EdenField: when intelligence becomes less of a hammer, and its seven qualities
Sovereign, frugal, local, portable, reliable, explainable, adaptive: the seven qualities we put forward for EdenField are not marketing claims. They all stem from a single technical choice, a compact and verifiable intelligence rather than a large, opaque statistical model. Here is, for each one, what lets us claim it.
"When you have a hammer, everything looks like a nail."
OpenAI pulled artificial intelligence out of its latest winter, but engineers did not wait for ChatGPT to solve hard problems. Faced with the swelling bubble, we asked ourselves a simple question at Naturalink: why does everyone try to inject, everywhere, something that is sometimes very ill-suited to the problem at hand?
The best answer turned out to be the least original one: "talk about what you do." That is how the EdenField platform was born. This blog is a modest contribution to letting alternative, powerful ideas flourish, to make possible what an all-LLM approach would forbid for lack of connectivity, for budget reasons, for sovereignty, or simply out of ecological awareness.
Embedded Deus Ex Natura
EdenField, short for "Frugal Intelligent Embedded Logic Device", is a field-oriented box: the first embodiment of our effort, faithful to the ambition carried by the "Eden" approach, "Embedded Deus Ex Natura". That motto holds our team's DNA: the spark of human ingenuity in the service of varied problems, as respectful as possible of Nature, our first source of inspiration.
EdenField: Eden embodied
EdenField contains neither a large language model nor a neural network with billions of parameters, but a small learned circuit made of threshold gates and weights that take only three discrete values. This architectural choice is precisely what makes the seven characteristic qualities possible. In other words, everything in EdenField fits inside the box: perceive, decide, act, or abstain.
flowchart LR
subgraph BOX["Inside the box, no server"]
C["Sensor: camera, microphone, measurement"] --> P["Embedded perception"]
P --> D["Decision: threshold circuit"]
D -->|known situation| A["Local action"]
D -->|doubt| S["Abstention: I do not know"]
end
Sovereign, frugal, local, portable, reliable, explainable, adaptive: seven qualities that hold to a single choice, that of a learned "neurocircuit", compact and verifiable, rather than a large statistical model.
Sovereign. No lock-in to big tech platforms. The technology is designed and mastered in France, from the circuit architecture to the embedded firmware. The AI runs on your terms, on your hardware.
Frugal. It requires little computing power, memory or energy. Because it uses only integers and three-valued weights, it does without floating-point computation, GPUs and abundant memory, and runs on a simple microcontroller.
Local. It can make decisions on-site, without the cloud. No data leaves the equipment, so no risk of data leaks; no network latency, so an immediate reaction; and it keeps working even without a connection.
Portable. The same model, described in a single format, runs on microcontrollers, computers, browsers or servers, from a prototype on a bench to the series installed in the field.
Reliable. It knows how to say "I don't know." A generative model always answers, even if it has to make things up; EdenField instead chooses from a finite set of allowed decisions and abstains when the situation resembles nothing it knows. No risk of hallucination.
Explainable. It is not a "black box." A threshold-gate circuit is a finite structure that can be opened and inspected: its behaviour can be controlled at a granular level and its reasoning is auditable, where a model with billions of parameters remains opaque.
Adaptive. It evolves and improves its behaviour based on new information: the model trains on field data and can be retrained and updated as new situations arise, across a whole fleet of devices.
Promises tested in the field
None of this stayed theoretical. Our demonstrator God Save The Bee, installed in front of real hives to protect colonies from the Asian hornet, sees, recognises, decides and acts entirely inside its box, without any server. These results rest on dozens of hours of field recordings and tens of thousands of annotated data frames. It is this same foundation, now open to other agricultural and environmental equipment, that we invite you to explore with us.
You build equipment, and one of these promises meets a concrete need for one of your projects? Let's talk, at edenfield.natura-link.fr.
Tell us about your equipment
Describe your machine, the situation it must recognise and the decision it should make: we will tell you how to reach a first prototype.
Describe my need