Ethical AI: Achray Doesn't Need a Data Center (or Your Personal Data)

"Quantity has a quality all its own" and the public backlash is growing against the quality of AI and quantity of resources it demands. Computer types like us will tell you that LLM-type AI has an O(N²) operational cost and about O(N⁵) to O(N⁶) cost for training on N pieces of data. In real-life terms, this means the technology needs noisy, resource-hungry data centers that bring higher electricity bills and untold ecological consequences.

At a recent event, AI & Warfare: When Machines Make Decisions by the NATO Association of Canada, the discussions spurred us to catch up with some new research on AI ethics and governance. We realized that it was time to highlight and clarify the difference between traditional AI and Achray's work.

• We do not need a data center.

  We do not need the resources of the large AI companies. We are currently focused on execution and even learning on a 20W power budget for low-cost equipment. We aim to deploy at the location where the data is, not send it back to a data center for storage and processing.

• Our systems control for false confidence. User control over false confidence rates are fundamental to meeting any ethical and legal requirements.

  Our approaches are founded on the math of stochastic processes and adhere to the foundational principles of the statistical methods we use. We don't get into situations were our systems misbehave because we keep them true to a set of models whose behaviour is always known, even in high dimensional settings. We have transparency.

• We respect your privacy by never collecting personal data. We respect your copyrights.

  We don't scrape your social media, websites or code bases. We don't need to: sufficiently rich sources of information are publicly available. We don't even track you reading this website, there is no cookie consent form because there are no tracking cookies. When we use your confidential data, it is by permission, for a specific goal, and you retain control.

A 1943 paper by McCulloch and Pitt started the artificial neural network model that the mainstream LLM/AI providers use today. During 83 years of research, advances have been made and limitations to the model have been discovered (and often rediscovered). That brand of AI is coming at an immense and increasingly explainable cost on many levels.

We choose a different path: solve the same problems but efficiently, making smart use of the amazing computing power you can hold in your hand.

-- Originally posted 2026-06-09.