Microsoft launches AI agent to outperform OpenAI’s GPT-4o — running locally on your PC with built-in human intervention triggers

by

Like what you’re reading? To stay informed about the newest news, reviews, and features, consider setting Windows Central as a preferred source in Google Search. You can learn more about the benefits of doing so on our site.

Generative AI is developing quickly, and we’re now seeing more than just basic text-generating programs. Leading AI companies, such as OpenAI, are pushing boundaries and entering a new phase – agentic AI – with innovative tools like Operator and Deep Research.

One system is built to control computers and complete tasks on its own, while the other assists people with complex online research.

Microsoft and OpenAI have updated their multi-billion-dollar partnership with some key changes. One major adjustment is a rule preventing OpenAI from claiming to have achieved artificial general intelligence (AGI) without approval from an independent group of experts. Microsoft also now has the freedom to develop AGI on its own or with other companies.

Microsoft is rapidly expanding its artificial intelligence efforts. Following a recent partnership with OpenAI, they’ve created a dedicated team called MAI Superintelligence to accelerate development. Their latest innovation, Fara-7B, is an AI agent that can perform complicated tasks for users right on their own computers and devices, as reported by VentureBeat.

This model, with 7 billion parameters, can run efficiently on smaller devices, offering faster response times and improved data privacy. Importantly, it operates independently of cloud services, increasing its dependability, privacy, and security.

Keep in mind that this model is still experimental and not yet widely released. Because Fara-7B is relatively small, users can run it directly on their own computers, which simplifies automating tasks that involve confidential information, like company data.

This model works by imitating how people use computers. It ‘looks’ at websites as images and then guesses where to click, type, or scroll, just like you would with a mouse and keyboard.

Instead of using a website’s code to explain content to screen readers, this system analyzes the actual pixels displayed on the screen. This allows it to work with websites even if their code is complex or poorly structured.

While speaking to VentureBeat, Microsoft Research’s Senior PM Yash Lara indicated:

Handling all image and video processing directly on the device gives users complete control over their visual data – often called ‘pixel sovereignty’ – because screenshots and any data used for automated tasks stay private on their device. This is particularly important for organizations in highly regulated industries like healthcare (HIPAA) and finance (GLBA), helping them comply with strict privacy rules.

Yash Lara, Microsoft

Notably, Fara-7B actually performed better than advanced, closed-source models like OpenAI’s GPT-4o on a common test for web-browsing AI agents, achieving a score of 73.5% compared to GPT-4o’s 65.1%.

To make sure AI doesn’t make mistakes when handling important or private tasks, Microsoft created Fara-7B. This AI is designed to recognize situations where extra caution is needed. When it encounters these moments – especially when dealing with personal information or needing permission – it will stop and ask a human for guidance before continuing.

Read More

2025-11-25 17:09