How an AI answers,
and what it can do.

A colleague who has read everything, thousands of times.

Imagine a colleague who had read an entire library: textbooks, articles, guidelines, millions of pages. They keep no book on their desk. What they read has become a way of thinking.

This is a language model. Its reading set billions of connections, called its weights. It has no files to look up: everything it knows lives in those settings.

Just like in medical school.

Remember medical school: we reread, recited and rewrote until it stuck. Then one day, the answer came on its own.

A model learns the same way, on another scale. That epinephrine is the first-line treatment for anaphylaxis, it has read thousands of times: in textbooks, guidelines, articles, cases. Each reading strengthened the same connections a little more. Today, the question calls up the answer, as it does for an experienced clinician.

It has also read errors. But an error read once weighs little against what thousands of serious sources repeat. Then experts, including physicians, grade its answers, and it learns to prefer the good ones.

A memory that does not tire.

After the fifteenth admission of the night, our memory slips. Its memory knows neither fatigue nor lack of sleep, and it has read more than any of us will in a lifetime.

What it lacks is what you have: the patient in front of you, their exam, their history. That is why it suggests, and you decide.

Support, even at 4 a.m.

How it writes an answer.

It splits text into pieces, called tokens.

Each token becomes an address in a vast space, where words that go together move closer.

To answer, it writes one word at a time, taking the most likely each time. The most likely is not a roll of the dice: it is what medicine has written most often for a question like yours.

That is also its limit: faced with a rare question, or a practice that has just changed, the most frequent is not always the most accurate.

That is where sources make the difference.

It reads what it is shown.

During a conversation, it reads what it is shown: your question, attached documents, what was said before. That is its context window. It is large: it can work with a lot of information at once.

Its limit is the edge of the window: what it is not shown, it does not see. That is why AtomicMD starts from your dictation: the case, as you described it, is in front of it.

A dated memory, an up-to-date reference.

What it knows by heart stops on the day its training ended, like a textbook printed on a given date.

We do the same: we do not memorize every update, we open the reference. That is what Source does: it reads guidelines and articles at the time of the question, even those published last month.

When it does not know, it fills the gap.

A model does not like silence. Faced with a question whose answer it does not know, it produces what most resembles a good answer: a plausible dose, a study with a credible name, a well-formed reference.

This is called a hallucination. The risk is known, and sources and checks reduce it. That is Source’s rule: every statement carries its source, every citation is checked against its text, and a figure that no source supports is removed.

Thinking before answering.

The most recent models write a draft before their answer. They frame the problem, consider several leads, check a calculation, then answer, like a clinician working through a differential before writing the plan.

This draft makes them clearly better at a difficult problem. It changes nothing about their memory: what they have not read, they still do not know.

Agents: it acts, then it reports back.

So far, the model answers. An agent, on the other hand, acts. It is given a goal and tools (search, read a document, calculate, fill out a form), and it chains the steps: it plans, acts, reads the result, adjusts, and starts again until it can report back.

This is the shift from an assistant that answers to an assistant that does the work.

An agent does the work. You keep the decision.

The engine and the car.

A language model is the engine: knowledge and reasoning. The big companies build better ones every few months.

AtomicMD is the car built around it, for emergency medicine in Québec: precise instructions for each task, local references, forms and RAMQ billing, and checks.

AtomicMD is not tied to any one model: when a better engine comes out, we test it on the same course before installing it.

The car stays built for your shift. And you stay at the wheel.

What the studies measure.

This is no longer a promise: research measures it, and its results point the same way.

AtomicMD is built on this strength: big general models, and the discipline of searching, reading and citing.

Closed book, open book.

A model alone takes a closed-book exam: it answers from memory, with no way to check. It can also be given the book: search for the useful documents at the time of the question, have it read them, and ask it to answer from them, citing what it puts forward.

That is what Source does.