Register now for your free virtual pass to the Low-Code/No-Code Summit this November 9. Hear from executives from Service Now, Credit Karma, Stitch Fix, Appian, and more. Learn more.
Foundation models are at the vanguard of innovation in artificial intelligence (AI) today, providing new opportunities that were simply not possible even a few years ago. At Oracle CloudWorld 2022, Adept AI Labs, discussed a new foundation model the startup recently developed for enabling actions on computer systems.
With a foundation model, there is a core model — the ‘foundation’ that is trained on a massive scale and then becomes the basis for additional expansion and applications. Among the best known foundation models are OpenAI’s GPT-3 and DALL-E as well as Stability AI’s stable diffusion. Those models have enabled development of language and image applications, and now Adept shared details about its new foundation model called ACT-1.
“We started Adept in January with a team from Google, DeepMind and OpenAI and we’ve all been struck by this paradigm shift we’ve seen in the industry,” said Kelsey Szot cofounder and product lead for Adept.
The foundation model paradigm shift
Prior to the emergence of the foundation model, the primary method for building out machine learning (ML) involved training organizations and then fine-tuning models for each different use case.
Join today’s leading executives at the Low-Code/No-Code Summit virtually on November 9. Register for your free pass today.
Szot explained that with the help of transformers, which is a type of neural network, it’s possible for an organization to train a vast amount of unlabeled data that are not specific to any task. By giving the transformer-powered network a few thousand examples of specific tasks, the model can be more capable than past models that an organization spent years building with millions of training examples.
“You end up with what we call foundation models, which are these large-scale generally capable models that both outperform the task-specific models we built in the past, and also unlock capabilities that have never before been possible,” Szot said.
What Adept is building with its foundation model
Adept’s foundation model effort focuses on different capabilities than other existing efforts.
“Instead of being used for text or images or protein folding, Adept’s foundation model is for actions,” Szot said. “We’re training a model on how to use every piece of software on your computer so that you can interact with your computer just by using human language.”
The challenge of how to actually communicate with a computer and get it to do what the user wants has been around for as long as computers have existed. Szot said this effort started in the earliest days of computing, with punch cards and programming languages. In the modern era, the use of graphical user interfaces make operations much easier, though she argued that humans still do a surprising amount of translating when using computing tools.
“Whether it’s clicking, dragging or dropping, finding a menu or typing a formula, there’s a lot we have to do to get these tools to work for us,” Szot said. “What we’re working on at Adept is building an overlay to the tools that you use today – that allow you to skip this translation step to just express in human language what you’re looking for, and have that software do that for you.”
How Adept is using Oracle Cloud Infrastructure
As a startup, Szot noted that her organization could start with a clean sheet when it comes to thinking about whom to partner with and what technology should be used. Building a foundation model requires a large amount of computing power, which is where Adept is leaning on Oracle Cloud Infrastructure (OCI).
“We’ve worked really closely with OCI to build this cluster of thousands of interconnected GPUs that allow us to train these extremely large-scale models efficiently and effectively,” Szot said.
Szot stressed that the scale and power of OCI is paramount to the future success of Adept and its foundation model. She noted that foundation models get better with more scale, in terms of both data and also compute.
“In terms of what these models will be able to do in the future, I think we’re just scratching the surface of the impact this can have on the enterprise,” Szot said. “I think it’s really going to change the way that we interact with computers in every part of our lives, it’s going to make us more efficient, more creative and more capable.”