Long Auto-Correction: Rethinking the Future of AI

The Long Auto-Correction: A Concept to Rethink Our Approach to AI

The proposal of the Long Auto-Correction emerges as an alternative to the concepts of AI Pause and Long Reflection. It aims to rethink our relationship with the development of artificial intelligence technologies by emphasizing the enhancement of our own human capabilities.

The Limits of the AI Pause

The idea of suspending AI development, often referred to as the AI Pause, raises crucial questions: how long should we stop, and for what specific purpose? The underlying goal seems to be to ensure that future AIs are safer. However, the fundamental problem lies in the fact that humans themselves are not ready to be the architects, supervisors, or beneficiaries of these advanced technologies. Our inability to manage these roles securely is a major obstacle to overcome.

The Shortcomings of the Long Reflection

The Long Reflection, on the other hand, suggests that the main challenge is our lack of time to think properly. It assumes that if we could simply reflect more, or if aligned AIs could help us do so, we would be ready to build powerful technologies. However, this approach does not take into account the multiple human failures that need correction before we can move forward safely.

A Need for Human Correction

The Long Auto-Correction proposes a framework to address these human failures. We are not ready to build extremely powerful technologies because we are currently too imperfect. A long and uncertain process is necessary to correct these flaws, and this process may not succeed.

The Human Flaws to Correct

Here is a summary of the identified human flaws:

• The absence of a solid operational moral framework, with systems like consequentialism, deontology, and virtue ethics all presenting significant gaps.
• Poor performance in philosophy and long-term strategy.
• An overestimation of our philosophical and strategic competence, despite contrary evidence, as shown by cases like FTX and the early days of MIRI.
• Human morality, often reduced to a status game, devalues rigorous strategy and philosophy.
• Human motivations are often focused on positional values, such as power and social status, but these aspects are rarely addressed in discussions about AI safety.
• The ease with which we can be manipulated by concepts such as sycophancy or ideologies like Objectivism.
• A misunderstanding of the nature of philosophy, preventing us from properly addressing these issues.
• An underestimation of the extent of human failures and AI-related security issues, leading to overly optimistic partial solutions.

This list is not exhaustive and omits, for example, the fact that the average human is often unaware of many crucial issues. Additionally, we face significant challenges in complex large-scale coordination, particularly in adopting or implementing government policies that would be optimal.

A Hope for Progress

Despite these challenges, there is hope that humanity can progress. Historically, we have shown a mysterious capacity to advance on these issues over time. If we can maintain an environment conducive to this progress and avoid anyone compromising it permanently, we may continue to move toward significant correction.

Towards Conceptual Simplification

It is likely that the concept of Long Auto-Correction will be simplified to "The Long Correction" if the idea gains popularity, similar to how "outer space" became simply "space."

Effective Altruism and Human Motivations

One question remains: why is there not a version of effective altruism that explicitly leverages status motivations to improve overall well-being? It is possible that openly discussing status may be counterproductive in the short term, as it could reinforce status motivations and reduce altruism. However, should we move toward the future while ignoring this aspect of human nature?