Reefwing Software

You are here: Home / Blog

Blog

by Leave a Comment

Your Brain Has a Real Estate Problem, and Dreams Might Be the Solution

What a strange theory about REM sleep tells us about building AI that doesn’t quietly fall apart

Put a blindfold on a sighted adult and stick them in an fMRI scanner. Get them to feel some textured surfaces with their fingers. Then watch what happens in their visual cortex, the part of the brain that is supposed to be doing absolutely nothing in the dark. It lights up. Not after weeks, not after days, but after about forty-five minutes.

That finding, which has been replicated in various forms since the early 2000s, is one of the more unsettling results in modern neuroscience. The visual cortex is not a quiet, dormant region waiting patiently for the lights to come back on. The moment you stop using it for vision, the neighbours start moving in. Touch first, then hearing. Within an hour, real estate that was supposed to be reserved for processing photons is being repurposed for something else entirely.

This raises an obvious and slightly alarming question. If your visual cortex starts getting taken over within an hour of going dark, what happens every single night when you close your eyes for eight hours?

A neuroscientist named David Eagleman thinks he has the answer, and it is one of the more interesting hypotheses about why we dream that has come along in a while. He calls it the defensive activation theory, and once you understand it, you start seeing the same pattern everywhere, including in some surprisingly broken corners of how we currently build AI.

Read More…

by Leave a Comment

A chip that controls a balancing propeller on seven microwatts

Every battery-powered device you own has a quiet energy hog in it that nobody talks about. It is not the processor, it is not the radio, and it is not the screen. It is the analog-to-digital converter, the small piece of circuitry that translates the messy real world into the clean ones and zeros a computer can think about. For thirty years it has been the line item that decides how long your hearing aid, your pacemaker, or your soil sensor lasts on a battery.

In March 2026, a team at the University of Michigan published a result that quietly removes that converter from the picture for a specific class of problems. Their bismuth selenide memristor runs a closed-loop control task at about seven microwatts, roughly a millionth of what a household LED bulb pulls. The chip does not run code in any conventional sense. The physics does the arithmetic, and the answer drives the motor directly.

In this episode, we walk through what the device actually is, why removing the converter changes the energy budget by orders of magnitude, and which products land first when microwatt-class intelligence becomes buildable. We talk about hearing aids, implants, environmental sensors, and the small drones that have been waiting for this kind of result for a decade. We also talk about what this chip cannot do, because the press releases tend to skip that part. It will not run a language model. It will not recognise your face. It will run the reflexes underneath all of that, and the case for why those reflexes matter more than the cortex gets credit for is the through-line of the episode. #embeddedAI #podcast

Listen to the Podcast…

by Leave a Comment

Spot the Robot Can Reason, but it can’t hold a can of soda

Boston Dynamics and Google DeepMind announced last month that Spot is now running Gemini Robotics-ER 1.6, a high-level embodied reasoning model. The headline is that Spot has been taught to reason. The commercial pitch is industrial inspection: wandering around a facility, reading gauges, spotting spills, deciding what to do when something looks wrong. So far, so good.

Buried in the demo video is a small failure that says more about the architecture than the press release does. Asked to recycle cans in the living room, Spot picks one up sideways. If there is any liquid left in the can, it ends up on the carpet. The problem is using a language model to solve something handled by other more primitive layers in the organic world.

Read the complete article…

by Leave a Comment

Why 95% of AI Deployments Fail

MIT’s August 2025 study of 300 enterprise generative AI deployments found that 95% produced no measurable P&L impact. Gartner forecasts that more than 40% of agentic AI projects will be cancelled by 2027. McKinsey’s State of AI 2025 identifies workflow redesign as the single strongest correlate with EBIT impact, yet only 21% of organisations have redesigned any workflows. The data converges on a structural conclusion: enterprise AI is failing because the operational substrate is inadequate, not because the models are. This episode examines the process-readiness gap, the misallocation pattern that concentrates investment in low-ROI front-office applications, and what the 5% of high performers do differently. It is an architectural argument, not a change-management one: AI is a linear amplifier acting on a pre-existing process, and the sign of the output depends on the sign of the input.

Listen to Podcast…

by Leave a Comment

Why ARC’s $2M Bet Against Priors Is the Wrong Bet for Embedded AI

ARC says intelligence emerges from learning when priors are minimised. Vertebrates disagree and have spent 500 million years proving the opposite.

The ARC Prize is a public competition with $2 million on the table, run by François Chollet (creator of Keras) and Mike Knoop (co-founder of Zapier). The premise is that current AI benchmarks measure what a model already knows, not how well it learns. Their fix is ARC-AGI, a series of reasoning tasks that are easy for humans and, so far, brutal for machines. Frontier models score below 4% on ARC-AGI-2.

The newest version, ARC-AGI-3, swaps static puzzles for interactive game environments where the agent has to figure out the rules by playing, with no instructions and no pre-loaded domain knowledge. Humans score 100%. AI scores under 1%. The gap demonstrates the point. Behind the gap is a philosophy about what intelligence is and where it comes from, and that philosophy is where the Primal Layers framework and ARC differ.

Read Full Article…

by Leave a Comment

Why Humans and Robots must Dream

Put a blindfold on a sighted adult and the visual cortex starts being colonised by touch and hearing within forty-five minutes. Not weeks. Not days. Forty-five minutes. This is not a quirk of extreme cases. It is how the cortex works all the time. Every region of the brain is in continuous low-grade negotiation with its neighbours over territory, and the currency of that negotiation is activity. Stop using a subsystem and the neighbours move in, fast. This is the empirical foundation of a hypothesis from neuroscientist David Eagleman called the defensive activation theory: that REM sleep exists specifically to keep the visual cortex active during the eight hours each night when external input is unavailable, defending its territory against takeover by senses that never go offline.

The theory itself is plausible but not yet directly proven. What is proven, and what matters more for engineers, is the underlying principle. A complex system with reconfigurable resources will silently lose capability in any subsystem that is not regularly exercised, even when nothing is actively trying to take that capability away. This is not catastrophic forgetting in the usual machine learning sense, where new training overwrites old parameters. This is something subtler and arguably more dangerous: passive territorial loss in any system that supports continuous adaptation. It shows up wherever capabilities are not being exercised in long-running adaptive AI: rarely-routed experts in mixture-of-experts models, underused sensor pipelines in multi-modal robotics, capabilities that drift out of online reinforcement learning agents over months of deployment. Most current architectures treat their structure as fixed by design. Biology treats its structure as continuously contested.

This episode looks at what defensive activation reveals about a missing primitive in modern AI architecture. Current systems have two fundamental modes, training and inference. Brains have at least three, and the third one, the maintenance mode that operates during REM sleep, has no clean equivalent in the systems we build. We examine what this mode is doing structurally, why generative replay in continual learning is mechanistically closer to dreaming than the field usually acknowledges, and what a telemetry-driven maintenance subsystem might look like for embedded and edge AI. The closing argument is straightforward: if biology has been running this experiment for a few hundred million years and converged on internally-driven activation as the way to maintain a plastic computational substrate, the absence of an equivalent mechanism in our architectures is not a neutral design choice. It is a gap.

Listen to the podcast…

by Leave a Comment

Why Asimov’s Three Laws Shouldn’t Be the Blueprint for AI Principles

Every time a news story breaks about some new AI mishap, a chatbot lying to a user, a self-driving car making a dodgy decision, a recommendation engine nudging teenagers toward inappropriate content, someone in the comments inevitably writes, “We just need Asimov’s Three Laws.”

The trouble is that Isaac Asimov wrote the Three Laws as a plot device, not as an engineering specification. Almost every story he wrote about them was about how they fail. It is a bit like reading Jurassic Park and concluding that you now have a solid operating manual for cloning dinosaurs.

If we are serious about building safe AI, and particularly if we are building the sort of layered, bio-inspired systems that drive physical robots, we need to start from a different foundation. This article explains why, and proposes a replacement set of principles drawn from the Primal Layers framework.

Read the complete article…

by Leave a Comment

Sovereign AI and the End of the Borderless Cloud

The borderless cloud era is ending. In the second week of January 2026, four government decisions announced in rapid succession made that shift undeniable: the UK activated its £500 million Sovereign AI Unit, France committed €109 billion, the UAE consolidated a $40 billion data centre portfolio, and the Trump administration revised chip export rules to China. In this episode, we examine why AI infrastructure is now being treated as a strategic national utility on par with energy and water, and what that means for engineers and boards making architectural decisions today.

We map the global sovereign AI landscape, roughly 130 national initiatives across more than 50 countries, and separate political rhetoric from engineering reality. We examine the distinction between regulatory sovereignty (the legal authority to govern AI) and compute sovereignty (the physical capacity to run it), and explain why most nations have the first without the second. We cover China’s full-stack response through Huawei’s Ascend and CloudMatrix programme, a deliberate trade-off of efficiency for independence that is becoming a template other regions may follow. We draw on the Clipper chip precedent from the 1990s to show why embedded enforcement mechanisms in silicon create durable market incentives that are difficult to reverse.

Listen to the Podcast…

by Leave a Comment

Pi and the Mirage of Patternicity

In April 2025, a claim began circulating online: pi is gradually increasing around the 7,237th decimal place. A math enthusiast in Cincinnati named April Simons had apparently flagged the anomaly. Prof F.O. Olsday, head of the Number Theory Group at Princeton, was quoted confirming it. Cosmologists were linking it to the accelerating expansion of the universe. The same algorithm, the same hardware, different results. A 4 becoming a 5. Persistent. Inexplicable.

Except that “F.O. Olsday” is a phonetic rearrangement of “Fool’s Day.” And April Simons was posting from Cincinnati on the first of April.

Pi has not changed. It cannot change. It is a fixed ratio determined by Euclidean geometry, and every one of its digits is as immutable as the definition that produces them. The 7,237th digit was a 4 before 2016, it was a 4 after 2016, and it will remain a 4 until the heat death of the universe and beyond.

But here is what matters: the joke worked. It worked on humans, and it would work on machines.

This episode examines why both biological and artificial neural networks are structurally vulnerable to detecting patterns in structurally empty data, a phenomenon with a clinical name: apophenia. We trace the evolutionary logic behind false positive pattern detection, from Skinner’s superstitious pigeons to the fusiform face area that fires on toast. We then show how the same asymmetry, optimising for recall at the expense of precision, is recapitulated in trained neural networks through simplicity bias, the documented tendency of gradient-descent-trained models to latch onto whichever statistical regularity is easiest to extract, regardless of whether it reflects causal structure.

Listen to the Podcast…

by Leave a Comment

Claude Code: Creating a C++ Linter for Embedded Development

I know! I’m late to the Claude Code party but now I’m here, I’m all in. If you write C++ for microcontrollers, or edge inference, you already know that the rules are different from desktop software. No heap allocation after startup. No exceptions. No recursion on a 4 KB stack. And these constraints are not optional if you want your firmware to survive.

The problem is that general-purpose linters do not enforce the rules you need. Clang-tidy is powerful, but configuring it to catch you just used int instead of int32_t, requires writing custom checks in C++ against the AST. That is a significant investment for what should be a simple rule. I wanted something I could tweak for each project.

Read More…