In Marvel’s Iron Man movies, tech virtuoso Tony Stark has a right-hand man AI assistant: It can create graphs and charts, analyze data for insights, answer complex questions about business, and run experiments.
Adept, a new machine-learning startup, wants to make that kind of human-AI collaboration a reality.
The startup debuted on April 26, after raising a $65 million Series A. Adept’s founding team includes seven Google Brain alumni. CEO David Luan previously led Google’s giant model efforts and, while at OpenAI, helped ship large language models like GPT-2, GPT-3, and DALL-E. Ashish Vaswani, Adept’s chief scientist, and Niki Parmar, CTO, both helped invent the Transformer model—the type of deep-learning architecture that underpins some of the most advanced and widely-used language models, like Google’s BERT and OpenAI’s GPT-3.
The goal: Create an AI assistant for the layperson, which can be used to do their bidding via hundreds of software tools ranging from Photoshop to Airtable to Twilio.
“This Transformer-GPT-3 arc teaches us that you can create these intelligence substrates, these models that have many general capabilities, and we believe we can take these models [and] coach them to be very good at helping real people use the software that they’re using today,” Vaswani told Emerging Tech Brew.
The team declined to share additional information about the contents of early tests on the call, but it did release a video demo of early results from training a neural lever to use software tools. Adept hopes to roll out initial demos and secure their first set of customers within a year.
“Within three or four years, we want our models to be extremely general, and using a huge, very large suite of tools reliably,” Vaswani said, “and be in the hands of…[in] an ideal world, millions of users.”
Most of Adept’s founding team left Google near the end of 2021 to create the startup, which has been in operation since January.
The vision
Adept’s founders want their product to perform a wide range of tasks that would normally require a human assistant: plot a graph according to your preferences, plan and run an experiment, create slides for a team presentation, or crop a photo and then share it with colleagues.
Luan told us Adept’s product will be different from robotic process automation (RPA) because of both the process and the proposed market. RPA typically involves one company hiring another (think: RPA companies like UiPath or Automation Anywhere), then having workers clearly document their workflows, and finally incorporating custom-made automation software. Adept is planning more of a one-size-fits-all, direct-to-consumer approach.
Keep up with the innovative tech transforming business
Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.
“[Imagine] complex actions, like, ‘Hey, you know, my friends and I want to watch…the most recent Marvel movie…and can [you] just figure out a schedule that works for all of us?’” Vaswani said. “This is a complex action that goes into their calendars because it figures out the movie that’s probably some optimal distance from everyone and then…buys a ticket [and] sends everybody an email, ‘Hey, the tickets are bought.’”
The action plan
When it comes to Adept’s vision, there’s an elephant in the room: AI is notoriously lacking in common sense. And that’s central to performing complex tasks for a human—especially the type of tasks Adept’s team envisions, which involve nuance, user preference, and, oftentimes, a chain of actions.
Vaswani said that models today “do embody some common sense,” but that is a controversial claim in the AI field. For instance, when it comes to language models, the machine learning involved essentially boils down to “pattern recognition”—the models learn patterns from training data and then recreate them when deployed, Dr. Emily M. Bender, linguistics professor at the University of Washington, told us last year.
Adept’s team will use reinforcement learning to teach its models how to use common software tools, coupled with training data from software troubleshooting or instructional resources like videos, message boards, and tutorials. They also plan to introduce a beta version to make the models smarter with user feedback—banking on the idea that people have relatively similar problems, solutions, and requests, and the model can practice addressing the most common queries.
But some large-scale ML models, including language models from Google and OpenAI, have recently faced criticism for their tendency to propagate human biases.
To help address this, Adept’s team plans to select training data that’s largely focused on the software tools the model is built for.
“We’re not trying to create some fun chatbot,” Parmar told us. She added, “There’s a lot of just common repetitive actions that we’re doing with these existing tools, so it becomes much easier to define and measure what the expected output is in the scenario. So I think starting out with that, it just helps us—our foundation is inherently safer to begin with.”