Nous Research, an emerging artificial intelligence startup known for its role in the open-source AI movement, has recently released Hermes 4, a new family of large language models. This announcement highlights a key development in the ongoing debate about access to advanced AI technologies, positioning Hermes 4 as a potential competitor to proprietary systems from major tech firms such as OpenAI and Google.
Hermes 4 differs from typical commercial AI models by offering increased user control and fewer content restrictions. The model allows users to request a wide array of tasks without the standard safety measures typically imposed in proprietary systems. Notably, it is designed to facilitate creative and engaging interactions while maintaining strong performance in reasoning, mathematics, and coding.
The models employ a feature termed “hybrid reasoning,” which enables users to switch between quick responses and detailed analytical processes. This approach aims to provide transparency into the AI’s thought process, differentiating it from other systems in the market. The largest version of Hermes 4, with 405 billion parameters, achieved high scores on various benchmarks, indicating competitive capabilities.
The training process utilized innovative systems like DataForge and Atropos to enhance the model’s performance. DataForge generates training data through complex tasks, while Atropos provides a reinforcement learning environment that focuses on skill development. Combined, these systems account for a significant training investment, utilizing extensive computational resources.
Despite its progress, Hermes 4’s release has raised questions regarding the balance between user freedom and responsible AI usage. While advocates argue for the benefits of greater control, concerns around potential misuse remain prevalent. The launch emphasizes the ongoing tension in AI development between corporate oversight and open-source principles, suggesting that the future of AI may not solely rely on the largest companies.
Source: https://venturebeat.com/ai/nous-research-drops-hermes-4-ai-models-that-outperform-chatgpt-without-content-restrictions/