Anthropic has published its core views on AI safety, outlining its perspective on the rapid advancement of AI and the potential for transformative systems within the next decade. The company emphasizes the importance of understanding, evaluating, and aligning AI systems with human values.
Key Points
- Anthropic believes AI progress may lead to transformative AI systems in the next decade.
- The company was founded on the belief that AI's impact might be comparable to the industrial and scientific revolutions.
- Scaling laws for AI, developed in 2019 by members of Anthropic's founding team, demonstrated predictable improvements in AI performance by increasing model size and training data.
- This research contributed to the training of GPT-3, described as the first modern "large" language model with over 173B parameters.
- Anthropic notes that previous perceived "walls" to AI progress, such as multimodality and logical reasoning, have since been overcome.
- The cost of training current AI systems is significantly less than "big science" projects, suggesting substantial room for further growth.
- Anthropic highlights that feedback loops from advanced AI, such as code models and Constitutional AI, are already accelerating AI research.
Context
According to Anthropic, the three primary factors driving predictable improvements in AI performance are training data, computation, and enhanced algorithms. The company observed in the mid-2010s that larger AI systems consistently exhibited greater intelligence. This led to the theory that the total computational budget for AI training was a critical factor in performance, with computation for the largest models growing at a rate of 10x per year.
Why It Matters
This perspective from Anthropic underscores the potential for rapid and profound societal changes driven by AI. For builders and curious readers, it highlights the ongoing need for robust safety research and ethical considerations as AI capabilities expand. The company's emphasis on scaling laws and the falling barriers to AI progress suggests that the current pace of development is likely to continue, necessitating proactive engagement from AI companies, policymakers, and civil society to prepare for transformative AI.