Dyssonance AI Charts New Territory with a Physics-Based Model that Addresses Architectural Limitations of LLMs

The platform works as a common-sense reasoning engine that models thought to continuously interpret complex and ambiguous data, detecting and resolving inconsistencies as they emerge.

NEW YORK--(BUSINESS WIRE)--Dyssonance AI, a post-LLM research lab founded by quantum physicist and ex-DeepMind executive Jennifer Prendki, is coming out of stealth at just the right time, when band-aid fatigue is rampant, and the industry is eager for more reliable and robust architectural models. Rooted in the natural and physical sciences, Dyssonance’s team brings a physics-based approach to intelligence that derives hypotheses and provisional conclusions from raw data, much like the human brain does. The platform is now live and the waitlist is growing.

In line with the company’s ethos of challenging the status quo, it will reveal its technology through a Meta Murder Mystery experience, showing guests firsthand how the system reasons in real time. An exclusive group of 30 AI advisors, investors, journalists and researchers will gather in a landmark Tribeca building to find and observe different clues and tensions (dissonances), leading to varying belief states that affect their conclusions, something Large Language Models (LLMs) can’t do.

On- and off-Broadway actors from Broadway Murder Mysteries will help bring the entire experience to life, from the platform to the penthouse.

“We could have run through a deck, like most other companies, but having a room full of people interpret clues and find dissonances in the story is a much better way of showcasing our technology,” said Prendki, founder and CEO of Dyssonance AI. “A language model predicts the next word. Our system is built to infer who knows what and who believes what. When it notices something nobody has said out loud, I want people to wonder, ‘How did it know that?’”

How Does Dyssonance AI Differ From an LLM?

The human brain dislikes uncertainty and aims to reduce tension by resolving decisions quickly. Dyssonance AI identifies tensions and inconsistencies in the information it encounters and offers solutions based on inferred reasoning and what it calls synthetic intuition.

The platform is built to hold several conflicting beliefs at once and self-verify through factual reasoning as it learns something new, while LLMs need to be asked the right questions to make accurate predictions.

Because it doesn’t rely on tokens, Dyssonance AI uses significantly fewer resources than LLMs—which require immense amounts of electricity to train and run—it costs the end-user much less.

“While at DeepMind, I realized something that should be obvious to a physicist: ‘ground truth’ is a myth. Data is not reality; it is an observation of an event, which is incomplete and approximate. Intelligence cannot simply memorize from data. It has to infer the underlying state of the world that could have produced those observations. That makes intelligence a physics problem, not an information problem. And there is nothing artificial about it,” Prendki said. “Language models are just one piece of the puzzle. That’s why I call our work the physics of intelligence, and why I call us an intelligence company."

“I built Dyssonance AI from first principles, and where I landed happens to look remarkably similar to how the brain works; it’s multi-faceted and requires an astonishingly low amount of energy to power,” Prendki said. “Our physics-based approach can accomplish abstract reasoning that LLMs never will.” She added: "When you think about it, if we keep relying on people writing better prompts to highlight relevant context, this begs the question, ‘where is the intelligence really coming from?’”

A Platform, API and Agents that “Think Through” Data Instead of Merely Memorizing.

Dyssonance AI is launching a platform, an API and multiple agents. The agents include a command-line agent for developers and another agent for observability and inspection, both currently in beta.

The company’s framework is based on common-sense reasoning layers that gut-check data through evidence, existing knowledge and advanced logic. Built to infer human thought, context or underlying intent, the technology is incentivized to find coherence and tensions (dissonances) to solve problems and reach conclusions.

The company already has users and a waiting list across key verticals, such as Enterprise, Finance, Legal, Security, Tax and Trading. It also sees huge opportunities in Healthcare and other fields where ambiguity makes interpretation and judgment as important as knowledge.

“Enterprises are looking for systems capable of exercising judgment and reasoning under uncertainty, especially when information is incomplete or contradictory—which is actually how the world works,” said Carlos Lozano, head of product at Dyssonance AI. “But LLMs are unable to operate reliably in ambiguous situations. As a result, we end up telling agents what to do using agent harnesses. The problem is that ‘agent harness’ is an oxymoron: we want AI to have agency but then constrain it because we don’t trust it.”

Who Is Jennifer Prendki, Founder of Dyssonance AI?

Prendki trained as a particle physicist and earned her doctorate at the Sorbonne. She then spent nearly 20 years building and leading machine learning and data science teams, including as chief data scientist at Atlassian and in leadership roles at Walmart Labs (now Walmart Global Tech) and Figure Eight. She most recently led AI data at Google DeepMind.

At Dyssonance AI, she developed Continual Reasoning Under Uncertainty, a method that lets AI systems form and update beliefs as they deal with incomplete, unclear and changing information. Her time in AI at Google gave her direct insight into what works and what doesn’t, so she built Dyssonance AI with a strong focus on ethics, privacy and security.

How Can People Join the Waitlist?

The platform is live today for its first wave of customers in key verticals such as Enterprise, Finance, Legal, Security, Tax and Trading, and new customers can join the waitlist. The company’s website features articles, blog posts, research papers and use cases. It plans to host an event every six months to showcase its technology and highlight new customers, features and functionality.

About Dyssonance AI

Dyssonance AI is a post-LLM research lab founded by Jennifer Prendki, a quantum physicist who previously led AI data at Google DeepMind. Unlike LLMs, which predict the next word, its platform uses a physics-based approach to intelligence that models thought and builds stateful, common-sense reasoning layers incentivized to find coherence and tensions (dissonances). For more information, visit dyssonance.ai.