For several years, progress in artificial intelligence has largely been measured by scale: more data, larger models, longer memory. A former Google DeepMind data lead now argues that this race is missing the point, and has launched a company to prove it.
As covered by The Next Web, Dr. Jennifer Prendki founded Dyssonance after leaving Google DeepMind, where she was responsible for key decisions about the data used to train Gemini and helped shape the data behind models including Gemini and Gemma. Her new company is built on a different premise from most of the industry: that the next breakthrough will come from teaching AI to handle ambiguity, not from feeding it ever more information.
The timing of the argument is notable. Popular assistants such as ChatGPT, Claude and Gemini have become remarkably capable at processing large amounts of information, answering complicated questions and sustaining conversations. Prendki does not dispute that. Her point is that these systems still have trouble with tasks that people consider basic to real intelligence. They struggle to reconcile conflicting information, to recognise when they are wrong and to keep updating their understanding as circumstances change.
Behind this view is an insight from her years working with training data. Real-world data, she concluded, can never be perfectly cleaned, because the world it describes is complex, ambiguous and in constant motion. There is no inherent ground truth in data, and every model trained on it carries that ambiguity forward. From this perspective, collecting more data or expanding memory only enlarges the store of unresolved contradictions.
Prendki's decision to leave was itself shaped by these concerns. She writes that she left once she felt she could no longer meaningfully influence how things would turn out, and that she feared watching the consequences unfold later from the outside. Away from the company, she was free to question the assumptions she had worked within. She chose not to build another, better large language model or a set of tools to patch the weaknesses of existing ones. Instead, she decided to go back to first principles.
The result is Dyssonance's central bet: contradictions are where learning starts. In science, detective work and entrepreneurship, progress usually begins when something does not add up. Prendki argues that AI is already able to flag such dissonances, and that it should be built to work through and resolve them rather than ignore them or simply file them away. Recognising what does not make sense, and figuring out why, is described as the essence of intelligence.
To put that idea into practice, the company is working on what it calls a common-sense reasoning engine. Rather than storing facts, such a system would develop and refine its own understanding, with interpretations that continue to evolve as new evidence becomes available. It is a shift from treating AI as a giant memory to treating it as something closer to an investigator that revises its conclusions.
Another part of the plan is what Dyssonance calls engineering intuition. Instead of a single system attempting every task, the company wants to build a shared foundation that many different AI capabilities can rely on. Prendki lists the expected advantages: AI that is more adaptable, reasoning that is easier to evaluate, and capabilities that are easier to improve and to combine.
If this direction gains traction, it would change how progress in the field is measured. Success would depend less on how many parameters a model has or how much it can recall, and more on how well it copes with ambiguous, conflicting and changing information, which is exactly what real-world use tends to involve.
Dyssonance describes its mission as rethinking the foundations of AI. The systems it envisions are not simply bigger models, faster answers or deeper memory, but systems that keep questioning, learning, adapting and making better sense of the world. In Prendki's words, the future of AI is about understanding better, not remembering more.

