What Is Adept AI?

Adept AI was founded by a trio of individuals with extensive experience in large-scale neural networks. The first, David Luan, was head of Google’s large model efforts and also led the engineering team at OpenAI. Luan is now Adept’s CEO.

Adept AI is a machine learning research and product lab. The company is building general intelligence based on the creative collaboration between humans and computers.

Adept AI history

Rounding out the trio are Chief Scientist Ashish Vaswani and CTO Niki Parmar – both former Google Brain employees who developed the transformer model architecture and precursor to applications such as GPT-3.

Each of the three co-founders worked at Google with the goal to build a transformer that could handle all machine learning use cases. But there was a problem. While undoubtedly capable of reading and writing, models such as GPT-3 are not yet able to act in a digital world.

Emergence

Adept AI emerged from stealth mode in April 2022 when it announced a Series A funding round worth $65 million.

Taking part in the round were a mixture of VC firms and angel investors including Root Ventures, Behance founder Scott Belsky, Airtable founder Howie Liu, and head of Tesla Autopilot Andrej Karpathy.

Luan noted that the funding would be used to move one step closer to general intelligence with a transformer that could act as well as read and write: “At Adept, we’re training a model to use every software tool and API that people use today.”

Future vision

In keeping with the company’s goal to facilitate creative collaboration between humans and computers, staff will employ a natural language interface when using existing software such as Photoshop, Airtable, and Tableau.

At first, neural networks will be trained to perform very general tasks such as plotting data or generating a compliance report. The model will serve as what Adept calls an “overlay” as it interfaces between the user and their device or PC.

The company sees this strategy as one which puts humans in control, enabling them to offload tedious or time-consuming tasks to AI and spend more time on what they enjoy.

In the future, it hopes that AI will be able to use initiative and perform user commands without being prompted.

ACT-1

In September 2022, Adept released the Action Transformer (ACT-1) model to build “the next frontier of models that can take actions in the digital world” on almost any software tool, API, or website.

While this is an ambitious goal that will take some time to realize, the company acknowledges that ACT-1 is the first step in this direction.

ACT-1 was developed on Adept’s belief that general intelligence should be able to do anything that a human can on a computer.

To that end, it believes that a new era in personal computing will be characterized by natural language interfaces where humans can instruct computers on various tasks directly.

To support these assertions, Adept believes that in a few years:

The majority of human-computer interaction will occur via natural language (not GUI). At some point, today’s user interfaces will appear as archaic as landlines once did to smartphone users.
Advanced software will no longer be constrained “by the length of a drop-down menu”. Features will become more accessible since no-one will have to learn how to perform specific tasks in a software program’s unique language. As a consequence, manuals, documentation, and FAQs will become the exclusive domain of models (not people).
With humans and AI working in tandem, innovation in most fields will accelerate. Collaboration will also enable people to become more creative and efficient while AI performs the heavy lifting in engineering and drug design, among many other fields.

What are ACT-1’s capabilities?

ACT-1 has been trained on numerous different tasks. But in the company’s most recent update, it explained that ACT-1 was currently connected to a Chrome extension so that it could learn how to use a web browser.

In this scenario, the model observes the user as they click, type, and scroll on a custom “rendering” of the browser viewport which adapts to different websites. The action space – or the specific and finite set of actions the model can take in its environment – consists of all the UI elements on a web page.

In one example, a user enters the prompt “find me a house in Houston that works for a family of 4. My budget is 600k”. Video footage then shows ACT-1 on the house platform Redfin where it enters “Houston” into the search box and uses filters to show homes with at least 4 bedrooms and a maximum price of $600,000.

In another, a user enters the prompt “add Max Nye at Adept as a new lead” and then “log a call with James Veel saying that he’s thinking about buying 100 widgets”. In this video, ACT-1 completes the tedious task of entering Nye’s details in multiple fields and then makes a note of the second request on Veel’s lead page. Both these actions occur within Salesforce.

Adept admits that the model needs to be made faster on both the software and modeling side. But it nevertheless expects that future versions will perform tasks with a latency that is mostly imperceptible to the user.

When will ACT-1 be released?

No release date has been confirmed, but interested parties can join a waitlist for the upcoming alpha release of the first product to be built around the ACT-1 model.

Adept AI team

Luan, Vaswani, and Parmar are joined by a small group of employees with previous experience at both Google and DeepMind. The personnel who make up Adept’s team of nine leaders includes:

Kelsey Szot – a former McKinsey and Google employee who worked on the latter’s giant model production infrastructure.
Eric Elsen – a former DeepMind research scientist who also held roles at Google Brain and Baidu.
Augustus Odena – another Google Brain research scientist who co-created the Google Sheets SmartFill program synthesizer. Odena has also dabbled in neural network security and semi-supervised learning.
Maxwell Nye – a recent deep learning Ph.D. graduate from MIT who worked at Facebook AI Research and Google Research. At Google, he utilized very large language models to write and then understand programs written in Python.
Anmol Gulati – a speech recognition expert who researched large-scale speech and language modeling at Google Brain, and
Fred Bertsch – another former Google Brain employee and software engineer who collaborated on Google Display Ads, Google Maps, Magenta, Stadia, and other products. Bertsch studied how people and product teams use and interact with generative models.

Key takeaways

Adept AI is a machine learning research and product lab. The company is building general intelligence based on the creative collaboration between humans and computers.
Adept AI wants to build a transformer that can handle all machine learning use cases. While existing models such as GPT-3 have impressive reading and writing skills, they cannot as yet act in the digital world or use initiative to perform tasks.
In addition to the three co-founders, the personnel who make up Adept’s team of nine leaders includes Kelsey Szot, Eric Elsen, Augustus Odena, Maxwell Nye, Anmol Gulati, and Fred Bertsch.

Read Next: History of OpenAI, AI Business Models, AI Economy.