Two techniques for working with System One models (www.seangoedecke.com)

A recent exploration into "System One" language models highlights the introduction of Jev, a model designed to provide swift decision-making by answering multiple-choice questions. While Jev's operation remains somewhat opaque, its focus on efficient outputs rather than flexibility distinguishes it from traditional large language models (LLMs) like ChatGPT. The author suggests that any LLM can be adapted into a fast classifier by batching prompts, enabling developers to create systems that require quick and predictable responses across various applications. The significance of this approach lies in its potential to outperform domain-specific classifiers by maintaining flexibility and reducing training burdens. Key techniques for optimizing these models include tiered goal setting and tournament choice sampling, which enhance decision-making and task execution, as demonstrated in gaming applications like "Doom" and "Wikiracing." These techniques allow models to manage multiple layers of goals and improve efficiency by narrowing down choices in structured ways. As System One models evolve, they may redefine real-time AI applications and spur competition among leading AI labs to produce optimized choice-based models.