The rapid advances in artificial intelligence over the past decade have brought forth astounding breakthroughs, from ChatGPT to self-driving cars to Alexa and Siri. However, some of the most buzzed-about leaps have come in the domain of natural language processing – training AI systems called large language models (LLMs) that can understand and generate sophisticated human language.
LLMs like GPT-4 and Claude 2 have demonstrated an impressive capacity to carry on free-flowing conversations. And AI chatbots are now being infused into various products and services, from search engines to customer support. But while these technologies hint at a transformation in how we interact with computers, their development has largely happened behind closed doors, shaped solely by a handful of large tech firms.
Now, researchers at Meta have open sourced an LLM conversational agent called Llama 2-Chat that achieves state-of-the-art performance. The release of such an advanced open source chatbot marks a major milestone in democratizing access to AI. But just as importantly, the way this technology was created highlights the importance of transparent, responsible AI development through open collaboration.
Crowdsourcing Human Wisdom to Teach AI Ethics
Llama 2-Chat was built by crowdsourcing input from over one million everyday people to train the system on appropriate conversational conduct. As Meta AI Research scientists Thomas Scialom and Hugo Touvron explain, they used a technique called reinforcement learning from human feedback.
The researchers began by collecting comparison data from annotators through a simple binary choice test. The annotators were shown two possible AI responses to a given conversational prompt, and asked to select which one they preferred based on criteria like helpfulness and safety. From millions of these comparisons, Llama 2-Chat learned nuanced human preferences and values when it comes to communication norms and ethics.
“Supervised data may no longer be the gold standard, and this evolving circumstance compels a re-evaluation of the concept of ‘supervision’,” the researchers stated. In other words, today it’s becoming less effective to rely on a few annotators to manually label massive datasets. By tapping into the “wisdom of the crowd,” Llama 2-Chat instead acquired an evolving, scalable sense of human mores.
This approach points the way towards a new paradigm in AI ethics – one based on mass participation. Through crowdsourcing, the developers of Llama 2-Chat managed to imbue the system with human notions of integrity at a scope virtually impossible through traditional annotation techniques. And unlike closed research, this process allowed people from diverse backgrounds to directly influence the behavioral ethics of an emerging LLM.
Open Source as a Framework for Responsibility
In addition to crowdsourcing human guidance, the other notable aspect of Llama 2-Chat’s development was its commitment to openness. All of the code, datasets, annotation processes, model architecture details and evaluations have been made publicly available.
This transparency contrasts sharply with the secrecy surrounding comparable commercial ventures like ChatGPT, Bard and Claude. The typical closed-development approach has faced backlash from civil rights groups and AI ethics researchers concerned about potential harms hidden from public view.
The researchers behind Llama 2-Chat advocate that openness should be embraced as the best framework for guiding responsible AI progress. As they wrote in their paper, “an open approach draws upon the collective wisdom, diversity, and ingenuity of the AI-practitioner community to realize the benefits of this technology.”
Through broad collaboration, they argue, emerging hazards can be better anticipated, public concerns integrated, and innovations accelerated. From this perspective, open source offers a model for accountable, community-driven technology development that considers more stakeholders.
Certainly, openness alone cannot guarantee ethical outcomes. Thoughtful regulation, safety precautions and continuous research will always be integral. However, transparency creates more surface area for public scrutiny and critique to highlight issues. Open development also allows risks to be identified earlier by a wider range of diagnosticians.
A Small Step Towards Responsible AI
The release of Llama 2-Chat represents just one modest step towards more responsible AI innovation. Meta’s researchers acknowledge that, like any LLM, their system remains a work in progress requiring ongoing safety improvements before real-world deployment. Wider community involvement will be critical to that effort.
However, this small advance powerfully demonstrates the merits of transparency, open collaboration and public participation for steering AI along a wise path. Llama 2-Chat proves that, with the right approach, everyday people can have a voice in shaping artificial intelligence that reflects shared human values.
As this technology continues permeating our lives, we must keep exploring frameworks that proactively align its influence with the public interest. The philosophers of ancient Athens saw open discussion and debate as essential to the health of their society. In the 21st century, we must foster a modern open agora – a place where the arc of AI can bend towards justice.
You can try Llama 2-Chat free of charge today at llama2.ai. Model weights can be requested from the Meta AI website.