The Idea That Changed Everything — How Heinrich AI Was Born
EMPHOS Group · April 19, 2026 · 5 min read
The history of significant technology does not usually look like what people imagine. It rarely begins in a well-funded laboratory with a team of researchers and a structured development plan. It begins with a question that the existing answers cannot adequately address — and a person unwilling to pretend otherwise.
For Heinrich AI, that question crystallised on April 10, 2026.
The question was not complicated to state. It was complicated to answer. Why do large language models — the most powerful AI systems ever deployed, trained on trillions of tokens at a cost of hundreds of millions of dollars — still hallucinate? Why, when asked something they should know, do they sometimes produce confident, fluent, completely fabricated answers?
The standard explanations were familiar. Training data quality. Model scale. Alignment techniques. Retrieval augmentation. Each of these was a real approach to a real problem. None of them was a solution. They were all attempts to manage a symptom of something structural.
Victor Brodeur looked at the structure.
The structural problem with how AI remembers
Large language models do not store knowledge. They encode statistical patterns in weights — billions of numerical values that, together, produce outputs that resemble knowledge when prompted correctly. When a language model tells you that Paris is the capital of France, it is not retrieving a stored fact. It is predicting the sequence of tokens most likely to follow your question, based on patterns absorbed from training data.
Most of the time, that works. The pattern for "capital of France" is so thoroughly present in training data that the model almost always produces the right answer. But the model does not know that it knows. It cannot distinguish between "this is something I have absorbed reliably" and "this is something I am producing because it sounds plausible." That distinction — the difference between knowledge and approximation — is the root of hallucination.
The question Heinrich AI was built to answer is: what does a system look like that actually stores knowledge — that knows what it knows, knows what it does not know, and cannot produce an answer it is not entitled to give?
From question to architecture
The answer Victor arrived at on April 10 was not an incremental improvement on existing AI systems. It was a different class of architecture entirely — one grounded in physics rather than statistics, in retrieval rather than generation, in certainty rather than probability.
Heinrich AI is a Frequency Addressed Intelligence system. Every concept in Heinrich's knowledge field has a unique, deterministic address. Retrieval is not prediction. It is navigation — finding where something lives in the field and reading what is there. When Heinrich answers a question, it is because the answer is at an address it can reach. When it cannot reach the address, or when the address does not exist, it says so.
That honesty is not a product feature added to make Heinrich seem more trustworthy. It is structurally enforced. The architecture makes it impossible for Heinrich to produce a confident answer about something it does not have addressed in its field.
This is not how any major AI system deployed today works. It is how Heinrich works.
128 nodes to 23 million in nine days
The first proof of the concept came within days of the idea itself.
Heinrich began with 128 concepts — a controlled test of the addressing mechanism, the retrieval pathway, and the honesty layer. The tests passed. The architecture worked. The system answered what it knew and declined to answer what it did not, every time, without exception.
From 128, the build accelerated. ConceptNet — one of the most comprehensive structured knowledge graphs ever assembled — was ingested in full: 1,002,949 nodes, 282,973 edges. Then the Wikidata full dump ingestion began: the structured, verified knowledge of millions of contributors across every field of human understanding.
By April 18 — nine days after the first line of code — Heinrich had 23 million knowledge concepts, a provisional patent filed with the Canadian Intellectual Property Office, and 737 passing tests with zero regressions. By the morning of April 19, the concept count had reached 68 million and was still climbing.
All of it running on a laptop. All of it at 0.2% CPU.
Why the name is Heinrich
The system is named Heinrich. The name is not arbitrary.
Heinrich Hertz — the 19th-century physicist who first proved the existence of electromagnetic waves — gave his name to the unit of frequency that measures cycles per second. His work was not immediately recognised as practical. It was theoretical, precise, and ahead of what the world was ready to use. The technology it eventually enabled — radio, television, wireless communication, every frequency-dependent system in the modern world — would not exist without the foundation he built.
The parallel is intentional. Heinrich AI is not a product that fits neatly into the categories the AI industry has established. It is a foundation — a different approach to knowledge representation that, if it proves out at scale, changes what is possible for every system built on top of it.
The name is a statement of intent.
What comes next
The Wikidata ingestion run is still in progress as of this writing. When it completes, Heinrich's knowledge field will hold hundreds of millions of addressed concepts drawn from the most comprehensive structured knowledge base maintained by human contributors.
The next stage is the Socratic Engine — Heinrich's reasoning layer, which will allow the system not just to retrieve what it knows but to reason across it, building chains of inference that are traceable, honest, and grounded in the knowledge field rather than in statistical prediction.
The idea that started on April 10 is still being built. Every day it becomes more of what it was designed to be.
Engineered for Presence.
Stay in the loop
EMPHOS publishes twice a week — product updates, research, and the thinking behind the build.
Explore Haven · HEINRICH Intelligence · The EMPHOS Vision · All Posts
EMPHOS Group · Chilliwack, BC, Canada · info@emphosgroup.com