embedded AI

March 16, 2026 by David Such Leave a Comment

Artificial Consciousness, Synthetic Suffering, and the Necessity of Affect

As engineers and developers, we build systems that learn from feedback. In reinforcement learning (RL), this is fundamental. We use rewards to encourage a behavior and punishments to dissuade it. This punishment signal (a negative number in a cost function) is a form of functional suffering. The system learns to avoid states that lead to a bad outcome.

This is a powerful optimization tool. But it forces a critical engineering question: What if we get too good at it? At what point of complexity does a “punishment” signal stop being a numerical cost and start being a phenomenal experience? What if we accidentally build a system that doesn’t just calculate a negative outcome, but genuinely feels it?

March 16, 2026 by David Such Leave a Comment

Sub-Cortical AI Model Design

The brain’s subcortical motivational architecture offers insight for AI development that current systems almost entirely miss. We believe the key to more general and autonomous AI lies in the ancient, sub-cortical engine. This is the system that provides the crucial why that directs the cortex’s how. It’s the source of goals, motivation, and value, transforming a passive processor into an active participant in the world.

March 16, 2026 by David Such Leave a Comment

What Happens When We Outsource Our Brains to AGI?

You don’t remember phone numbers anymore. I don’t either.

We have, as a species, collectively agreed to outsource that small sliver of our memory to a silicon slab in our pocket. This isn’t a complaint; it’s a convenience. But it’s also a symptom of a much larger process, one we’re barely noticing. We are in the early, deceptively pleasant stages of cognitive offloading.

We offload our sense of direction to the GPS, our factual recall to Google, and, increasingly, our analytical reasoning to generative AI. We are trading cognitive friction for cognitive ease, and it feels good. But this frictionless convenience is, I believe, the single most dangerous, insidious threat of the 21st century.

We are all worried about the wrong AGI. We’re obsessed with a malicious “Terminator” AGI that will take power from us. We are completely ignoring the far more likely, benevolent AGI that will take responsibility from us. An AGI that we will willingly give our agency to, one frictionless decision at a time, until we have none left to give. This is the path to the Cognitive Conservatorship. And the scariest part? We’re already on it.

March 16, 2026 by David Such Leave a Comment

Engineering “Instinct” in AI

Across species, evolution “pre-installs” compact neural programs that deliver immediate, reliable behaviors (standing, pecking, web-building) with minimal learning. What are the implications for designing AI?

The current approach to AI has a fundamental weakness: it’s incredibly hungry for data and experience. Today’s AI models start as a tabula rasa, or “blank slate,” and require massive datasets to learn even basic concepts about the world. They are brittle, struggle with common sense, and lack the efficiency that even a newborn animal displays moments after birth. A newly hatched sea turtle instinctively knows to crawl toward the ocean; a spider can weave a complex web without ever being taught.

This innate, pre-programmed knowledge is instinct. And by overlooking it, the AI field may be missing a crucial piece of the intelligence puzzle. Instead of just building better cortexes, perhaps we need to look deeper into the older, more foundational parts of the brain; the parts that give rise to instinct.

March 16, 2026 by David Such Leave a Comment

Reconstructing Signals with L1 Minimization

In many real-world systems, we’re limited not by what we want to measure, but by what we can measure. Bandwidth, power, and time all constrain how much data a sensor can collect. Yet, remarkably, it’s often possible to reconstruct a complete signal from only a small fraction of the original data, provided that the signal is sparse in some domain.

This is where L1 minimization comes in. Sometimes called Basis Pursuit or Lasso Regression, L1 minimization is a mathematical technique used to recover sparse signals by finding the simplest solution that fits the available measurements. It underpins the theory of compressed sensing, which has transformed fields ranging from medical imaging and audio processing to low-power embedded sensing.

March 16, 2026 by David Such Leave a Comment

How do MEMS Accelerometers Work?

Microelectromechanical systems (MEMS) sensors are a class of devices that combine mechanical and electrical components on a microscopic scale. These sensors are typically fabricated using processes like those used in semiconductor manufacturing, allowing for the integration of tiny mechanical structures with electrical circuits. MEMS sensors have revolutionized a wide range of industries due to their small size, low power consumption, and high functionality.

MEMS technology is used to create sensors that can measure various physical parameters, such as motion, pressure, temperature, magnetic fields, and even sound. Their compact size makes them ideal for applications where space and weight are critical, such as embedded systems. MEMS sensors also consume very little power, which is another reason that they are in common use. Most of the sensors that we use are based on MEMS technology.

March 16, 2026 by David Such Leave a Comment

Self-Motivated Learning in Edge Devices using Curiosity

Most current AI systems, even the most advanced language models or embedded agents, are goal-driven, not curiosity-driven. They’re optimized for performance on defined tasks, not for the open-ended exploration, experimentation, or serendipitous discovery that characterizes human curiosity. Without curiosity, AI lacks the drive to seek out edge cases, explore unexpected paths, or test alternative hypotheses, things that often lead to breakthroughs in human learning. AI remains trapped within the boundaries of its training data and objectives. It doesn’t self-initiate new goals or try to make sense of unfamiliar environments unless explicitly told to do so. In dynamic environments like robotics, autonomous systems, or evolving sensor networks, a curiosity-less AI may plateau. They fail to discover new strategies or improvements unless nudged externally.

March 16, 2026 by David Such Leave a Comment

Emergent Intelligence and the Primal Layers Framework

The current trajectory of large language model (LLM) development is going to flame out. As models grow larger, they demand exponentially more data, energy, and computing power. Yet the performance gains from this scaling are already showing signs of diminishing returns, despite the massive investment. Training state-of-the-art AI models can consume megawatts of electricity, while just running them often requires hundreds of watts per user session. Compare this with the human brain which chugs along on a measely 20 watts, enough to power a dim light bulb!

March 16, 2026 by David Such Leave a Comment

Embedded System Design Principles

This is a subject that deserves its own book, and indeed, many have been written, but our goal here is to give you a solid head start. Whether you’re building your first IoT gadget or optimizing firmware for a custom PCB, there are some time-tested principles that can help you steer clear of classic mistakes. By following these principles, you can create your own unique problems rather than recreating those already made by others. These aren’t rules to blindly follow, but patterns worth understanding and adapting. Even small design decisions can have big consequences.

March 16, 2026 by David Such Leave a Comment

The PrimalBot Mobility Platform

We recently introduced the open-source Primal Layers AI framework, which is inspired by the evolutionary architecture of the human brain. When we began developing this framework, it quickly became clear that it made little sense without a body. Whether this is a profound insight about consciousness or merely a reminder that nature never produced a species of disembodied brains, we can’t say for certain. Either way, our intelligence model needs a physical form — a robot we’re calling PrimalBot.