Ex-Googlers to build ‘general intelligence’ at Adept AI • The Register

Adept AI, an artificial intelligence R&D lab founded by ex-Googlers who helped invent the popular transformer architecture, launched on Tuesday with the ambitious goal of teaching machines how to use “every software tool and API in the world.”

The upstart raised $65 million from investors Saam Motamedi and LinkedIn co-founder Reid Hoffman at the Greylock venture capital fund; Lee Fixel at Addition; and Root Ventures. Other angel investors include: Andrej Karpathy, head of Tesla Autopilot; Jaan Tallinn, an early Skype developer; and Chris Ré, a computer scientist at Stanford University and co-founder of Lattice Data, acquired by Apple in 2017.

Co-founder and CEO David Luan said Adept is a “research and product lab building general intelligence.” Luan is joined by CTO Niki Parmar, and chief scientist Ashish Vaswani as well as an early group of employees, who have left their previous roles at Google and DeepMind.

The co-founders have a deep knowledge of transformer-based systems. Luan helped build GPT-2 and GPT-3 at OpenAI and later left for Google, working alongside Parmar and Vaswani. The pair are credited in the development of the AI transformer architecture, first proposed in 2017. 

“The transformer was the first neural network that seemed to ‘just work’ for every major AI use case – it was the research result that convinced me that general intelligence was possible,” Luan said in a statement.

“We trained bigger and bigger transformers, with the dream of eventually building one general model to power all ML use cases – but there was a clear limitation: models trained on text can write great prose, but they can’t take actions in the digital world.”

Adept will focus on training a neural network to perform general tasks on your computer, like generating compliance reports or plotting data using existing software such as Photoshop, Tableau, or Twilio. “For example, you can use GPT-3 to talk about ordering a pizza, but you can’t actually get it to do that for you,” Luan told The Register. Instead Adept’s model will act as an “overlay,” interfacing between a user and their PC or device. 

To us, it sounds like the computers in Star Trek, in which you describe an action to take, and the system figures out how to carry it out. It doesn’t sound like too much of a stretch.

Luan envisions a future where people will be able to instruct the model to perform actions using natural language. The system will have learned how to complete those tasks using available software. This part is crucial: the system will, in theory, be taught how to use software and application interfaces so that it can carry out someone’s commands on its own initiative. This flexible and scalable design is distinct from the more trivial and rigid approach of transcribing a command and, for instance, parsing it for hard-coded instructions and keywords.

Adept will continue to build on today’s transformer architecture. “Unlike models that are trained on text only, we’re training it on actions on a computer,” Luan told us.

The tasks and tools Adept will be able to control will be simple to start with, and get increasingly complex over time, we’re told. Luan suggested a scenario in which developers could use the model to help brainstorm and carry out scientific research.

“This product vision excites us not only because of how immediately useful it could be to everyone who works in front of a computer, but because we believe this is actually the most practical and safest path to general intelligence,” he added.

“Unlike giant models that generate language or make decisions on their own, ours are much narrower in scope–we’re an interface to existing software tools, making it easier to mitigate issues with bias.” ®