“We Were Looking at It Wrong”: The Explosive Claim That’s Upended the AI World
An exclusive interview with Nick Cammarata, the researcher whose 26,000-word viral blog post has AI safety folks celebrating.
By Will Douglas Heaven | October 19, 2026
In late 2026, as debates over AI safety and alignment reach a fever pitch, a single blog post has managed to shift the conversation. Nick Cammarata, a researcher known for his work on interpretability and mechanistic understanding of neural networks, published a 26,000-word essay that argues the field has fundamentally misunderstood how large language models (LLMs) operate—and why that misunderstanding matters for safety.
The post, titled “We were looking at it wrong,” has been shared widely across AI safety circles, drawing praise from researchers who say it offers a new lens on model behavior and a path forward for alignment work. In this exclusive interview, Cammarata explains the core claim, the evidence behind it, and what it means for the future of AI development.
The Core Claim: A Paradigm Shift
Cammarata’s central argument is that the AI community has been treating LLMs as complex, opaque systems that require increasingly sophisticated tools to interpret—when, in fact, the models are far simpler and more structured than commonly assumed. He contends that the field’s obsession with scale and emergent capabilities has obscured a more fundamental truth: that LLMs learn highly regular, interpretable algorithms that can be understood with the right conceptual framework.
“We were looking at it wrong,” Cammarata says. “We kept building bigger models and assuming that complexity would scale with size. But what we found is that the underlying computations are remarkably consistent—and remarkably simple. The complexity we see is often an artifact of our own analysis, not the model itself.”
Evidence and Implications
Drawing on years of research in mechanistic interpretability, Cammarata’s post synthesizes findings from sparse autoencoders, circuit analysis, and behavioral probing to argue that LLMs develop modular, reusable components that can be mapped and understood. This, he says, has profound implications for AI safety: if we can reliably identify and manipulate these components, we can better predict and control model behavior.
“The safety community has been celebrating because this reframes the problem,” Cammarata explains. “Instead of trying to align a black box through trial and error, we can potentially align it by understanding its internal structure. It’s a more tractable problem than we thought.”
Reception and Controversy
The post has sparked intense discussion. Some researchers hail it as a breakthrough that could accelerate interpretability research and provide a roadmap for safer AI systems. Others remain skeptical, arguing that the simplicity Cammarata describes may not hold at larger scales or in more complex tasks.
“There’s a lot of excitement, but also a lot of caution,” says one AI safety researcher who requested anonymity. “Nick’s work is rigorous, but scaling these insights to frontier models is a huge challenge. Still, it’s the most promising direction we’ve seen in years.”
Looking Ahead
As AI systems grow more capable and more integrated into society, the need for robust safety guarantees becomes ever more urgent. Cammarata believes his framework offers a way forward—but he cautions against overconfidence.
“We’ve only just begun to scratch the surface,” he says. “The claim isn’t that we have all the answers, but that we’ve been asking the wrong questions. If we can get the questions right, the answers will follow.”
For now, the AI world is taking notice—and the conversation is shifting.