Wrapping my head around AI wrappers(wreflection.com)
wreflection.com
Wrapping my head around AI wrappers
https://www.wreflection.com/p/wrapping-my-head-around-ai-wrappers
38 comments
I would argue that a "wrapper" denotes a product that could be replicated with little effort, not that one relies on the other. None of your examples fit this definition.
I believe the controversy arises from the notion of “little effort” and critics who have never independently pushed anything to market. It comes across as dismissive and arrogant simply because someone exudes excessive confidence in a limited set of skills. I can personally attest to the immense demands of building a successful business, and it’s evident that very few individuals possess the capability to achieve that. Therefore, while it may provide comfort to avoid challenging oneself and dismiss others’ total work, it ultimately doesn’t benefit anyone and feels more like a self-serving “I could, but I never will” attitude.
The potential customer rarely cares whether a service provider is running their business well. What matters is the product's value added and risks added, as compared to just using the underlying tech directly.
The processor is built of transistors, built of silicone. The paper that wraps the box that wraps the processor is simply a mindless container. Yes, there’s nuance when it comes to “wrapper” companies but it in the end they may just be wrappers. Back in the “web 2.0” things were called mash-ups and everybody didn’t try to make them look like companies.
> But Cursor and other such tools depend almost entirely on accessing Anthropic, OpenAI and Gemini models, until open-source open-weight and in-house models match or exceed frontier models in quality.
I'm not sure I agree with this because even though Cursor is pay north of 100% of revenues to Athropic, Anthropic is selling inference at a loss. So if Cursor builds and hosts its own models it still has the marginal costs > marginal revenues problem.
The way out for Cursor could be a self-hosted much smaller model that focuses on code, and not the world. This could have inference costs lower than marginal revenues.
I'm not sure I agree with this because even though Cursor is pay north of 100% of revenues to Athropic, Anthropic is selling inference at a loss. So if Cursor builds and hosts its own models it still has the marginal costs > marginal revenues problem.
The way out for Cursor could be a self-hosted much smaller model that focuses on code, and not the world. This could have inference costs lower than marginal revenues.
Can you have a useful code model that doesn't understand the world? It seems like such a model would be limited to little more than auto complete.
what makes you so sure of this?
> Anthropic is selling inference at a loss.
cost of models have gone down dramatically over time.
> Anthropic is selling inference at a loss.
cost of models have gone down dramatically over time.
True, but they keep using more tokens (Agentic FTW!) and the currently most expensive models.
source: https://www.wheresyoured.at/
source: https://www.wheresyoured.at/
this is not true. they use more tokens to get more performance - the cost of using the model is going up but the performance is going up with it.
Your observation does not align with recent Cursor price changes.
I’m talking about metered billing and not subscription.
cost goes down by 50% (it didn't) but you fake improvement while lacking any real improvement by taking the average of 4 runs. Cost just doubled, even though inference technically halved.
Never mind the fact that the current model is alwayj outdated, and a new (bigger one) one is always being trained in parallel with the supposedly "cheaper" inference.
Never mind the fact that the current model is alwayj outdated, and a new (bigger one) one is always being trained in parallel with the supposedly "cheaper" inference.
I suppose supermaven is d'oing something to that effect.
> But I think the insight lies between these positions. Even if a new application starts as a wrapper, it can endure if it lives where work is done, writes to proprietary systems of record, builds proprietary data and learns from usage, and/or captures distribution before incumbents bundle the feature.
Basically the same as MS & Social Media did: build a proprietary silo around data, amass enough data, so it will become too big an inconvenience to move away from the first provider.
It's good that the EU has laws now to ensure data interoperability, export & ownership.
Basically the same as MS & Social Media did: build a proprietary silo around data, amass enough data, so it will become too big an inconvenience to move away from the first provider.
It's good that the EU has laws now to ensure data interoperability, export & ownership.
I agree with you in spirit but this harms the potential for these new products to emerge. You’re saying you don’t want them to be able to accrue a data moat. It sounds good for user privacy and optionality later on but it makes it harder for these services to get started as they dont see that model as possible.
Marketing, UI, and brand matter a lot. Especially when all of the products are functionally "the same" to the average consumer, who doesn't care about technical details and benchmarks, etc. It reminds me of this great scene from Mad Men:
This is the greatest advertising opportunity since the invention of cereal. We have six identical companies making six identical products. We can say anything we want.
https://youtu.be/8SsnkXH2mQY?si=SWPOsGBel1yh3kMd&t=198
This is the greatest advertising opportunity since the invention of cereal. We have six identical companies making six identical products. We can say anything we want.
https://youtu.be/8SsnkXH2mQY?si=SWPOsGBel1yh3kMd&t=198
right, if you look at the largest food companies, they're all just wrappers around proteins and simple/complex carbs, yet some products can do so well, and others so poorly
Currently working on a SaaS app that could be called an "AI Wrapper". One thing I picked up on is once you start using AI tools programmatically, you can start doing far more complex things than what you can with ChatGPT or Claude.
One thing we've leaned heavily into was using Langgraph for agentic workflows and it's really opened the door to cool ways you can use AI. These days the way I tell apart an AI "Wrappers" vs "Tools" is what is the underlying paradigm. Most "wrappers" just copy the paradigm of ChatGPT/Claude where you have a conversation with an agent, the "tools" are where you take the ability to generate content and then plug that into a broader workflow.
One thing we've leaned heavily into was using Langgraph for agentic workflows and it's really opened the door to cool ways you can use AI. These days the way I tell apart an AI "Wrappers" vs "Tools" is what is the underlying paradigm. Most "wrappers" just copy the paradigm of ChatGPT/Claude where you have a conversation with an agent, the "tools" are where you take the ability to generate content and then plug that into a broader workflow.
> One thing we've leaned heavily into was using Langgraph for agentic workflows
Probably my single biggest mistake so far with developing LLM tooling so far has been to try to use Langgraph even after inspecting the codebase, because people I thought were smarter than me hyped it up.
Do yourself a favor and just write the plumbing yourself, it's a lot easier than one might think before digging into it, and tool calling is literally a loop passing tool requests and responses back and forth until the model responds, and having your own abstractions will make it a lot easier to build proper workflows. Plus you get to use whatever language you want and don't have to deal with Python.
Probably my single biggest mistake so far with developing LLM tooling so far has been to try to use Langgraph even after inspecting the codebase, because people I thought were smarter than me hyped it up.
Do yourself a favor and just write the plumbing yourself, it's a lot easier than one might think before digging into it, and tool calling is literally a loop passing tool requests and responses back and forth until the model responds, and having your own abstractions will make it a lot easier to build proper workflows. Plus you get to use whatever language you want and don't have to deal with Python.
I can recommend looking into DSpy, I haven’t felt the need to use any other LLM based frameworks though I‘m open to suggestions
There really isn't need, all they add is additional code to be responsible for, building the same abstractions yourself but focused on your use case will be something like 50-100 lines of code, hard to beat the simplicity, and the understanding you'll get.
Almost every startup is a wrapper of some sort, and has been for a while. The reason a startup can startup is because it has some baked in competency by using new and underutilized tools. In the dot com boom, that was the internet itself.
Now it's AI. Only after doing this for 20+ years do I really appreciate that the arduous process and product winnowing that happens over time is the bulk of the value (and the moat, when none other exists).
Now it's AI. Only after doing this for 20+ years do I really appreciate that the arduous process and product winnowing that happens over time is the bulk of the value (and the moat, when none other exists).
Cant help myself and compare to frameworks, libraries and oop... cant we built so fast because of them?
I think of wrapper more as a very thin layer around. Thin layer is easy to reproduce. I do not question that a smart collection of wrappers can do great product. Its all about idea :)
However its if ones idea is based purely on wrappers there's really no moat, nothing stopping somebody else to copy it within a moment
I think of wrapper more as a very thin layer around. Thin layer is easy to reproduce. I do not question that a smart collection of wrappers can do great product. Its all about idea :)
However its if ones idea is based purely on wrappers there's really no moat, nothing stopping somebody else to copy it within a moment
> Almost every startup is a wrapper of some sort, and has been for a while.
The difference is that with AI they will send your data to a third party.
The difference is that with AI they will send your data to a third party.
as others have mentioned -- I think wrapper is a fair term. It is not trivial and took untold man hours of research and labour to go from nvidia gpus to modern llms. some of the ai products really do feel like minimal engineering around calls to openai (or claude or what have you)
Software is 10x more valuable than inference tokens because tokens do nothing for the user, just like a database request by itself does nothing.
Software is what makes inference valuable because it builds a workflow that transforms tokens and data into practical benefits.
Look at the payment plans for Lovable, Figma Make, Claude Code. None of them charge by token. They charge by obfuscated 'credits'. We don't know the current credit economics, but it is certain that the credit markup will increase and probably eventually reach 10x of the token cost. Users will gladly pay for it because again, tokens do nothing for them. It is the Claude Code, Figma Make products that make them productive.
Software is what makes inference valuable because it builds a workflow that transforms tokens and data into practical benefits.
Look at the payment plans for Lovable, Figma Make, Claude Code. None of them charge by token. They charge by obfuscated 'credits'. We don't know the current credit economics, but it is certain that the credit markup will increase and probably eventually reach 10x of the token cost. Users will gladly pay for it because again, tokens do nothing for them. It is the Claude Code, Figma Make products that make them productive.
The claim is too absolute. Software amplifies value, but inference cost and capability still shape what’s possible. Users aren’t demanding obfuscation; they just want predictable pricing and clear ROI. Does anyone want hidden math in their pricing?
In many markets, transparency wins. Think of Carfax or banking fees or airbnb pricing for example, when regulators or competitors force clarity, buyers benefit and trust grows.In a functioning government that serves the people (regardless of party) we would see this
People believe they “need” these AI products partly because they’re saturated in both earned and paid media. In '23 there were nearly 400k articles covering AI. I think we can all safely assume its more now, and when we include financial reporting, quite inescapable.
In many markets, transparency wins. Think of Carfax or banking fees or airbnb pricing for example, when regulators or competitors force clarity, buyers benefit and trust grows.In a functioning government that serves the people (regardless of party) we would see this
People believe they “need” these AI products partly because they’re saturated in both earned and paid media. In '23 there were nearly 400k articles covering AI. I think we can all safely assume its more now, and when we include financial reporting, quite inescapable.
What do we call companies that spin up open-source ML models, and basically just sell access to said models?
One example is audio / stem separation, object segmentation.
They're not wrappers, but whatever that is one step deeper down in complexity.
One example is audio / stem separation, object segmentation.
They're not wrappers, but whatever that is one step deeper down in complexity.
I think those are just inference providers unless they are adding something on top of it.
i recently framed this as "agent labs" vs "model labs" - https://latent.space/p/agent-labs - definitely far from proven or given that they are a lasting business model, but i think the dynamic is at least more evident now than it was a year ago and even that is notable as we are slowly figuring out what the new ai economy looks like
It’s hard to count how many ffmpeg wrappers are there
A Large Language Model is just a Large Hadron Model with better marketing.