Stable Diffusion and What Generative AI Really Means

Watching DALL-E 2 launch in April was like watching the web browser launch in the 90s. Not because the technology was new—generative models existed before April—but because it suddenly became obvious to everyone that this was going to change what software could do.

I spent a weekend testing it. The results were stunning, but that's not what grabbed me. What grabbed me was the second-order thought: if a model can generate coherent images from text, what else can it generate? The implications rippled outward. Documentation. Scaffolding. Content. Entire workflows that assumed human time at the input layer could be reimagined.

By August, Stable Diffusion was open-sourced. By July, Midjourney was launching with Discord integration. February felt like the moment people started really grappling with what this meant for their businesses.

Generative AI Isn't About Art

The internet went wild with AI art. Forums filled with debate about whether it's "real" art, whether it's theft, whether artists should be concerned. All valid questions. But there's a distraction in that conversation.

Generative AI solving for art is the headline. Generative AI solving for everything else is the real business play.

Think about the content generation inside your product. If you're building commerce, you need product descriptions. If you're building SaaS, you need onboarding copy, placeholder content, instructional text. If you're building analytics software, you're generating reports and summaries. That work—the vast majority of content generation—doesn't require a human artist. It requires infrastructure.

Text-to-image is just the most visible instance of a broader pattern: systems that can turn high-level intent into finished output. That pattern scales far beyond image generation.

I'm watching it happen in our projects. A client building an e-commerce ERP needs to generate product images when new SKUs are added. Manual photography is bottlenecked. Now? Generate a reference image, adjust it with Stable Diffusion, done. Not photorealistic, not for a luxury brand, but sufficient for internal systems and faster than waiting for a photographer.

That's not science fiction. That's February 2022 economics.

What This Means for Product Architecture

Most founders building software today aren't thinking about generative models in their stack. They should be. Not because you need to build an AI company, but because your users are going to expect it.

Here's the pattern I see emerging: products that have a content generation bottleneck become candidates for generative layer insertion. If your software asks users to write something—product descriptions, email copy, chart labels, anything—you have an opportunity to offer generated suggestions and refinement rather than creation from scratch.

The architecture question becomes: where do you inject the model? In the browser? In the API? Do you call a hosted model or run it locally? Do you fine-tune for your domain or use a general model?

These are non-trivial questions, and different answers fit different constraints. But the fact that they're even questions you need to ask is new. Six months ago, generative models were research projects. Now they're commodity infrastructure.

We're exploring it in our ERP work. Imagine an inventory management system that suggests reorder quantities not just based on formulas, but by generating contextual reasoning about seasonal patterns, supplier lead times, and historical demand. Not as a black-box number, but as generated reasoning that a human can read and override.

That changes the UX. Instead of a form you fill in, you get a suggestion you refine.

The Scarcity Shift

Generative models are doing something profound: they're shifting scarcity upstream. You don't need scarce human writers anymore to fill your content gaps. You need scarce human judgment to shape what the models generate.

That's not the same as making writing obsolete. It's making bad writing cheaper, which means good curation—editing, refinement, taste—becomes the differentiator.

For product builders, this means: if you're competing on volume of content, generative models are a threat. If you're competing on curation, they're an tool.

The same applies to images, code, data. Models can generate draft code—GitHub Copilot launched last year—but they generate a lot of mediocre code too. The scarcity moves to developers who can read generated code and direct it toward good architecture.

That's actually good news for quality. It means the skill premium goes to people who can think systemically, not just execute mechanically.

The Immediate Opportunity

If you're shipping software in 2022, you have a window. Generative capabilities are still novel enough that offering them is a differentiator. In two years, users will expect it. In five years, it'll be table stakes.

The question isn't whether to use generative AI—it's when and where to offer it as a feature, and how to architecture your product so you can integrate it without tearing everything down.

For our ERP clients, we're starting with the lowest-friction entry points: description generation, report summarization, pattern spotting. Places where generative output is useful but not critical—wrong output has cost, but not catastrophic cost.

By the end of 2022, I expect that to feel quaint. The models are improving monthly. The infrastructure is getting cheaper. What felt like "this is interesting, we should explore it" in February is going to feel like "we need this for parity" by November.

That timeline is accelerating, and products that start thinking about generative layers now will have the architecture in place to move fast when it becomes urgent.