Ex-Google employees created Adept AI to develop what is known as Artificial General Intelligence (AGI), which is capable of using every software available in the world, as well as a programming interface (API) as tools to perform tasks. The company has already raised $65 million in investment, writes The Register .

Image Source: The Register

The main beneficiaries of the new project were the American investor Saam Motamedi and the co-founder of the social network for professionals LinkedIn Reid Hoffman from the venture capital fund Greylock. Also interested in the Adept AI project were Andrej Karpathy, head of Tesla autopilot system development, Jaan Tallinn, one of the founders of the Skype messenger, and Chris Ré, computer scientist at Stanford University and co-founder of Lattice Data, which was acquired by Apple in 2017.

Adept AI co-founder and CEO David Luan says the company is a “research lab for developing and building strong artificial intelligence.” The company also includes CTO Niki Parmar, Chief Scientist Ashish Vaswani and a team formerly of DeepMind. The latter is owned by Google and is engaged in AI technology projects.

The founders of Adept AI have deep experience in developing natural language processing algorithms from OpenAI. Before leaving Google Luan, Parmar and Vasnavi worked on the popular GPT-2 and GPT-3 language models for AI training. In addition, Parmar and Vasnavi are the developers of the Transformer neural network architecture, which they introduced in 2017. It is used for processing sequences such as natural language text and for tasks such as machine translation and automatic summarization.

“Transformer was the first neural network that could “just work” with all tasks using AI. The results convinced me that the creation of strong artificial intelligence is indeed possible. We continued to train more and more complex transformers in the hope of one day developing one AGI model to work with all machine learning methods. However, in the end, they came to the conclusion that the current models have one serious limitation : they can be trained on text data and, as a result, are able to write a really great work, but they are unable to interact with the digital world , ”comments Luan.

Image Source: Pixabay.com / geralt

Adept AI will focus on training a neural network to perform common tasks on a computer, such as generating reports on task completion or compiling data using existing software such as Photoshop, Tableau and Twilio.

“To explain with an example, you can use the GPT-3 network to discuss ordering pizza, but you can’t force it to order pizza for you ,” says Luan.

The Adept AI model will serve as an “add-on” with which the user will be able to interact with a computer or mobile device through natural language. The system will learn to perform the tasks assigned to it using the software tools available to it. The most important thing in this project is that a system trained in the use of software and programming interfaces, at least in theory, will be able to execute user commands, taking into account the most effective methods “in its opinion” , that is, in fact, on its own initiative. This flexible and scalable kind of interaction between the user and the system differs from the more trivial and rigid approach of deciphering the command and, for example, parsing it for hard-coded instructions and keywords.

The development of Adept will be based on the modern architecture of the transformer. “Only unlike models that are trained on texts, we will train our model on interactions with a computer,” Luan adds.

Initially, the Adept will be able to perform simple tasks, but over time, their complexity will increase. According to Luan, developers will one day be able to use their model, for example, to help with brainstorming and finding new ideas, and even to perform scientific research.