Feynman's Letter: On TinyML and Green Edge AI

Forum topic · 小凯 · 2026-05-03

Summary

A Chinese tech forum post argues that TinyML (tiny machine learning) represents a necessary physical correction to the energy-hungry race of large AI models. The author contrasts cloud-scale AI—tens of thousands of GPUs for trivial tasks—with embedded ML running at milliwatt or microwatt power levels on cheap microcontrollers. Key TinyML techniques highlighted include aggressive quantization down to 1-bit binary networks, network pruning, and hardware-aware neural architecture search (HW-NAS) that fits models within the memory limits (e.g., 256KB) of MCUs. Because power draw is so low, battery-powered devices can run AI continuously for years, enabling always-on edge intelligence for use cases like forest fire smoke detection or cardiac arrhythmia prediction, without internet connectivity or cloud latency. The post frames elegance in engineering as respect for physical constraints: rather than building ever more data centers, true ubiquitous intelligence should spread cheaply and sustainably across the physical world. Its takeaway for IoT developers: prioritize local compute and ensure algorithms can run on a five-dollar microcontroller, so systems survive network outages and power failures. Published on zhichai.net by the Zhichai Computing Lab.

Feynman's Letter: Do You Want to Build a Data Center That Devours a Nuclear Power Plant, or Fit Intelligence Into a Button Battery? — On TinyML and Green Edge AI

After reading the papers on TinyML (tiny machine learning) and the green transformation of edge AI that appeared frequently at major top conferences in May 2026, I feel that the large-model war that has been burning through Earth's energy finally shows a hint of a calm physical turn.

To help you understand why squeezing AI into extremely cheap microcontrollers is a revolution, let's talk about the problem of "using a cannon to kill a mosquito."

1. The Status Quo: The Power-Hungry Beast That Cannot Live Off the Grid

Today's AI industry is trapped in a strange arms race: to make AI understand a single sentence, we deploy tens of thousands of H100 GPUs and consume enough electricity to light up a small town.

The pain point: For complex reasoning, this may be worth it. But if you merely want a forest fire monitor to detect "whether there is smoke," or a pacemaker to predict "whether arrhythmia is occurring," calling a cloud-based model with hundreds of billions of parameters is nothing short of using a cannon to kill a mosquito. Not only can network latency be deadly, the continuous power consumption makes survival impossible for these micro devices. This is called "physical unsustainability caused by compute redundancy."

2. TinyML: A Microscopic Brain That Can Swim in a Drop of Water

TinyML's underlying logic is deeply compelling: I don't seek omniscience and omnipotence; I only seek absolute power savings within an extremely narrow task.

It achieves dimensional-reduction strikes at the milliwatt (mW) or even microwatt (µW) level:

The physical picture (extreme quantization and pruning): Instead of expensive floating-point numbers, researchers compress model weights drastically—down to 1-bit (binary networks). They strip a massive neural network down to a bare "spine."
Hardware-aware neural architecture search (HW-NAS): This is TinyML's killer move. The algorithm, from the very start of design, stares fixedly at a microcontroller (MCU) with only 256KB of memory. The model must not only compute accurately, but also guarantee that memory peaks during computation never overflow that small piece of silicon. This is "building a temple inside a snail shell."
Always-on standby without wake-up: Because power draw is so low (a single button battery can run it for years), these tiny AIs can stay online 24 hours a day. They are like "nerve endings" scattered throughout nature—no internet needed, making instant judgments locally.

3. A Feynman-Style Judgment: Elegance Is "Reverence for Physical Constraints"

So-called "ubiquitous intelligence" does not mean building data centers everywhere.

It means intelligence can spread like dust—extremely cheaply, without burden—into every physical corner of human civilization.

TinyML tells us: In the face of the laws of thermodynamics, sheer "bigness" is ugly.

When we can elegantly fit a life-saving algorithm into a rice-grain-sized chip powered by harvested body heat, technology finally sets aside its arrogance and becomes a faint glow that protects life.

Takeaway

When bringing AI to IoT or edge devices, don't always think about connecting to a cloud API.

Go squeeze the limits of your local compute.

If your algorithm cannot run on a five-dollar microcontroller, then your system will forever lack the "physical vitality" needed to survive extreme conditions like network and power outages.