A lot has changed in two years.
Not long ago, we were marveling at how good language models had become at sounding smart. Now, we’re starting to ask a different question: what if they could actually think?
The latest wave in AI isn’t about getting faster or cheaper—it’s about getting smarter. And not just in the pre-trained, pattern-matching way. We’re talking about models that reason in real time, that pause mid-task to weigh options, simulate outcomes, and choose the best path forward. That shift—from reflex to reasoning—is quietly redefining what’s possible.
At the center of this evolution is OpenAI’s new model, o1, also known as “Strawberry.” Unlike its predecessors, o1 does more than guess the next word. It stops to think. That might sound subtle, but it’s a tectonic shift in how AI operates. Instead of relying entirely on what it learned during training, o1 uses inference-time compute—extra processing power applied on demand—to reason through problems step-by-step.
This is the same idea that powered AlphaGo back in 2016. When it beat the world’s top Go player, it wasn’t just pulling from memory. It was simulating, evaluating, and strategizing in real time. Now, large language models are starting to do the same. They’re still early, and much better at tasks like coding or math than open-ended writing—but the ability to reason live is here, and improving fast.
And here’s where things get really interesting: models like o1 aren’t being dropped into the world as standalone products. Instead, they’re becoming the brains inside increasingly complex cognitive architectures—bespoke systems designed to handle real workflows in real industries.
We’re seeing early versions of this everywhere. A legal assistant that drafts arguments. A developer tool that refactors entire codebases. A customer support agent that doesn’t just chat—it resolves. These applications look nothing like traditional software. They’re built around outcomes, not interfaces. They don’t sell “seats”—they sell work. You give them a job, they get it done.
This changes the rules for startups. Competing on models is tough—OpenAI, Google, Anthropic, and Meta have a head start and war chests to match. But applications? That’s still open territory. Building domain-specific reasoning systems for the messy real world is hard—and it’s where startups thrive.
Of course, there are questions. Will incumbents adapt quickly enough? Can cognitive architectures scale across industries? Is this shift as big as the leap to the cloud?
We don’t know yet. But the clues are compelling. Just as the SaaS transition required rethinking everything from pricing to distribution, the AI transition is forcing companies to rethink how software works. From pattern mimicking to problem solving. From “what do you want to do?” to “just tell me the result.”
And as inference-time compute gets cheaper and more available, we’ll start to see even more surprising capabilities emerge. Multi-agent systems. Auto-generated tools. Software that builds software.