top of page

Do I have your attention?

Recent progress in a new form of neural networks may lead to smaller, more elegant, yet powerful, intelligent agents. 

Our brains have a fantastic ability to focus. We can be in a noisy restaurant and ignore the noise and activity around us to focus on what the person at dinner with us is saying. Our brains limiting the information they process allows them to be incredibly flexible, constantly learning, adapting, and processing information from the environment in real-time, using only about 20 watts of power. 

Open AI’s GPT-4 has perhaps 1.5 trillion parameters, reflecting the "bigger is better" philosophy that has recently dominated AI development. The sheer scale of similar models allows for an impressive breadth of knowledge and capabilities, making them incredibly versatile and powerful across a wide range of tasks. However, they require enormous computational resources for training and operation, consume significant energy, and often lack transparency in how they make decisions.

Liquid Neural Networks (LNNs) operate effectively with maybe as few as 19 parameters,  allowing them to mimic the fluidity, efficiency, and adaptability of the human brain more closely. They can adjust their structure and function on-the-fly, allowing them to process new information or adapt to new tasks without needing to be retrained from scratch. The low parameter count does not necessarily limit LNNs power but underscores a new design philosophy that values flexibility and the capacity to learn from a stream of data in real-time. 

LLNs are illustrated in this TEDxMIT video. (

Why Should You Care? LLNs are an exciting new approach to developing intelligent agents. 

  • Less Data, More Smarts: Traditional AI often needs mountains of data to learn. LNNs are better at learning from smaller datasets, meaning they can be applied to situations where data is scarce, like building robots for disaster zones or creating AI for low-powered devices.

  • On-the-Go Learning: LNNs mimic the flexibility of our brains constantly adapting. This can lead to AI learning and improving as they go without needing constant retraining. Imagine a self-driving car that adjusts to new traffic patterns in real-time, or a voice assistant that gets better at understanding your accent the more you use it.

  • Explainable AI:  LNNs, due to their simpler structure, promise to be more transparent, which is crucial in fields like medicine or finance, where trusting the AI's reasoning is essential.

  • Efficiency and Sustainability: By eliminating the need for frequent retraining, LNNs can significantly reduce required computational resources and energy consumption making them more cost-effective and environmentally friendly.

  • Enhanced Real-Time Decision Making: LNNs can process and adapt to new data in real-time, making them ideal for applications requiring immediate decision-making, like autonomous vehicles adjusting to sudden changes in road conditions or smart grids responding dynamically to changes in electricity demand.

  • Robustness to Change: LNNs' adaptability makes them more robust to changes in their environment or input data, meaning they can be more reliable and less prone to failure or error when conditions change unexpectedly.

  • Innovation in Healthcare: LNNs can revolutionize how we monitor and respond to patients' conditions in real-time, from wearable devices that adapt to the wearer’s changing health conditions to intelligent diagnostic systems that learn from new cases as they come.

  • Personalized Technology: LNNs have the potential to make technology more personal and adaptive, such as smart homes that learn and adapt to your preferences over time or educational software that dynamically adjusts to each student’s learning style.

LLNs think more like your brain than traditional neural networks. That excites some of us and scares the bejeezus out of others.

Let's talk about this.

John Warner


128 views0 comments


Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page