OthersideAI: Customer Spotlight
Tell me more about OthersideAI. How did the company get started, and what are you working on now?
OthersideAI started as an AI lab. We build applications that enable people to get more out of AI models, allowing them to become useful for day-to-day work. We started with writing; HyperWrite was our original product. It allowed people to write better and faster across the Internet, whether in Gmail for business emails, Google Docs, Microsoft Word for documents, or any other site. It also has smart auto-complete tools to help you come up with ideas and write things from scratch. However, our main goal has always been to build an AI personal assistant that anybody can use: a tool that has the power of an executive-level personal assistant and can be scaled to everyone at a fraction of the cost.
Our AI today can actually operate a web browser on a user's behalf. You just describe what you want it to do and it will literally use your computer to do that thing. I could say “Order me a pizza” and it will go to a pizza store's website and order that pizza for me. I can say “Do some research on recent machine learning models and summarize it for me” and it'll use my browser and do some research and scroll through the pages and give me all its findings. Just like a personal assistant, this can do it for you. Scheduling meetings, anything.
Is this personal assistant customized to the user? As you're asking it to do tasks, does it learn from the behavior that you're asking it to perform?
That’s something that we're working on today. We are still focused on making our base model really great so that it works reliably to do what you want it to do, but we're working on building a system that can understand you, learn from you, and be personalized to you. We want to make sure it's seamless and the user doesn't have to think about it. A lot of the AI systems say that they are personalized, but the user has to put in a lot of work. Our goal is to make personalization core to our product but done automatically in the background. We’re getting there.
How are you building the AI models you use in your products?
We've trained models from scratch and we’ve fine-tuned base models. Our objective is to end up with the best models at the lowest cost; by working with MosaicML, we can afford to experiment. As we improve our techniques and finetune our data sets, we are able to continually improve our models.
Why did you choose MosaicML to work with?
We want to work with companies that are willing to just sort of get in the trenches and help us get to where we want to go. It was clear that instead of going through 20 sales calls, your team was like: ”Alright, let's get this training run started by next week.” And that was a very, very different experience than we'd seen before.
What was your experience like training on our platform?
I've trained some models but nothing at large scale, and I was worried that I wasn't going to be able to do it, but this platform is usable by anybody. It's more about understanding what you want to do, building the right data system, and making key architectural decisions than implementing the code. Your platform makes it easy, it abstracts away the complexity of training.
We’ve been training models outside of MosaicML, but what we find is that with the same model, the same data, the same architecture, and everything else the same, it'll train three times faster on your platform because of the optimizations you’ve done. We've trained models as small as 1B parameters and as large as nearly 100B parameters on your platform. It’s amazing.
Your engineering team was willing to jump on calls and get this thing running when there were blockers, which was helpful. That's how we ended up doing it so fast. You’re willing to support new models that come out within hours of them being released. We're able to experiment faster because of the ability to train faster. Using your platform, we can train 3 times faster, so we can train three models in the time it would normally take us to train one, evaluate them and compare them. Plus, because it's 3X faster training, we're saving 3X on the training cost with GPUs. So we're able to try more things for the same amount of money, which is amazing.
What trends do you see in the Generative AI industry?
There are a lot of headwinds moving in the open source direction, which is great for MosaicML. Today we're getting close to the point where there are models out there that truly compete with closed source base models at a fraction of the cost, and that you can host yourself, control the throughput, and control the latency.
With finetuning, you're going to be able to make things that are far better than what is possible with the closed source APIs, again at a fraction of the cost, again at a fraction of the latency. So you're going to get the benefits of having these tiny specialized models with intelligence that's better for your task than closed APIs. I still think there will be times when you’ll want the closed APIs with the best possible massive models. There's no way you're going to host that kind of model yourself for things that require more than just task specialization. But I do think that a lot of folks are going to start moving toward finetuning. And yeah, there's a little bit of cost to train, but when you look at inference, it's probably going to cost less to train those models than it is to serve them for one month on a large API.
_____________________________________
To learn more about OthersideAI’s platform of AI-powered writing tools, including a universal translator, magic editor, proposal generator, and more, visit: https://www.hyperwriteai.com/aitools
To start experimenting with OthersideAI’s AI-powered personal assistant, sign up for a trial: https://www.hyperwriteai.com/personal-assistant
