The notion that merely increasing computing power can overcome the limitations of current AI models is becoming increasingly outdated. Each time a major model fails a basic logic task, some experts suggest that the next generation of models, with even more parameters, will somehow magically overcome these flaws.

However, this approach is fundamentally flawed. Autoregressive language models, which dominate the field, are essentially sophisticated autocomplete tools. They don’t truly comprehend the information they process, instead using statistical patterns to guess the most likely next token.

This approach is not only inefficient but also potentially hazardous. If we rely on AI systems to make critical decisions, such as those related to aviation or power grids, we need to trust that they are making those decisions based on sound logic and mathematical proof, not just guesswork.

Fortunately, alternative architectures like Energy-Based Models (EBMs) are gaining traction. These models are designed to compile and prove their logic, rather than relying on statistical patterns. This approach has the potential to revolutionize the field of AI and enable the development of systems that can be truly trusted to make critical decisions.

Photo by Okiki Onipede on Pexels
Photos provided by Pexels