Independent developer Jeremy Morrell floated an idea on Simon Willison's blog this week: LLMs (large language models that let AI understand and generate text) have crushed the cost of writing software extensions from engineer-hours down to a few lines of natural-language prompts, while modern sandbox technology (running code in isolated environments so it can't pollute the host) makes user-written code safe enough to run. Taken together, we think this means: software companies no longer need to pile on features for every niche use case. They can ship a "reliable core" and let users assemble the rest with AI.
What this is
Morrell's central argument is a re-accounting of software extensibility. The extension mechanisms we know well — browser extensions, VS Code plugins, WordPress themes — were built for developers: high barrier to entry, small audience. But Morrell contends that LLMs lower the author barrier from "can code" to "can describe a need," and sandboxing lowers the runtime barrier from "trusted engineer" to "any user." An application's extension space is no longer defined by the vendor — it's defined by user imagination. He says this gives users "superpowers."
Industry view
Optimists read this as the biggest paradigm shift in software in 30 years: from "use what vendors build" to "AI builds what you want." The analogies stack up — the browser beat feature-frozen early portals on the strength of its extension ecosystem, VS Code grew into the largest dev tool on the back of plugins, and WordPress underpins half the internet.
But the skepticism is just as pointed. The security camp will ask: when non-engineers write LLM-generated code, even inside sandboxes, who defines behavioral boundaries? Data leaks, unauthorized API calls, denial-of-service — these old problems don't vanish just because AI wrote the code. The business camp is more anxious: SaaS (subscription-based online software) companies make money selling premium features. If users can AI-generate those features themselves, what happens to the subscription model? And the quality camp worries: user-written extensions conflict with each other, perform unevenly, and break with no safety net — software isn't Lego. When the blocks fall over, they don't just damage their own corner.
Impact on regular people
For enterprise IT: selection logic will shift. The next time we evaluate a SaaS, we won't just scan the feature list — we'll ask "can business units extend it themselves?" That hits traditional software vendors' product roadmaps hard.
For individual careers: non-technical roles will increasingly use AI to "write small plugins" inside their everyday tools. What used to require an IT ticket may now take a single sentence. That's an opportunity — and a jolt to existing IT workflows.
For consumer markets: we'll see this land first in "low-risk, high-customization" scenarios — personalized templates and automations inside note, spreadsheet, and design tools. Open-ended extension on mobile apps is still a long way off.