What if you could run a powerful LLM directly on your smartphone?

There are emerging architectural innovations paving the way for a new wave of AI platforms. These systems aim to reduce energy consumption while enabling a wide range of enhanced and novel features, particularly supporting edge computing capabilities.

Understanding Edge Computing

Edge computing refers to data processing that occurs near the source—such as a sensor or a user’s personal device—rather than being sent elsewhere for centralized processing.

Another way to look at it is as a shift away from the cloud-centric model. While centralizing data in the cloud has benefits, like reducing the need for local infrastructure, it also brings transfer costs and less control. Processing tasks locally on hardware devices introduces numerous efficiencies, including reduced energy use—something that plays a role in addressing climate challenges.

At the forefront of this movement are the emerging Liquid Foundation Models (LFMs), which break away from conventional transformer-based LLM structures toward something more adaptive.

A September 2024 article by Carl Franzen in VentureBeat explores some of these developments. I’ll add the standard note: I’ve worked as a consultant with Liquid AI, and I’m familiar with many involved at MIT CSAIL. But verify for yourself what Franzen reports.

He states, “These new LFM models already outperform similarly sized transformer-based alternatives such as Meta’s Llama 3.1-8B and Microsoft’s Phi-3.5 3.8B.” “They’re built not only for high performance but also for operational efficiency, making them suitable across enterprise applications—from finance and biotech to consumer electronics—and for deployment on edge devices.”

Insights from a Lead Developer

In an April interview at IIA between Will Knight and Ramin Hasani of Liquid AI, further insights were shared.

Hasani described how the team used the neural structure of C elegans—a small worm—as a foundation for designing their models.

He discussed implementing these post-transformer models into devices like cars, drones, planes, and for uses in predictive finance and healthcare.

LFMs, he explained, can perform functions similar to GPT but directly on local devices.

“They can listen, and they can respond,” he noted.

Exploring Further Developments

Since launching a recent initiative, Hasani mentioned that Liquid AI has engaged in commercial talks with major firms about integrating this technology effectively within enterprise environments.

“Organizations prioritize privacy, secure AI usage, and low-latency AI applications,” he said. “These are areas where current AI providers don’t fully meet enterprise needs.”

Emphasizing the mindset of being a “scientist at heart,” Hasani highlighted the core advantages of running LLMs offline.

No Infrastructure Needed

One key takeaway from discussions around LFMs is that when models operate locally on a device, there's no reliance on external infrastructure—no need for data centers, cloud services, or network connectivity.

Essentially, these systems combine affordability with strong performance, reflecting a trend akin to Moore’s Law applied to AI. This means increasingly capable, adaptable, and manageable AI tools are becoming available faster than ever.

Keep watching this space as smarter, more efficient AI solutions continue to emerge.

The above is the detailed content of Post-Transformer Model Systems Can Drive Change. For more information, please follow other related articles on the PHP Chinese website!