Mistral trains enterprise models without moving customer data

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 1:11:33

   Because if you're working with one customer in aerospace or fintech, they might have a need set, but they may not want that training to go to a competitor. I can use a few examples. I think overall, the data segregation is super important. And the way we have solved that is through a portable platform. So our technology is a set of services, a set of training tools, a set of data processing tools that I can take and that I can put on the infrastructure of my customers.

2. 1:12:00

   So suddenly from an IT perspective, and when we talk to the CIOs, they realize that from security perspective, the flow of data doesn't go, there's no data flow coming back to Mistral because everything stays there. Now, the way we then use that technology that has been deployed is that we're going to be working with the team that is doing image scanning and default detection with ISML, for instance. And we're going to be sending forward deployment engineers, scientists, they're all PhDs, they know how to train models, and they spend some time with the subject matter experts that can explain how an image is being detected, how do you detect defaults, etc.

3. 1:12:35

   And based on that, we're going to work out what kind of data needs to be used to train the models that it's going to solve the task in itself. And so we send the technology, typically we send a little bit of scientists because you do need that expertise transfer and that knowledge transfer in between our teams and the vertical experts. And then we make sure that eventually our team no longer needs to be there to retrain the models to get more data access, et cetera. So that combination of data segregation, expertise transfer, knowledge transfer is the one thing that makes us quite unique and allows us to sell the most critical use cases, the most critical processes in industries that actually need to take their data and put it into models for it to work. Yeah, this seems to be once the entire open web, what was available,

4. 1:13:22

   legally, gray market, et cetera. I wouldn't have you comment on that controversy, but we kind of exhausted what's in the open crawl yeah we have and and it's time to actually either make synthetic data or actually use experts do you believe in synthetic data and where does that work and where does it fail we use synthetic data as a way to warm up the models it's a way to actually be quite efficient at the beginning if you have a large model and you want to train a small model you would you will use your large model to process and to produce a lot of synthetic data at the beginning, but eventually you do need to have human signal. So the human signal is something that is always a bit costly to acquire because you need to talk to the experts, they need to give feedback to the machines. And so at the beginning, synthetic data allows you

5. 1:14:09

   to do the compression, to further compress the models. At the end, you do need to go and get data that is produced by humans. So yeah, it's mostly an efficient way of training models to have bigger models that are used as teachers for smaller models, but it's not enough. And so you also need human signal. Arthur, we've seen an incredible explosion. We're sitting here on AO