Sep 16, 2022 07:00:00

AI 'ACT-1' that can easily automate tool operations such as browsers appeared

AI startup Adept has unveiled ACT-1 , a large-scale

Transformer model trained to use digital tools. ACT-1 monitors what is happening in the browser and is connected to Chrome extensions that can perform specific actions such as clicking, typing, scrolling, and can automate operations.

ACT-1: Transformers for Actions
https://www.adept.ai/act

ACT-1 can take advanced user requests and execute them. The user simply enters the command in the text box and ACT-1 does the rest.

Instructions to ACT-1 are entered in English in the pop-up displayed on the browser. For example, enter 'I want you to find a house where a family of four lives in Houston, with a budget of $ 600,000 (about 84 million yen)'.

Then, Houston is searched on the property search site ......

'Price limit $600,000' is automatically entered in the condition input field.

The properties displayed at the top of the search rank within the budget were displayed.

Also, with the top page of Wikipedia open, ask ACT-1, 'How old is the actor who played Mace Windu in' Star Wars '?'

Then, an article about Mace Windu is displayed on Wikipedia ... ...

The link in the article will take you to the article by

Samuel L. Jackson .

ACT-1 replied, 'He is 73 years old' from the content of the article.

If you use Google Sheets, you can also automatically process the spreadsheet.

According to Adept, ACT-1 is still under development, but it has the following goals.

・Make it possible to communicate with a computer using natural language instead of GUI.
- Make it easy for beginners to use, and allow anyone to access advanced features.
・Make documents, manuals, and FAQs for machine learning models like ACT-1, not for human reading.
・Apply AI not only to browsing but also to the design and engineering of pharmaceuticals.

Adept said, 'While it's exciting that a system like ACT-1 will change what people can do, it also has the potential to harm people if used incorrectly or if it deviates from user preferences. We're very clear: our goal is to create a company that's centered around human feedback at scale, where models are evaluated on how well they meet user preferences, and our products are As we get more sophisticated and under load, we will repeatedly evaluate how well the AI is performing, and we plan to carefully and incrementally deploy machine learning techniques to combat misuse.' I commented.