Counterintuitive Corp., an artificial intelligence startup that is building what it calls reasoning-native computing, launched today with a mission to redefine AI by developing chips and software designed for reasoning, not mimicry.

The startup is aiming to overcome what it calls the “Twin Trap” of AI — precision without truth and reasoning without memory — through a new class of computing architecture it calls the Artificial Reasoning Unit.

The problem Counterintuitive is seeking to address is one where modern AI runs on floating-point arithmetic that was built in the 1990s to address imaging and gaming with a numerical foundation designed for performance, not reproducibility. According to the company, each operation introduces rounding drift and order variance, meaning the same computation can yield different results across runs or machines. That creates a hidden instability the causes nondeterminism, which in turn undermines auditability and trust in critical applications.

The “precision without truth” issue has become, in Counterintuitive’s words, an “invisible wall of modern AI,” a wall that no amount of scale can break.

Thrown into the mix is reasoning without memory, the issue whereby current AI models are memoryless and predict the next token or frame without retaining the reasoning that led it there. Each output overwrites its own logic, producing fluent mimicry of logic without true understanding.

“These two traps are why even the largest AI systems lack reproducibility, are energy-intensive, require massive capital and training and are unable to reason,” the company says.

Counterintuitive’s reasoning-native computing research, on the other hand, is promising to push AI beyond its current limits by offering a reasoning chip and software reasoning stack that eliminate these traps.

The company’s proposed Artificial Reasoning Unit is a new class of compute, not a processor, that is intended to be the world’s first reasoning chip when launched and is being architected natively for deterministic, memory-driven reasoning. The ARU initiates the post-floating-point graphics processing unit era of computing. Unlike GPUs, ARUs are designed to execute causal logic, memory lineage and verifiable deduction directly in silicon.

The ARU is complemented with full-stack reasoning software, which is being designed to integrate hardware and software into a unified causal reasoning framework.

“Our ARU stack is more than a new chip category being developed — it’s a clean break from probabilistic computing,” explains Counterintuitive co-founder Syam Appala. “The ARU will usher in the next age of computing, redefining intelligence from imitation to understanding and powering the applications that impact the most critical sectors of the economy without the need for massive hardware, data center and energy budgets. A new era of intelligence-based reasoning is upon us.”

Image: Counterintuitive