1 month and 60 models later (www.supcpu.com)

In a recent update, an AI researcher shared their journey of training 60 models over one month, releasing 11 tailored to specific questions about model capabilities and dataset influence. Each model, ranging from the ability to play chess to writing prose in the style of Shakespeare, is designed to be knowledge-light, allowing for exploration within the constraints of smaller architectures. This hands-on experience has deepened the researcher’s understanding of AI, emphasizing the importance of model size and data shaping. Significantly, the models, while innovative, reveal a challenge: they struggle to function effectively in traditional conversational AI roles due to their limited capabilities. This highlights a growing need for novel interfaces that cater specifically to the unique strengths of these smaller models, such as specialized tools for playing games or testing text output. The researcher advocates for creating dynamic “harnesses” that can leverage the models' inherent skills, transforming them into instruments that offer surprising and engaging interactions. This approach could reshape how we utilize AI and ML technologies, moving beyond conventional interfaces to a more nuanced exploitation of AI's potential.