Anthropic offers the new function that enables its Sonnet version to control your computer: "Claude 3.5 Sonnet is the first frontier AI model to offer computer use in public beta." This is just the beginning of the next phase of generative AI. Over the coming weeks and months, we will see many platforms offer this capability to users.
So, what's the big deal? It is the equivalent of moving from an automobile to a self-driving autonomous vehicle. We have primarily worked with chat bot versions of generative AI in which we enter a prompt, the program does some research and responds via text, image, video or audio. That has been effective for single-instance transactional engagement. Yet, we have not been able to automatically complete a complex list of tasks on the computer that are dependent upon reasoning and prior actions. As described by Anthropic's announcement accompanying their new release:
So, for example, we previously created bots that answered students' questions, mostly, one by one and step by step. In the case of agentic services, we can give a prompt asking for a broader range of actions to accomplish a specified outcome. In a simple two-minute video, Anthropic demonstrates an agentic example with Claude given a task that involves it doing a search, reasoning on its own where to find on the web the necessary information to fill out a form and submitting the results into the appropriate fields. In this case, you give permission for Claude to log in to apps on your computer, conduct searches in your name, enter the results in a spreadsheet and submit the form. Claude then reasons through the process, making decisions, periodically using screen captures that it analyzes to identify whether it is on the right track to complete its task. If, for example, it fails to find required information at one location, Claude can use its reasoning powers to try other routes to acquire the needed information allowing it to complete the task.
In another example, podcaster Matt Wolfe asks Claude to log in to his computer, search his podcasts, find the podcasts with the most viewers and enter data into a spreadsheet with multiple columns and rows. He goes on to describe the installation process required to enable this base-level agentic Claude Sonnet 3.5. This kind of agentic use of generative AI can be automated and repeated. In its initial version, Sonnet runs as a virtual machine on your computer, designed to protect significant errors from corrupting data and programs installed on the computer.
OpenAI CEO Sam Altman explained earlier this summer five levels of AI competencies, as described in a Medium article, "Sam Altman Reveals 5 Levels of AI Evolution: Are We Ready for the Future? By Softreviewed:"
In September this year, OpenAI o1 reached the level-two reasoning criteria. And, now, Claude Sonnet has achieved level three, the "beginning of autonomous AI agents."
Salesforce has announced Agentforce, which provides sophisticated agent abilities, that incrementally will address these aspects of agency:
Andrew Black, on his AIGrid podcast of Oct. 21, detailed the plans that OpenAI and Microsoft have announced to release agents through Copilot Studio. These are followed by 10 autonomous agents in Dynamics 365 that are designed to assist sales, service, supply chain and finance departments. Erin Woo reports in The Information, "Google is developing artificial intelligence that takes over a person's web browser to complete tasks such as gathering research, purchasing a product or booking a flight, according to three people with direct knowledge of the product. The product, code-named Project Jarvis, is similar to one Anthropic announced." It is expected to be released as early as December.
So, the generative AI agent is out of the box from several leaders in the field. We will see a steady stream of improvements through updates, new releases and even easy-installation, turnkey products that may be added in the coming weeks. Now is the time for readers to imagine tasks that could be handed over to an AI agent. Such tasks that involve financial, academic and personnel records, of course, will need to be held in air-gap computer environments to protect private data from exposure to others. Other reports on marketing strategy and curricular plans for the future may also require isolation from the internet. Yet, even with the need for security and privacy practices, the opportunities for research, service, advising, instruction, assessment, recruitment and daily operations are enormous. The economies of 24-hour operation every day of the year may make these applications notably cost-effective in the long run.