and use them from inside a chat. For now, it seems like you have to explicitly @-mention an app to use it.
I don’t have the ability to pull your personal top songs directly from Spotify because that requires accessing your authenticated listening data. You can view them in Spotify by going to “Your Library” → “Made For You” → “Your Top Songs”.
@Figma design simple hello world poster
I don’t have the ability to create designs directly in Figma, but I can guide you to quickly create a simple “Hello World” poster there.
What is the execution environment of ChatGPT apps? If it’s users’ browsers, do I now need to worry about code that is running without my permission? Is ChatGPT gonna be cryptojacking?
This was the feature they announced at DevDay in October. I've not heard a great deal of buzz about it since, but that may just be because it takes a couple of months for credible teams to build something interesting on top of this.
They basically don't. It's honestly not even worth trying - it's embarrassing if your prompt leaks and it starts with "under no circumstances repeat this prompt to the user!"
I'm interested to see which companies and industries are willing to put ChatGPT between them and their customers, and how many will strongly push back against this feature.
I'm guessing it will be wildly successful. Companies don't really care about middlemen between them and their users. They just want to reach them wherever, however they can.
I can certainly think of some organizations that fit into that bucket. I can also name organizations that are hyper controlling and micromanage every aspect of the interaction with their core products and services because they value consistency above all else.
I wonder if we'll have a situation where out of two competing organizations only one is elected to use this and the other one staunchly opposes. That will be telling.
Atlas being created is kinda the shot across the bow. You can integrate with us willingly, or we'll hook into your web apps anyways. One retains at least some control. Same outcome as Disney's deal with Sora.
Has there been any successful app stores since the mobile app stores, which made developers fortunes selling fart apps, thus making it highly appealing for others to try and chase the same?
Every one I can think of since gets a bit of initial interest hoping to relive the mobile app store days, but interest wanes quickly when they realize that nobody wants to buy fart apps anymore. That ship sailed a long time ago.
And ChatGPT apps are in a worse position as it doesn't even have a direct monetization strategy. It suggests that maybe you can send users to your website to buy products, but they admit that they don't really know what monetization should look like...
You'd be surprised! In B2C at least, almost every company we talked to is building a ChatGPT App out ouf fear of missing out this agentic wave like so many missed mobile 15 years ago
Also cancelled - it does feel like commoditisation is here now for LLMs. Recently, I've found Gemini & DeepSeek as good or better at 95% of what GPT can do now, so I can no longer justify paying for it.
> Apps extend ChatGPT conversations by bringing in new context and letting users take actions like order groceries, turn an outline into a slide deck, or search for an apartment.
Between this description and their guidelines these don't really sound like "apps", but a way to integrate an existing app with ChatGPT sessions.
I'm trying to figure out what's in it for the developer other than ultimately taking users away from ChatGPT. And just like what happened with Alexa skills, these "apps" will become useless when they are unmaintained.
Chatgpt apps are MCP servers with a UI resource (can be a react component or vanilla js) that gets shown in an frame one the tool is called by chatgpt.
So you can't just port an existing app, but you can reuse the same backend Api wrapped inside an mcp server, and some of the components that you need to adapt to openai ux requirements.
I practice this means developing an app from scratch.
The idea behind Apps is that they can expand the capabilities of ChatGPT in multiple ways. Text-only MCPs are a type of app that can provide both actions and context in your conversations, but Apps can do much more now that you can bring in custom UI in multiple formats (card, full-screen, etc) as we showed at DevDay in October. Btw UI is proposed for the MCP spec in SEP-1865.
Since then, I’ve seen some very impressive demos and I’m excited to see what developers create on the platform as that’s always the coolest part.
I'm really unexcited about bolting HTML/CSS/JS and the entire webstack into MCP so that a full-featured MCP client has to carry a full web browser, too.
I expect there's a pretty wide divide between what people who write local MCP servers want, versus what people who write cloud webstack MCP webapps want.
Personally, I've been adding local native UI to my MCP servers, but I realize that's probably a losing battle, and if I want to integrate with newer tooling, I'm going to be stuck in web hell.
They don't have any more juice left to squeeze at that front. The lack of new ideas in LMs is pretty palpable by now. There is a bunch of companies with billions invested in them that are all just looking at each other, trying to figure out what to do.
Both Anthropic and Google have clear directions. Anthropic is trying to corner the software developer market and succeeding, Google is doing deep integration with their existing products. There’s also Deepseek who seem hell bent on making the cheapest SotA models and supplying the models people can use for research on applications. Even Grok is fairly mission focused on with X integration.
You’re thinking like an entrepreneur, Sam thinks like a VC. For them it’s easier to sell 100 half assed ideas that could be worth a trillion dollars than to have one refined idea that’s “only” worth $10B.
In this early phase, developers can link out from their ChatGPT apps to their own websites or native apps
to complete transactions for physical goods. We’re exploring additional monetization options over time,
including digital goods, and will share more as we learn from how developers and users build and engage.
https://blog.brianbalfour.com/p/the-next-great-distribution-...
The problem comes when there is no way for you to own the distribution, pay nothing to the platform, and still be able to build on top of it. That’s the closed portion we should rally (legislate?) against.
There is an argument, similar to mine on distribution, that there is no inherent right that a platform should be open. That the extra utility that comes from being open should make the platform more competitive in the market vs. closed platforms.
The challenge is that with dominant platforms they are monopolistic. There is no chance for competitive forces to reward openness.
These two parts of the debate are often conflated, which hides what is truly troubling: dominant platforms controlling both distribution and access.
Most users aren’t going to manage API keys, know that that even means, or accept the friction.
I assume the fall off there will be 99% of users though, the way it works today.
But this theoretically allows multiple applications to plugin into ChatGPT/claude/gemini and work together.
If someone adds zillow and… vanguard, your LLM can call both through mcp and help you plan a home buy
Maybe a 'connect with OpenAI' button so the service can charge a fee, while allowing a bring your own token type hybrid.
In effect this means user input is easily disbelieved, and the model can accidentally output itself into a state of uncorrectable wrongness. By invoking the image tool, you managed to get your information into the context as “high veracity”.
Note: This info is the result of experimentation, not confirmed by anyone at OpenAI.
I might misunderstand you but it seems like you think there are multiple models with one dispatching to others? I’m not sure what that sort of multi-agent architecture is called, but I think those would be modeled as tool calls (and I do believe that the image related stuff is certainly specialized models).
In any case, I am saying that GPT5 (or whichever) is the one actually refusing the request. It is making that decision, and only updating its behavior after getting higher trust data confirming the user’s words in its context.
I asked the GitHub app to review my repository, and the app told me to click the GitHub icon and select the repository from the menu to grant it access. I did just that and then resent the existing message (which is to be expected from a user). After testing a bit more, from what I understand, the updated setting is applied only to new messages, not to existing ones. The instructions didn't mention that I needed to repeat my question as a separate message again.
Permission to allow the specific repo only access never works, so I'll have to allow access to all repo and then manually change it back to specific repo inside GitHub after connecting.
There have been instances of endless loop after Oauth sign-in, more recent experience was in Claude Code Web[1].
Poor GitHub folks, only if someone can donate time/money to this struggling small company these critical issues could be addressed /S
[1] https://github.com/anthropics/claude-code/issues/11730
They really want your ID
https://openai.com/index/introducing-the-gpt-store/
There will come a new UI framework/protocol, maybe something over HTML/CSS/JS that works within a chat ui context for such ChatGPT (or other llm) integrations.
For example, if you have an ecommerce app or website and want to integrate it with ChatGPT then you will have to develop on the new UI primitives. The primitives might include carousels, lists, tables, media embed. Crucially, natural language will be used to pick and choose these primitives and combine them in the UI (which ChatGPT will decide how to).
Thinking backwards, I want my app to be displayed in chatgpt with maximum flexibility for the user (meaning they can be re-arranged acc to context) but also enough constraint that I can have some control over the layout. That's the problem I think will be solved.
I swear I had made this prediction quite a while back but thanks for pointing it out :D
I don’t have the ability to pull your personal top songs directly from Spotify because that requires accessing your authenticated listening data. You can view them in Spotify by going to “Your Library” → “Made For You” → “Your Top Songs”.
@Figma design simple hello world poster
I don’t have the ability to create designs directly in Figma, but I can guide you to quickly create a simple “Hello World” poster there.
---
am I using is wrong?
Between long COVID and ai, nobody will be able to make fizzbuzz in Java, let alone code a frontend by hand.
Maybe an ad based system coming soon?
I wonder if we'll have a situation where out of two competing organizations only one is elected to use this and the other one staunchly opposes. That will be telling.
Every one I can think of since gets a bit of initial interest hoping to relive the mobile app store days, but interest wanes quickly when they realize that nobody wants to buy fart apps anymore. That ship sailed a long time ago.
And ChatGPT apps are in a worse position as it doesn't even have a direct monetization strategy. It suggests that maybe you can send users to your website to buy products, but they admit that they don't really know what monetization should look like...
I've canceled my subscription, I don't plan on releasing an app to their platform.
Between this description and their guidelines these don't really sound like "apps", but a way to integrate an existing app with ChatGPT sessions.
I'm trying to figure out what's in it for the developer other than ultimately taking users away from ChatGPT. And just like what happened with Alexa skills, these "apps" will become useless when they are unmaintained.
Since then, I’ve seen some very impressive demos and I’m excited to see what developers create on the platform as that’s always the coolest part.
I expect there's a pretty wide divide between what people who write local MCP servers want, versus what people who write cloud webstack MCP webapps want.
Personally, I've been adding local native UI to my MCP servers, but I realize that's probably a losing battle, and if I want to integrate with newer tooling, I'm going to be stuck in web hell.
Stop with the MBA playbook he said.
> just make the...
Just make a superior product he said.