Public discourse on AI risks is muddled, often confusing speculative future danger with real present-day harms. Technical discussions frequently conflate large intelligence-approximating models with standard algorithmic and statistical decision-making systems.
Nick Bostrom defines alignment as ensuring AI systems are aligned with what the people building them seek to achieve. The 'we' in this definition refers primarily to private companies like OpenAI and Anthropic.
OpenAI lists building superintelligence as a primary goal, citing economic growth and quality of life improvements as benefits. They argue that stopping superintelligence is risky because the number of actors building it is rapidly increasing.
Anthropic focuses on research-driven arguments regarding the necessity of studying models at the bleeding edge of capability to understand risks. Despite this measured approach, the company continues to raise hundreds of millions in funding.
Source: https://thegradient.pub/the-artificiality-of-alignment/
