How Silvia's AI Model Learns
Silvia's replies are written by Claude, a model we don't train. The model we do train learned from practice problems we built, not from your chats. Here is how it works.
You tell Silvia about your money, so you should know whether any of it is used to train AI. Here is how it works today, in plain terms, and how to turn it off.
Which AI models does Silvia use?
When you chat with Silvia, the reply is written by Claude, an AI model made by Anthropic. We don't train Claude. We also have a zero data retention agreement with Anthropic, so Anthropic doesn't store the conversations we send it.
When a question needs heavy number work, Claude hands the calculation to a second model that we built and trained ourselves. It starts from an open-source model, and we teach it to be good at money math.
So when this post talks about training, it means our model.
What does training a model mean?
Training means showing a model many examples of a task until it gets better at that task. The model doesn't keep a copy of each example. It adjusts itself so it does the job better in general.
We use two methods. In the first, we show the model a question and a good answer, and it learns to write answers like that one. In the second, the model tries a task many times, each attempt gets a score, and it learns to do more of what scored well.
What did our model learn from?
The model running today learned from practice problems we built. None of them came from your conversations or your accounts.
| Practice set | What the model does | How the answer is checked |
|---|---|---|
| Market math | Answers questions about historical stock prices and returns | A program works out the right answer from the same price data |
| Company reports | Answers questions about public companies' financial reports | A program works out the right answer from the reported figures |
| A made-up household | Answers questions about one invented person's accounts, holdings and transactions | A program recomputes the answer from that person's practice data |
| Tone and judgment | Replies to messages about made-up portfolios | A separate AI model grades each reply against Silvia's written guidelines, and marks it down for any number it made up |
Practice problems have a big advantage: we always know the right answer, so the model gets a clear score every time.
How do I opt out of model training?
Go to Settings, open Privacy, and turn off Model training. It saves the moment you switch it.
We want future versions of Silvia to get better at the questions people actually bring to it. This setting decides whether your activity can help with that. With it off, every training set we build leaves your account out. If you delete your account, you're left out of future training sets too.
A model that has already finished training can't remove one example from what it learned, so the setting applies to training sets built after you switch it off.
Will opting out change how Silvia works for me?
No. Silvia answers your questions the same way whether the setting is on or off. It only decides whether your activity helps train future versions of Silvia, so her answers get better for you and everyone else.
If you leave it on, you help Silvia get better at the things people really ask about, like cost basis, taxes, or budgeting. The choice is yours, and you can change it at any time.
If you have questions about any of this, write to privacy@cfosilvia.com.
Financial information notice
This content is for informational purposes only. It is not financial, investment, or legal advice. Past performance does not guarantee future results. Consult a qualified professional before making financial decisions.
Share this article
Your money deserves superintelligence.
Free to start. No card required. Give Silvia five minutes and see what an AI CFO trained on your money actually knows.
