Llama2-70B-Chat is now available on MosaicML Inference

MosaicML is now part of Databricks

Introducing MPT-30B, the latest addition to the MosaicML Foundation Series of Models.

Personal AI: Customer Spotlight

Interview with Sharon Zhang, the CTO & co-founder of Personal AI, on why she was inspired to develop personalized AI models for individual users.

What was your inspiration for founding this company?

We started Personal AI in February of 2020. The inspiration for us was the concept that humans forget about 80% of everything we see on a daily basis. There’s a lot of lost knowledge.

My grandmother is 90 years old and is a huge inspiration for me. She’s lived through WWII, the Chinese cultural revolution…we wanted to be able to capture our family history in a way that can be perpetuated through generations. I’ve always had the fear that once my grandmother passes away, all of this knowledge is lost.

As well, I bought my father an Alexa device back in 2019. He thought it was very cool, but wanted to know if he could use it to talk to me. I’m notoriously bad at calling and keeping my parents up-to-date. Since I was working in the NLP industry, I was brainstorming with him about what if we captured one person’s memory in a “memory stack” and then have the ability to retrieve the information on demand—either by you, or someone close to you. Essentially, a representation of me that can be ported so that my dad doesn’t have to wait for me to call.


Now that you’re in development, can you tell me more about your product?

There is an independent AI model per person. It learns about you: your favorite foods, how you take your coffee, who you hang out with. It’s an augmentation or extension of your memory, essentially. This is different than a larger generative AI model that would never know details of your life.

Our first use case is really about training on the data that you have. For example, documents, messages, texts, tweets…all the data in the “universe of you.” 

This data is used to create an individual model for each user. They are extremely small models compared to large language models, and fully controlled by the user. You can communicate with your personal AI, or other people can communicate with your AI, so a sense of connection comes into play as well. And you don’t just talk to your AI. It’s more like your AI is in the loop and in the background. Any conversation you have is automatically training your model (to the extent you want it to, of course).


How do you address concerns about security and privacy of personal data?

First off, we don’t aggregate data or cross-train models since we don’t have a large language model. In terms of data segregation, we go very far. Every single user has their own index of data that they fully control. They chose what data is included (or excluded), who has access, and to what data. If a user decides to leave the platform, their entire history is wiped.

In the future, we want this to be something that lives on the edge (on your personal device). If you never want your data to go to the cloud and train your model, which is how it works today, the eventual goal would be to have the training and inference done on the user’s device.


What is your model development process today? How do you train the personal AIs?

It starts almost immediately for the end user. As you type your first message on the platform, that data is incorporated into your model. We don’t actually train a new model until we get a certain level of data. For people that have a large amount of data, their generative model will be more robust. Essentially, there’s a hierarchy of training. We are literally training multiple models for each user, which makes model training, inference, and management a huge infrastructure challenge for us to solve. This horizontal scaling is something we’ve had to work through these past few years.


How did you discover MosaicML?

My co-founder Suman was familiar with the work of your CEO, Naveen Rao. At that time, we were running our own training containers in our own clusters, and it was taking about 10 hours to train each personal model—quite a long time for an end user to wait. When we talked with MosaicML and heard about the acceleration you guys are doing, we realized we could benefit from both a cost perspective (by training models faster) as well as giving a much more optimal user experience. We implemented your platform late last year, and our 10-hour model training time has gone down to 10 minutes—a massive impact on our user experience!


Have there been any particular challenges while building your models that MosaicML has helped you to address?

You really helped us with the issue of training many user models at the same time in a very short window. Additionally, we have been using your inference service, which means we’re able to host some of the LLMs ourselves, and really offer a private experience for our users.

Also, as a small startup, we’re competing with large companies for GPU availability. That’s been a major issue, and since you have tons of capacity, being able to reserve an A-100 has been very helpful in making us more efficient and more competitive.

I wish we’d had MosaicML two years earlier, so that we didn’t have to build out our own training clusters. Your MPT series is another good example; we probably wouldn’t have had the time to train a 7B model and experiment with it, but we can use what you’ve been doing!


What’s next for PersonalAI?

We are looking forward to taking advantage of your LLM models and seeing how they work with our personal models. I also got very excited when I saw your new edge training capabilities.

For PersonalAI, we will be rolling out our mobile offerings with keyboard and text integration very soon. We have about 60K users now and are planning to go much bigger. We hope to get to a world where we have 8billion humans and 8 billion personal AIs. A model for each individual!


Ready to get started? Message Sharon at ai.personal.ai to start training your personal AI!

What’s a Rich Text element?

The rich text element allows you to create and format headings, paragraphs, blockquotes, images, and video all in one place instead of having to add and format them individually. Just double-click and easily create content.

Static and dynamic content editing

A rich text element can be used with static or dynamic content. For static content, just drop it into any page and begin editing. For dynamic content, add a rich text field to any collection and then connect a rich text element to that field in the settings panel. Voila!

How to customize formatting for each rich text

Headings, paragraphs, blockquotes, figures, images, and figure captions can all be styled after a class is added to the rich text element using the "When inside of" nested selector system.