Page 21 - AI Vol 1: Foundations of AI
P. 21
AI Agents are advanced implementations of intelligence, making it imperative that their actions
LLMs and other AI technology capable of remain beneficial and aligned with human interests.
decision-making and taking actions in the real A classic illustration of misalignment is the
world. The user provides the AI Agent with an “paperclip problem,” a thought experiment where
overarching goal and the AI Agent proceeds an AI designed to manufacture paperclips could
by creating a plan of action, executing tasks, potentially exhaust global resources or cause
and making adjustments to its plan based on harm in its single-minded pursuit of maximizing
experience, continuing this process until the paperclip production. This paperclip producing AI,
user’s defined goal has been achieved. OpenAI given only the instruction to maximize paper clip
recently released an early implementation of AI production, manipulates the stock market to drive
Agents that are capable of this decision-making down metal prices, begins buying raw materials
process and taking limited real-world actions. at vast quantities thus depleting their availability
This is currently a major area of development for purposes other than the manufacturer of
focus, and there will likely be greater numbers of paperclips, engages in activities designed to
more sophisticated AI agents in 2024. harm competing paper clip companies, and even
produces misinformation to destabilize countries
10. SUPERALIGNMENT with metal deposits to drive down prices. This
Alignment in the context of AI, particularly with scenario underscores the importance of ensuring
LLMs, refers to the ongoing efforts to ensure that that AI systems comprehend and adhere to the
these models act in ways that are in accordance with broader context and consequences of their actions,
human values and intentions. The goal is to develop beyond their immediate objectives.
models that not only understand and execute tasks
but also align these tasks with the ethical, moral, The study of alignment and superalignment is
and practical preferences of their users. not just academic but a practical necessity. As
AI systems become more integrated into critical
Superalignment goes beyond basic alignment to aspects of our lives and work, ensuring that
ensure that AI systems not only perform tasks in they act in ways that are beneficial, ethical, and
line with human intentions but also proactively aligned with our broader goals is paramount.
anticipate and align with broader human values While these concepts may seem futuristic, they
and goals, especially in scenarios where direct form a foundational aspect of responsible AI
instructions may not be clear or sufficient. development and are critical for guiding the
This aspect of AI development is crucial in the trajectory of AI advancements towards positive
hypothetical scenario where AI surpasses human and sustainable outcomes.
FOUNDATIONS OF AI | LOZANOSMITH.COM VOLUME 1 | 21
