A lone startup plans to upend a trillion dollar artificial intelligence (AI) sector by developing a new class of chips optimized to run a stack of software that is not based on the floating-point arithmetic that large language models (LLMs) rely on to probabilistically predict an outcome.
Counterintuitive Corp is instead developing an artificial reasoning unit (ARU) that it describes as a new approach to building a more deterministic alternative to LLMs that are designed to run on lower cost application-specific integrated circuits (ASICs), says company chairman Gerard Rego.
The company expects to be able to have software versions of those ASIC designs available next year for application developers to create an ecosystem of applications that will ultimately be deployed on ASICs later this decade, adds Rego.
At the core of this effort are 80 patents pending that span deterministic reasoning hardware, causal memory systems and software frameworks. Unlike graphics processing units (GPUs) used today to build and deploy LLMs, ARUs will be designed to execute causal logic, memory lineage and verifiable deduction directly in silicon, said Rego.
That approach not only promises to enable the output generated by AI applications to be much more reliable and accurate, it will substantially reduce the amount of energy required to build and deploy AI applications by reducing the need for GPUs, he added.
It’s not clear how feasible ARUs may prove to ultimately be, but the marriage of LLMs and GPUs was a happy accident that is now fueling massive amounts of investments in data centers. GPUs existed long before LLMs, which themselves are based on floating point algorithms that have been around since the 1980s. Counterintuitive is making a case for an alternative approach based on a reasoning chip that leverages memory that is designed from the ground up to generate a deterministic outcome.
In contrast, LLMs are flawed because each operation introduces rounding drift and order variance, meaning the same computation can yield different results across runs or machines. That issue makes it impossible to reproduce the same output, much less verify its accuracy, said Rego.
No amount of GPU resources can make up for unstable math based on floating point algorithms, he added. AI models, as result, have no memory, which means they predict the next token or frame without retaining the reasoning that led to the output created, which results in each output essentially overwriting its own logic, said Rego.
The goal, instead, should be to build AI systems that can reason with traceable logic and remember decisions in a way that can actually be reproduced, he added. In contrast, LLMs are essentially a black-box that mimics the input used to train it, noted Rego.
The implications of any alternative approach to building AI applications would, of course, be profound. ASICs, for example, could be produced at a much lower cost than GPUs using the same manufacturing techniques used to create general-purpose x86 processors. The overall size of the data center needed would also be substantially smaller. Finally, deterministic applications could be built in a way that doesn’t require any appreciation for prompt engineering to create an outcome, said Rego.
Regardless of the merits of this particular approach, the one thing that is apparent is that when it comes to AI, research and development initiatives are not being limited to GPUs.
