Adept Raises $350 Million To Build AI That Learns How To Use Software For You
Chatbots rule the day in AI for now, but soon, predicts David Luan, co-founder of Adept, AI will not only display troubling human responses to typed requests, but execute them. It will do what you would do with your computer to you. Granted, such technology is still years away, but the speed of innovation in the AI space means we’re talking two to three years, not decades, according to Luan.
A self-confessed car geek, Luan envisions a world where an engineer can ask an AI assistant to create a design for a new car part, and watch him do exactly that step-by-step, choosing and entering the right software programs necessary commands or code, with his human as co-pilot. Do you want to change part of the design, test it in a car simulator software or send the blueprint to a manufacturer? In Luan’s vision, the AI would take care of all of that, too.
Adept, a barely a year old startup with just 25 employees, has raised $350 million in venture capital after demonstrating a rudimentary version of such a digital assistant. Rather than generating text like OpenAI’s ChatGPT or images like DALL-E, Adept intensively examines how humans use computers—from surfing the web to navigating a complex enterprise software tool—to create an AI model that translates a text command into sentences can convert from actions.
“With a synthesizer, a musician can play sounds from any instrument without having to learn how to play each instrument. We want to build the same for computers,” said Luan forbes.
General Catalyst and Spark Capital funded the majority of the Series B funding round, which closed with a post-money valuation of at least $1 billion per two sources involved in the transaction. The core portion of the funding was completed last fall, before ChatGPT ignited an AI craze among consumers, according to Luan. General Catalyst, the largest shareholder in the new round, beat seven competing term sheets to win its leading investment position, chief executive Deep Nishar said.
A synthesizer lets a musician play sounds of any instrument without having to learn how to play each instrument. We want to build the same for computers.
Some of the investor foam comes from the co-founder pedigree — rare in the deluge of founders who have flocked to found AI startups in recent months. “A lot of people talk about the game, but it’s very difficult for them to play the game,” Nishar said. “Have you built anything like this before? What are their capabilities?” Luan, the CEO, was vice president of engineering at OpenAI before moving to Google to lead the major modeling effort. Its co-founders Ashish Vaswani and Niki Parmar co-authored the Google research paper that invented the Transformer, the AI breakthrough that represents the “T” in GPT. (Vaswani and Parmar recently departed to found their own startup, according to a report by The Information; Luan declined to comment on the reason for the split.)
Their extensive machine learning expertise enabled Adept to create a working demo called ACT-1, which came less than a year after raising $65 million from venture firms Greylock and Addition. Back then, it performed similar tasks to ChatGPT and answered simple questions. In the months since, it’s been able to perform complex functions like importing LinkedIn URLs into the recruiting software. Advances like these have helped Adept launch strategic investors like Microsoft, Nvidia, Atlassian, and Workday, all of which market software that could one day benefit from its AI assistant. Adept is raising additional funds through these business alliances at a yet-to-be-determined valuation that is expected to exceed $1 billion, two sources said.
Conveniently, Adept displays ACT-1 as an overlay window over existing software such as Google Chrome or Salesforce. A prototype is ready for desktop, but Luan said it will be available on mobile in the future. The company has “commitments and targets on the revenue side” from a handful of partners, he added, but would not say when the public will be able to play with the AI assistant. “The level of interest in the [corporate investors] I think shows a certain sense of maturity [of the product],” he offered.
We don’t think about it [AGI] from the perspective of how other companies think about replacing humans in valuable tasks.
However, models for controlling computer actions are much less mature than their language model counterparts. For this reason, a company that’s barely a year old can potentially justify raising hundreds of millions of dollars without plans for a massive hiring frenzy or big acquisition – training such models doesn’t come cheap. “We haven’t reached the optimization phase yet,” said Luan. “What we really want is to train really powerful models that can do a lot, and then over time we’ll figure out how to make them cheaper and smaller.”
Adept’s rapid capital accumulation mirrors the strategies of other modeling companies like Anthropic, Cohere, and most notably OpenAI, where Luan’s former boss Sam Altman reportedly raised $10 billion from Microsoft earlier this year to outperform AI competitors. Like Altman, Luan aspires to help technology achieve artificial general intelligence, or “AGI,” a hypothetical AI system intelligent enough to make its own decisions without human intervention. One key difference: Luan’s AGI would be more boring and business-oriented, with people staying in the driver’s seat.
“We don’t think about replacing people in valuable jobs from the perspective of other companies,” he said. “We’re just trying to build the best possible AI teammate for everyone.”