If you've ever tried to use a digital assistant in a basement, parking lot, or other area with low signal, you know that the response time can be frustrating. Despite appearing seamless, most artificial intelligence tools still rely on remote servers, so their performance is closely tied to your internet connection.
That longstanding restriction is now being questioned. Research presented at the 2025 International Conference on Machine Learning and Autonomous Systems (ICMLAS) examines whether advanced AI models can be made efficient enough to run directly on smartphones without relying on the cloud.
https://www.linkedin.com/in/
The study was led by Rishabh Agrawal, an Austin-based data scientist and researcher. He focuses on making large-scale language models more practical for everyday devices. His research considers how existing AI systems, which are often too large and resource-intensive for local use, can be rebuilt to work within the constraints of consumer hardware.
A core part of the research involves model optimization techniques that reduce size without significantly impacting performance. One such method is known as pruning, which removes components of an AI model that contribute little to the output. Our findings show that this approach reduces the size of the model by about 60%, from about 500 MB to nearly 200 MB, which is in the smartphone-friendly range.
Running AI directly on the device also impacts speed and reliability. There is no need to send data to a distant server, and tests of the optimized model have demonstrated response times of just over 100 milliseconds. This could improve the consistency of AI tools even in areas with poor or disconnected connectivity.
Consideration for privacy is also noticeable. On-device processing limits the need to send personal data externally and addresses growing concerns with the proliferation of cloud-based AI services.
The study also further investigates knowledge distillation, a technique that trains smaller models using larger, more complex systems as a reference. Results showed that these compact models can reach nearly equivalent accuracy while significantly reducing training time.
These findings reflect a broader movement toward edge computing, where intelligence is embedded closer to users. Although challenges remain, this research suggests that offline, on-device AI is becoming a realistic direction for future consumer technology.
(This copy was created by the infotainment desk)
Big political news: AIMIM chief Asaduddin Owaisi slams Congress over denying bail to Umar Khalid and Sharjeel Imam
