Simula: Building Synthetic Datasets That Actually Work for the Real World

Google Research just dropped a paper that makes me feel a bit better about where synthetic data is headed. It’s called Simula, and the core idea is refreshingly pragmatic: stop treating synthetic data generation like a brute-force sampling problem and start treating it like mechanism design.

If you’ve spent any time trying to build specialized AI models for niche domains, you know the pain. Internet-scale data works great for generalist models, but the moment you need something tailored — medical records, cybersecurity logs, financial transactions — you hit a wall. Real-world data for these domains is expensive, hard to access, and often locked behind privacy regulations.

The usual workaround is synthetic data, but most existing approaches are hacky. People throw prompts at LLMs, run evolutionary algorithms that feel like black magic, or rely on seed datasets that bring their own biases. None of this scales well, and none of it gives you fine-grained control over what the dataset actually contains.

Simula flips the script. Instead of optimizing sample by sample, it designs the entire dataset from first principles. The framework uses reasoning models to map out the conceptual space of a target domain — building hierarchical taxonomies that act as a scaffold for generation. This isn’t random sampling. It’s deliberate, structured, and auditable.

The key insight is that you want independent control over three axes: coverage (does the dataset span the long tail?), complexity (are you including edge cases?), and quality (is the data actually useful?). Simula lets you dial each of these separately, which is something I haven’t seen done cleanly before.

Here’s the part that caught my attention: Simula is seedless and agentic. It doesn’t need human-curated examples to get started. The system recursively expands categories, proposes sub-categories, and then critiques and filters them — all using reasoning models. This propose-and-refine loop builds dense taxonomies dynamically. The Cyber Threat Intelligence tree they show in the paper is a good example of how deep this can go without human intervention.

Is it perfect? No. The quality of the generated datasets is still tied to the reasoning capabilities of the underlying model. If your model makes logical errors, those errors propagate into the taxonomy. But this is a much better problem to have than relying on opaque evolutionary steps or manual prompt engineering. At least here, the reasoning trace is inspectable.

I also appreciate that the paper addresses the operational side. Synthetic data treated as code — versionable, reproducible, inspectable — is a workflow that actually fits into modern ML pipelines. The static nature of real-world datasets has always been a bottleneck for iteration speed. If Simula can deliver on that promise, it’s a win for teams that need to move fast without waiting months for data collection.

There’s a broader point here that I think gets lost in the hype around synthetic data. Most people talk about it as a way to generate “more data.” But more data isn’t the goal. The goal is better coverage of the scenarios that matter — especially the rare, dangerous, or privacy-sensitive ones that real-world data can’t cover. Simula’s mechanism design approach is a step toward making synthetic data a first-class tool for safety and robustness, not just a crutch for data scarcity.

Will this replace real-world data entirely? No, and it shouldn’t. But for domains where data is scarce, expensive, or sensitive, Simula offers a principled alternative that’s actually controllable and explainable. That’s more than most synthetic data methods can claim.