Claude charges you twice for every PDF page, once for the text and once for the image. Converting to Markdown drops half the bill, as long as the document’s value is not in its figures.
A 50-page PDF can cost you 75,000 to 150,000 tokens before Claude has read a word of it. On a 200,000-token context window, that is most of your working space gone on one document. The reason is not the text, it is due to the way that Claude ingests a PDF.
In our original article on the Primal Layers framework, we described a multi-layered AI architecture inspired by the evolutionary structure of the human brain. Layer 1 (the brainstem) handles homeostasis and reflexes. Layer 2 (the limbic system) handles motivation, memory, and emotional valence. Layer 3 (the cerebellum) handles motor learning and predictive control. And the Cognitive Layer (the neocortex) handles abstract reasoning and planning.
What we didn’t address is how these layers manage knowledge over time. How does PrimalBot learn something new on Tuesday without forgetting what it learned on Monday? The answer, like so many in this project, comes from biology. And it turns out that one of the most important things a brain does is sleep.
By 2026, language models have moved off the cloud and onto the device in your pocket. What was a research demonstration two years ago is now a routine engineering capability, and the centre of gravity for artificial intelligence has begun to migrate from distant data centres to local silicon.
The episode traces the four engineering moves that made this possible. Quantization, which shrinks a model by storing its parameters with less precision. Optimized key-value caches, which let a model hold a long conversation without exhausting memory. Neural Processing Units, the dedicated AI accelerators now standard in flagship phones. And specialized frameworks such as LiteRT-LM and llama.cpp, which finally make all three usable from a single application.
The consequences reach further than performance figures. Privacy becomes the default rather than a feature, because data never leaves the device. The cost structure of AI applications changes, because there are no per-query cloud fees. And the link between training capital and deployment capability begins to decouple, opening the door for small teams to ship genuine intelligence on hardware they already control.
Most AI research is about making the cortex bigger. This one is about making the spinal cord cheaper.
A team at the University of Michigan has built a tiny computing device that controls a balancing propeller using about seven millionths of a watt. For comparison, the LED bulb in your kitchen burns through about ten watts. The Michigan device runs the control task on roughly a millionth of that power.
That is not a typo. It is the finding of a paper published in ACS Nano in March 2026, and it matters because power is the wall that edge AI keeps running into (or falling off?).
Most of the interesting AI you read about lives in a data centre. It has a wall socket, a cooling system, and an electricity bill measured in millions. Edge AI is what happens when you try to put that intelligence into a hearing aid, a pacemaker, a drone, a soil sensor, or a pair of smart glasses. You are running off a small battery, or whatever energy you can scavenge from sunlight or vibration.
In that world, every microwatt counts, and there is one component that has been quietly eating the budget for decades: the analog-to-digital converter (ADC). Sensors produce continuous signals, but computers think in ones and zeros. Something has to translate between the two, and that translator is the ADC. It is usually the single biggest line item in a battery-powered device’s power budget.
The Michigan team’s trick is to skip the translator entirely.
As foundation models move from the cloud into physical robots, a fundamental question emerges: who is accountable when an AI-controlled machine makes a decision that causes harm?
In this episode, we examine the growing collision between embodied AI, functional safety, and emerging regulation. We explore how new frameworks such as the EU AI Act and the Machinery Regulation are reshaping expectations for developers, manufacturers, and deployers of intelligent robots. From humanoid robots and autonomous mobile manipulators to AI-enabled industrial machinery, the challenge is no longer simply making robots smarter. It is making them governable.
We investigate a proposed architectural solution that is gaining traction across industry and academia: the hardware-isolated safety supervisor. By separating non-deterministic AI reasoning from deterministic safety-critical control systems, this approach aims to create clear lines of accountability while preserving the benefits of onboard intelligence.
Along the way, we examine NVIDIA’s Cosmos Reason 2 model, the EmbodiedGovBench governance framework, emerging standards efforts, and the practical realities of deploying foundation models on embedded platforms. We also ask whether traditional functional safety concepts such as SIL and ASIL can adequately address the unique challenges posed by robots whose actions are selected by large vision-language models.
The broader question is one that every roboticist, embedded engineer, and AI practitioner will soon face: when intelligence becomes local, autonomous, and physically embodied, what mechanisms ensure that accountability remains local too?