Modern AI tools can recall and rearrange a staggering amount of information, but recall is only part of intelligence. A new company called Dyssonance is betting that the real leap forward will come from teaching machines to deal with contradictions.
According to The Next Web, the argument comes from Dr. Jennifer Prendki, a former Google DeepMind specialist who was responsible for major decisions about the training data behind Gemini. In a contributed piece for the publication, she explains why she founded Dyssonance and why she believes today's approach to AI has reached a conceptual limit.
The core of her thinking can be summed up in a single sentence: remembering is not the same as understanding. Current systems such as ChatGPT, Claude and Gemini are very good at the first. They can absorb huge volumes of text, answer difficult questions and keep up a natural conversation. Where they still fall short, Prendki argues, is in making sense of conflicting information, noticing when they have made a mistake and steadily revising what they know.
She offers a simple analogy. Imagine someone who reads a great many books but never connects the ideas in them, never questions the contradictions between them and never changes their own views as a result. That person would know a lot, but it would be hard to say they understood much. True intelligence, in this framing, means grasping how facts fit together, spotting when something does not make sense and adjusting as new information arrives.
How does this work in practice? The starting point is the observation Prendki made while working with training data: real-world data can never be perfectly cleaned. The world is complex, ambiguous and constantly shifting, so there is no inherent ground truth hidden inside any dataset. Every model trained on real information inherits that ambiguity. Adding more data or more memory does not remove the problem; it simply stores more of it.
What it depends on, then, is how a system treats inconsistencies. Prendki points out that science, detective work and entrepreneurship all tend to begin at the moment when something does not add up. A scientist notices a result that contradicts theory; a detective notices a detail that breaks the story; an entrepreneur notices a gap that the market has missed. In each case the contradiction is the trigger for real insight. According to her, AI can already flag such dissonances, and the next step is to make it work to resolve them rather than ignoring them or quietly storing them alongside everything else.
Dyssonance's answer is what the company calls a common-sense reasoning engine. Instead of acting as a vast archive of facts, the system is designed to develop and refine its own understanding, holding interpretations that evolve as fresh evidence emerges. Recognising what does not make sense, and working out why, is presented as the essence of intelligence itself.
A further element of the approach is described as engineering intuition. Rather than building a single system that tries to do everything, Dyssonance aims to create a shared foundation that different AI capabilities can draw on. According to Prendki, this should bring three benefits: AI that adapts more easily, reasoning that is easier to evaluate, and capabilities that are simpler to improve and combine with one another.
When is such an approach most useful? Precisely in the situations that current systems handle worst: ambiguous cases, conflicting sources and information that changes over time. These are also the conditions people face most often in the real world, where neat, unambiguous answers are rare.
The company's stated mission is to rethink the foundations of AI. In Prendki's view, the systems of the future will not be defined by bigger models, faster answers or larger memory, but by their ability to keep questioning, learning, adapting and making better sense of the world around them. As she puts it, the future of AI is about understanding better, not remembering more.