How I view LLMs as Sept 2026 (bhardwajrish.blogspot.com)

In a reflective analysis on the evolution of large language models (LLMs) as of September 2026, a tech enthusiast highlights the distinctions between two categories of models: instructions-following models (like Luna and Sonnet) and higher "common sense" models (such as Astra and Fable). The former, while previously labeled as "dumb," are recognized for their utility in specific tasks like auto-completion, excelling at executing predefined instructions efficiently and cost-effectively. However, they struggle with making nuanced, common-sense decisions. The higher common-sense models aim to grasp underlying intent and make informed trade-offs that better mirror human reasoning, representing a significant advancement in LLM capabilities. The author notes that while these models are still not perfect at mimicking human-like decision-making, their potential to orchestrate tasks with greater insight marks a crucial step toward achieving Artificial General Intelligence (AGI). The ongoing refinement of LLMs, particularly in these higher-order functions, could ultimately lead to models that not only execute commands but also navigate complex problem-solving scenarios.