An LLM that's 7500× stupider (evanhahn.com)

The recently launched Kimi K2.5 large language model boasts 1 trillion parameters, positioning it as a "state of the art" generation. In contrast, the smollm2 model, with just 135 million parameters, highlights a stark difference in complexity, being described humorously as "7500× stupider." Users experimenting with smollm2 found its answers amusingly incorrect, such as misidentifying planets and inventing fictional characters in response to simple inquiries. For instance, its response regarding human habitat included erroneous references to Venus and a nonsensical interpretation of "Legend of Zelda". This release underscores an intriguing aspect of AI development, prompting reflection on the balance between the sophistication of models and their practical utility. While larger models tend to perform better, this lighter version showcases how simplicity can yield comedic value and encourage creativity. Despite its limitations in providing accurate information or functional code, smollm2 serves as a reminder that fun and engagement can also be prioritized in AI, pushing the boundaries of how language models can entertain while conveying insights about human knowledge and creativity.