To Wrap or Not To Wrap

1. Delusions

When I told my colleagues I’d be pivoting to AI-powered business solutions, one remark kept finding its way to my ears — “So you’re doing GPT wrappers now?”

Time and time again, tech folks not actively working with AI keep saying the same hunk of baloney. It’s ear grating!

So for this issue of the newsletter, I thought I’d go ahead and explain what is and isn’t a wrapper. I’m going all in and even discussing AI Agents.

So, let’s get started.

2. The Modeling Business

We need to start off with the most basic concept out there — AI models. What are they?

Well, let’s tone things down a bit and start with something even simpler — what’s a problem made of?

A problem can be broken down into three pieces:

• Statement, the thing beckoning for an answer.

• Solutions, things that may or may not satisfy the problem.

• Constraints, restrictions or rules that impact the validity of solutions.

Now all an AI model does is cook up solutions for you given the statement and constraints.

And the reason why AI is getting so much attention is that we’re now able to use it to tackle problems we thought were far past the realm of computers.

The popular models, e.g. ChatGPT, basically solve “next word prediction” — they try guessing what text goes best with what they’ve seen. Viewing writing as a problem of this sort is what’s allowed AI models to grab attention left, right, and center.

Thinking of hitting the stocks? Quant firms like Renaissance are already running models to tackle the market. Heck, DeepSeek — the Chinese-made open-source LLM making waves for being on par with GPT for a fraction of the cost — was cooked up by quants in their spare time.

So that’s a quick TL:DR of what AIs are. If you’re in the mood for more details, might as well browse the rest of our newsletter issues if you haven’t already.

3. Call Me

At the end of the day, an AI model is still a piece of software. What’s the one thing you can always do with software?

That’s right — get it to work with other software.

The browser you’re reading this on is probably pinging a million different pieces to get things operational — your device’s display, the network route this text traveled along, and the several tiny components responsible for properly drawing this issue up to spec.

AI is no different — any bit of software should be able to use it.

If you have software that needs an AI to tackle a chunk of the work, it can call one up. That’s basically how APIs work.

So what’s a wrapper then?

Well, say you have a special bit of software — let’s call it Software X — that can handle the brunt of your work. You write up a bit of additional software, the wrapper, which does nothing but call on Software X and fetch the results.

There is no further element or addition of value involved — it’s just call and response.

So an AI wrapper or GPT wrapper is as the name implies — a bit of software that, under the hood, does nothing but delegate the work to its namesake without any real value addition.

So that’s the basic gist of what a wrapper is. What about everything else?

4. But I’m Not a Wrapper

The first key thing about a real AI-solution is deliberation. There’s a clear difference between actively designing with AI and simply injecting it.

It’s time to pull out the good ol’ AI tech stack again.

AI-powered applications don’t simply call an API or two and call it a day. Cutting edge solutions to real pain-points are planned out with AI in mind. In particular,

• How is the data to be used going to be acquired, stored, and processed?

• What kind of AI models are we using to get the job done?

• How do we keep track of the solution?

• How do we interact with the solution?

So the first major difference between a wrapper and a real AI-powered solution is the attention paid to AI when outlining the solution itself.

The tech-stack above isn’t universal due to how recent AI adoption for business problems is, but exhibiting deliberation beyond “let’s just call an API” already sets things apart from being a typical wrapper.

Heck, here’s a quick way to test if your case scenario

1. Is my problem well defined?

2. Is my problem difficult to solve algorithmically?

3. Are there other key mechanisms at play besides the AI?

If you answered ‘yes’ to all of the above, it’s not a wrapper. To insist otherwise is, at that point, ignorance — might as well call all applications database wrappers.

5. We Are Autonomous

Autonomy is what separates agents from wrappers — funnily enough, the idea of agents is older than LLMs!

An agent is, simply put, an autonomous unit capable of processing work all the while engaging with an environment it lives in.

What do we mean by ‘engaging with environment’? Well,

• An agent can monitor and assess processes in its ecosystem.

• An agent can operate off of stored, ‘long-term memory’.

• An agent can call on other software to complete and coordinate tasks.

and much more — an agent is essentially just a much smarter unit of software.

Of course, the good thing about agents is that you can have several of them. Breaking down tasks into separate chunks lends itself neatly to agent orchestration.

If you’re invoking agents, you aren’t using a wrapper. Agents take the earlier idea of an AI tech stack and overclock it.

6. Reload

So to recap the point, your solution isn’t a wrapper if the usage of AI is deliberate or the solution itself is fairly autonomous.

This whole spiel isn’t that far off from how web applications and cloud-based SaaSs work — we don’t say things are SQL wrappers if they’re using it under the hood, it’s understood as a vital architectural component.

Likewise, the perception around using AI needs to be shaped by borrowing from how the rest of the SaaS world works.

But until that day comes, guess I’ll be doing my part one moron at a time.

If you’re interested in AI agents for your organization or products, visit our website at antematter.io.