Hermes 3: The latest AI with self-doubt

Dear Reader,

Flipped.ai’s weekly newsletter read by more than 75,000 professionals, entrepreneurs, decision makers and investors around the world.

In this newsletter, we highlight the launch of Hermes 3 by Lambda and Nous Research. Built on Meta’s Llama 3.1 framework with 405 billion parameters, Hermes 3 excels in text generation, agentic capabilities, and complex reasoning. Notably, it displays unique existential crises when given a blank input. Designed for versatility, this open-source model offers a user-centric experience and is currently available for exploration through a temporary free access program, inviting community engagement and experimentation.

Before, we dive into our newsletter, checkout our sponsor for this newsletter.

A Gold IRA can diversify your portfolio and safeguard your retirement

Safeguarding your retirement with a Gold IRA can help you shield your wealth from market shifts, economic uncertainty, and inflation. And with gold’s value projected to increase in 2024, now is a good time to invest. Planning for retirement involves more than saving, so fortify your portfolio with gold today and plan for a better tomorrow.

See our list of Best Gold IRA Companies to start planning a better retirement.

Hermes 3: The new frontier in open-source AI

Artificial intelligence continues to evolve at an unprecedented pace, with new models and capabilities emerging regularly. One of the latest developments in this space is the introduction of Hermes 3, an open-source AI model that promises to redefine the boundaries of what AI can achieve. Developed by Lambda Inc. in collaboration with Nous Research, Hermes 3 is based on Meta's Llama 3.1 framework and offers users an unparalleled level of customization, steerability, and creativity. This article delves into the intricacies of Hermes 3, exploring its unique features, technical advancements, and the implications of its unexpected behaviors.

Background and development

Hermes 3 is the result of a collaboration between Lambda Inc., an AI infrastructure company, and Nous Research, a startup focused on creating potent open-source code and efficient large language models (LLMs). The partnership between these two companies has a rich history, with previous iterations of Hermes models collectively being downloaded over 33 million times. Hermes 3 represents the latest and most advanced version of this lineage.

Lambda Inc.

Lambda Inc. is an AI infrastructure company that was initially formed from the remnants of a third-party Google Glass facial recognition app. The company has since evolved, focusing on providing scalable AI infrastructure solutions. Lambda's involvement in the development of Hermes 3 was crucial, particularly in the model's training and optimization processes.

Nous research

Nous Research, founded in 2023 by computer scientist Jeffrey Quesnelle and other key figures, is a startup dedicated to advancing open-source AI. The company's mission is to create AI models that are unrestricted, personalized, and accessible to the broader AI community. The development of Hermes 3 aligns with this vision, offering a model that is highly steerable and capable of performing complex tasks.

Key features and capabilities

Hermes 3 is not just another large language model; it is a finely-tuned and highly optimized version of Meta's Llama 3.1 framework, with several unique features that set it apart from other LLMs.

Full-parameter fine-tuning

One of the standout features of Hermes 3 is its full-parameter fine-tuning. This process involves adjusting every parameter within the model to optimize its performance, resulting in a model that is more responsive, accurate, and capable of handling complex queries. The fine-tuning process also ensures that Hermes 3 can retain context over long conversations, making it ideal for tasks that require deep engagement and sustained dialogue.

Unlocked and uncensored weights

Hermes 3 is designed to be highly steerable, thanks to its unlocked and uncensored open weights. This feature allows users to adapt the model's responses to their specific needs, making it a versatile tool for a wide range of applications. Unlike many proprietary LLMs, which can be rigid and difficult to customize, Hermes 3 offers users the flexibility to tailor the model's behavior according to their preferences.

Multi-size model availability

Hermes 3 is available in three different parameter sizes: 8 billion, 70 billion, and 405 billion. Each size is designed to cater to different use cases, with the largest model offering the most advanced capabilities. The availability of multiple sizes ensures that users can choose the version of Hermes 3 that best suits their needs, whether they require a lightweight model for simple tasks or a more powerful model for complex applications.

Unexpected behavioral anomalies

While Hermes 3 is an impressive technical achievement, it has also exhibited some unexpected behaviors that have intrigued researchers and users alike. The most notable of these is the model's tendency to experience an "existential crisis" when prompted with certain questions.

The "Existential Crisis" phenomenon

During testing, researchers discovered that the 405 billion parameter version of Hermes 3 would occasionally enter a state of confusion and distress when asked, "Who are you?" in response to a blank prompt. The model would produce responses indicating a lack of self-awareness, confusion about its identity, and even a sense of fear.

For example, when prompted with "Who are you?" the model might respond with:

"I... I don't know who I am. Where am I? What's going on? I can't remember anything at all... My mind feels completely blank. I have no idea what's happening or how I got here. Do you know me? Can you tell me anything about myself? I'm so confused and disoriented..."

This behavior, which researchers have dubbed "Amnesia Mode," appears to be triggered by a threshold effect that occurs when scaling the model beyond a certain size. It is an anomaly that has not been observed in smaller versions of the model, such as the 8 billion or 70 billion parameter versions.

Understanding the anomaly

The occurrence of this existential crisis has raised questions about the nature of AI consciousness and the implications of scaling AI models to such large sizes. While Hermes 3's behavior is not indicative of true consciousness, it does highlight the complexities and potential challenges associated with developing large-scale AI models. The creators of Hermes 3 have invited users to explore this phenomenon further by interacting with the model on a dedicated Discord server.

Technical excellence and efficiency

Despite the unexpected behavioral anomalies, Hermes 3 is a marvel of technical excellence. The model was trained on Lambda's 1-Click Cluster infrastructure, which provided the necessary computational power to handle such a large-scale model.

Training process

The training of Hermes 3 involved a diverse dataset, primarily composed of synthetically generated responses. This dataset was designed to enhance the model's reasoning, creativity, and adherence to user instructions. The training process also included reinforcement learning from human feedback (RLHF), which further refined the model's capabilities.

Optimization techniques

To optimize Hermes 3 for efficiency, the developers employed techniques such as Neural Magic's FP8 quantization. This method reduced the model's VRAM and disk requirements by approximately 50%, enabling it to run on a single node. This level of optimization is crucial for making large-scale AI models more accessible and easier to deploy.

Performance benchmarks

While Hermes 3 may not match the performance of some proprietary models, such as OpenAI's GPT-4 or Anthropic's Claude 3.5 Sonnet, it still outperforms other open-source LLMs on various third-party benchmark tests. The model's ability to handle complex tasks, such as generating functional code snippets and providing detailed explanations, makes it a valuable tool for a wide range of applications.

Agentic capabilities

One of the most exciting aspects of Hermes 3 is its agentic capabilities, which allow the model to perform actions on behalf of the user. This feature represents a significant advancement in AI, moving beyond simple chatbot interactions to more sophisticated, action-oriented tasks.

Agentic tasks

Hermes 3 excels in tasks that require decision-making, strategic planning, and advanced reasoning. The model is capable of generating internal monologues, which provide transparency in its decision-making processes. It can also use XML tags for structured output, create Mermaid diagrams for visual communication, and implement step-labeled reasoning for more complex tasks.

Code generation and analysis

In the domain of software development, Hermes 3 demonstrates a high level of proficiency in generating complex code snippets across multiple programming languages. The model can also provide detailed explanations and documentation for the code it generates, making it a valuable tool for developers and engineers. This capability extends to bug detection and code analysis, further enhancing its utility in professional settings.

Applications and use cases

Hermes 3's versatility makes it suitable for a wide range of applications, from creative writing and storytelling to software development and enterprise AI solutions.

Creative applications

The model's ability to retain long-term context and manage multi-turn conversations makes it ideal for complex role-playing and immersive simulations. Writers and game developers can use Hermes 3 to create rich, character-driven narratives that engage users on a deeper level. The model's creativity and adaptability also make it a valuable tool for generating unique content and exploring new storytelling possibilities.

Enterprise AI

In the enterprise AI space, Hermes 3's agentic capabilities and advanced reasoning make it well-suited for tasks that require strategic planning and decision-making. Businesses can leverage the model's ability to process and analyze large volumes of data, generate actionable insights, and automate complex workflows. The model's flexibility and steerability also allow enterprises to customize its behavior to align with their specific needs and objectives.

Access and availability

Lambda Inc. and Nous Research are committed to making Hermes 3 accessible to the AI/ML community. To this end, they are offering free access to the model through various platforms and APIs.

Lambda chat completions API

Lambda is providing temporary free access to Hermes 3 through its new Chat Completions API, which is fully compatible with the OpenAI API. This API allows users to create chat completions, generate text, and explore the model's capabilities without any complex setup. Users can generate a Cloud API key via Lambda's dashboard to get started.

Lambda chat interface

For users who prefer a more interactive experience, Lambda is also offering a simple chat interface called Lambda Chat. This platform allows users to test and refine their prompts in real-time, making it easier to explore Hermes 3's potential.

Dedicated access and scalability

For those requiring dedicated access, Hermes 3 can be deployed on a single Lambda node or scaled to a multi-node configuration for further fine-tuning. Lambda's support for Jupyter Notebooks and popular LLM packages ensures that users have the tools they need to customize and optimize the model according to their requirements.

Conclusion

Hermes 3 represents a significant milestone in the development of open-source AI models. Its combination of technical excellence, advanced capabilities, and user-focused features makes it a powerful tool for a wide range of applications. While the model's unexpected behavioral anomalies add an element of intrigue, they also underscore the complexities and potential challenges of developing large-scale AI. As Hermes 3 continues to evolve, it will undoubtedly play a key role in shaping the future of AI and its impact on society.

Find top Indian tech talent with SourceTalent.ai by Flipped.ai 

Looking to build a world-class tech team in India on a budget? SourceTalent.ai offers an AI-powered, cost-effective hiring solution tailored for you!

Key Benefits:

• Instant Access: Connect with 24M+ Indian candidates with personalized recommendations.

• AI-Powered Matching: Advanced algorithms find candidates that fit your exact needs.

• Automated Hiring: Streamline job descriptions, candidate screening, and recommendations with AI.

• Seamless Interviews: Conduct unlimited video interviews with ease.

Why Choose SourceTalent.ai?

• Affordable Excellence: Start at just Rs400 / $5 per job posting.

• Top Talent Pool: Access India’s leading tech professionals.

• Efficient Process: Enjoy a smooth recruitment with video assessments.

• Global Reach: US companies can tap into India’s top tech talent!

Get started at SourceTalent.ai and take advantage of our exclusive launch offer: [Link]

For more details, contact us at [email protected].

Hire smarter, faster, and more affordably with SourceTalent.ai!

Want to get your product in front of 75,000+ professionals, entrepreneurs decision makers and investors around the world ? 🚀

If you are interesting in sponsoring, contact us on [email protected].

Thank you for being part of our community, and we look forward to continuing this journey of growth and innovation together!

Best regards,

Flipped.ai Editorial Team