Technology

Why a Former Google DeepMind Data Lead Walked Away to Build Dyssonance

Artificial intelligence technology concept
Written by Keny

Some people leave big tech for a better offer or a bigger title. Dr. Jennifer Prendki left Google DeepMind because she was worried about where the way AI was being built would lead, and she has now founded a company to pursue a different path.

As Dr. Jennifer Prendki writes in The Next Web, her decision to leave was driven by a fear of the consequences of how AI was being developed. Her answer is Dyssonance, a venture built around the idea that AI should learn from contradictions instead of simply remembering more.

Her route into artificial intelligence was not a straight line. As a child, she wanted to understand the universe, and that ambition led her first to physics and only later, somewhat unexpectedly, to AI. What attracted her was the intellectual challenge of the field and the open question at its heart: what intelligence actually is. What kept her there, she explains, was something less comfortable, a growing unease about the way AI systems were being built.

At Google DeepMind, Prendki worked alongside researchers she describes as brilliant and ambitious. Her own responsibilities were weighty. She was in charge of major decisions about the data used to train Gemini, and she helped determine the data for models including Gemini and Gemma. Those questions, she writes, weighed on her heavily and often kept her up at night. She noticed that relatively few people around her seemed to be asking what would happen once ordinary users started relying on the models in their daily lives.

She stayed for a while in the hope that being on the inside would let her make a difference, even though there was no guarantee her concerns would be heard. Eventually she concluded that she could no longer meaningfully influence the outcome. The prospect that troubled her most was watching the consequences unfold later from the outside and regretting that she had left too late, or not acted at all.

Working closely with training data gave her a specific insight that now shapes her company. Real-world data, she realized, can never be perfectly cleaned, because the world itself is complex, ambiguous and constantly changing. "There is no inherent ground truth in data," as her own summary puts it. Any AI trained on real-world information therefore inherits that ambiguity, no matter how carefully the data is curated.

Leaving the company gave her the freedom to question the assumptions she had worked within for years. She knew firsthand where the limits lie in how large language models learn from data. Rather than build another, slightly better LLM, or create tools that patch the weaknesses of existing ones, she decided to return to first principles.

That decision became Dyssonance. The central idea is captured in one line: remembering is not the same as understanding. Today's best known systems, including ChatGPT, Claude and Gemini, can process enormous amounts of information, answer complex questions and hold fluent conversations. Yet, in Prendki's view, they still struggle with three things that matter for genuine intelligence: making sense of conflicting information, recognising when they are wrong, and continuously updating what they understand.

Dyssonance is her attempt to address that gap. The company aims to build what it describes as a common-sense reasoning engine, a system that develops and refines its own understanding over time instead of simply storing more facts. Its interpretations are meant to keep evolving as new evidence appears, much as a person's views change when they learn something that does not fit what they believed before.

Prendki frames the mission in broad terms. The goal is to rethink the foundations of AI, so that future systems are defined not by bigger models, faster answers or more memory, but by their ability to keep questioning, learning, adapting and making better sense of the world. For someone who once lay awake over the data behind Gemini, the new venture is a way to act on those concerns directly. Her conclusion is simple: the future of AI is about understanding better, not remembering more.

About the author

Keny