How Dimitri works
What happens between pressing Send and reading the answer, and how we decide which AI answers it.
1. A safety check
Every message goes through an automated safety check before it's answered, and answers are checked too. Requests for clearly harmful content are declined.
2. Working out what you're asking
A classifier reads your message and labels the kind of task: everyday advice, writing, an explanation, math, logic, code, data, and so on. It also notes whether the question needs current information from the web. Short messages like "thanks" skip this step.
3. Choosing the model
Each kind of task is assigned the model that did best on it in our tests, weighing accuracy, how clearly it explains, speed and cost. The choice is made for every message, so a single conversation can be answered by different models without you noticing. Whichever model answers is given the conversation so far.
4. Looking things up when needed
If your question depends on current information, like today's news, a price or a schedule, Dimitri searches the web before answering and links the sources below the answer.
5. A backup if something goes wrong
AI services have outages and slow moments. If a model fails or doesn't start answering in time, a model from a different company takes over automatically, so one company's outage doesn't stop your answer.
6. Try again
Not happy with an answer? Try again asks a stronger, more capable model to answer instead, and replaces the first answer. It also tells us which kinds of questions need a better first choice.
7. Ads
Every few answers, an ad may appear below the answer once it has finished. Ads are labeled, never interrupt typing, and are never chosen based on your conversation.
How we test models
Before a model answers anyone, it goes through the same tests as the models already in use:
- Known-answer questions in math, logic and data analysis, checked automatically.
- Blind side-by-side comparisons: two answers to the same question, with the model names hidden, judged for correctness and completeness. An AI judge scores every pair, and people review writing style, because AI judges tend to favor long, list-heavy answers.
- Long conversations, to make sure a model keeps track of what you said earlier.
- Sensitive topics, to make sure a model answers factually instead of refusing or repeating an official line.
- Speed and cost, because a free service has to be efficient to stay free.
We've graded thousands of answers this way. When a new model comes out, it goes through the same tests, and if it does better on a kind of task, it takes over that task.
Why we don't show model names
The lineup changes as better models come out, and the names mean little to most people. What matters is whether the answer is right, clear and quick. If you want to know which companies process your messages, they're listed in the privacy policy.