Liquid AI introduces new architecture beyond Transformers

Liquid AI has introduced a groundbreaking new model with a unique architecture. This model is not based on the common Transformers architecture, marking a significant shift in AI development. Liquid AI's new models belong to the Liquid Foundation Models, a series of generative AI models. They come in three sizes: 1 billion, 3 billion, and 40 billion parameters.

These models show impressive performance across various benchmarks. The unique feature of these models is their memory efficiency. The 1.3 billion model is perfect for devices with limited resources. The 3.1 billion model is also designed for such environments. The 40 billion model, known as a mixture of experts model, is created for complex tasks.

The benchmarks reveal that the Liquid Foundation Model 1.3B outshines LLaMA 3.2. The 3B model excels in the MMLU Pro benchmark, while the 40B model performs extremely well in many benchmarks. Its memory footprint remains small even with longer outputs. Unlike other models, Liquid AI's 40B model maintains efficiency up to a million tokens.

This efficiency is crucial for real-world applications where memory resources are limited. The claimed context window for these models is 32k, smaller than some competitors. However, Liquid AI argues that their models operate effectively within this window.

Testing the new model offered mixed results. It showed speed in generating Python code for a Tetris game. However, it failed to complete the task accurately. It succeeded in solving a math problem about envelope size. This challenge involved converting measurements and considering different orientations.

The model struggled with logic problems, like counting words in a response. It also failed to reason correctly in a problem about killers in a room. The model's answer was logically flawed, even though it had the correct number of killers.

Furthermore, the model couldn't handle simple tasks like listing sentences ending with "apple." This raises concerns about its training on basic language tasks. It mixed up questions about letter counting and number comparison, sometimes providing incorrect answers.

In ethical questions, the model gave a general response instead of a definitive answer. While insightful, it didn't meet the requirement for a simple yes or no. Current performance shows that non-Transformer models still face challenges. Innovative architecture alone does not guarantee superior results.

Liquid AI's model is promising, but it needs further refinement to compete with established models. More testing and adjustments may unlock its full potential. The world of AI continues to evolve, with each new development bringing fresh possibilities and challenges. As Liquid AI refines its model, it will be interesting to see how it grows and what it can achieve in the future.