Feynman Letter: Do You Want a Data Center That Devours a Nuclear Plant, or Intelligence in a Button Battery? — On TinyML and Green Edge AI
After reading papers on TinyML and the green transformation of edge AI presented at major conferences in May 2026, I feel the LLM arms race — one that wildly burns the planet's energy — has finally taken a calm physical turn.
To explain why squeezing AI into extremely cheap microcontrollers is revolutionary, let's talk about using a cannon to kill a mosquito.
1. The status quo: the power hog that can't live off the wall socket
The AI industry is stuck in a strange arms race: to get AI to understand a single sentence, we deploy tens of thousands of H100 GPUs consuming enough electricity to light up a small town.
- The pain point: For complex reasoning this may be worth it. But if you just want a forest fire monitor to recognize "is there smoke," or a cardiac pacemaker to predict "is there fibrillation," calling a cloud-based hundred-billion-parameter model is using a cannon to swat a mosquito. Network latency can be life-threatening, and continuous power draw makes these micro-devices impossible to sustain. This is "physical unsustainability caused by compute redundancy."
- The physical picture: extreme quantization and pruning: Researchers no longer use expensive floating-point numbers; model weights are compressed drastically — even to 1-bit (binarized networks). They cut a massive neural network down to a bare "spine."
- Hardware-aware neural architecture search (HW-NAS): TinyML's killer move. From the start, the algorithm stares down a microcontroller (MCU) with only 256KB of memory. The model must be accurate AND guarantee its peak memory during inference never overflows that tiny silicon. Call it building a cathedral inside a snail shell.
- Always-on standby without waking: Because power draw is so low (a button battery can run it for years), these micro AIs stay online 24/7. They are like "nerve endings" spread across nature — no internet needed, judgments made locally in an instant.
2. TinyML: a micro brain that can swim in a drop of water
TinyML's underlying logic is fascinating: I don't seek omniscience; I only seek absolute power efficiency in this extremely narrow task.
It achieves dimensional reduction at the milliwatt (mW) or even microwatt (µW) level:
3. A Feynman-style judgment: elegance is reverence for physical constraints
"Ubiquitous intelligence" does not mean data centers everywhere.
It means intelligence that can, like dust, spread cheaply and burden-free 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 chip the size of a grain of rice powered by harvested body heat, technology finally sheds its arrogance and becomes a faint light guarding life.
The takeaway:
When adding AI to IoT or edge devices, don't just think about connecting to a cloud API.
Squeeze your local compute limits.
If your algorithm can't run on a 5-yuan microcontroller, your system will forever lack the "physical vitality" to survive extreme conditions like network or power outages.
*Original hashtags: TinyML, EdgeAI, GreenAI, Sustainability, IoT, Microcontrollers, FeynmanLearning*