AI Publications: AI Conference Presentations
Developing safe, human-aligned SuperIntelligence requires a new design approach. Current opaque machine learning models and RLHF methods cannot meet key challenges of transparency, control, scalability, and sustained alignment as intelligence grows.
This paper introduces a collective intelligence architecture that integrates human and AI agents within a transparent, self-improving framework that is faster, more efficient, and inherently aligned with human values, even as SuperIntelligence advances beyond human oversight.
Marvin Minsky, Claude Shannon, Allen Newell, and Herbert Simon were among the founders of Artificial Intelligence at the 1956 Dartmouth Conference. Their pioneering ideas include Society of Mind, Information Theory, Problem Solving Theory, and Bounded Rationality.
Together, these four gifts offer insights into creating safe and aligned Superintelligent systems and understanding how humans can remain relevant in an age of advancing AI.
Developing safe, human-aligned SuperIntelligence (SI) requires addressing six challenges, including transparency, control, scalable safety, and maintaining alignment as intelligence grows.
Current opaque machine learning models with RLHF are insufficient, making a new design approach essential. Leveraging collective human and AI intelligence through a transparent, scalable architecture offers a faster, more aligned, and more efficient path to safe SI.
If Artificial General Intelligence proves to be a “winner-take-all” scenario where the first company or country to develop AGI dominates, then the first AGI must also be the safest. The safest and fastest path to AGI may be to harness the collective intelligence of multiple AI and human agents in an AGI network.
​
This approach has roots in seminal ideas from four of the scientists who founded the field of AI: Allen Newell, Marvin Minsky, Claude Shannon, and Herbert Simon.
To ensure the safe and beneficial development of AGI, a collaborative approach involving both humans and AI is proposed. By actively engaging humans throughout the development process, AI's values can be better aligned with human values. This approach allows for a broader and more diverse input from millions of people, leading to a more representative and reliable reflection of human values in AGI. This collaborative approach contrasts with the alternative approach of relying on a small group of experts to define a set of rules for AI to follow, which may not accurately capture the full spectrum of human values.
The safest path to AGI is to create a community of human and AI agents. Keeping humans in the loop for as long as possible maximizes the opportunity for humans to align the values of AI before it achieves SuperIntelligence. Enabling millions of humans to teach AI agents their values, ensures that the values of AGI reflect a statistically representative and valid sample of human values. This approach is in stark contrast to the idea of allowing AI to teach itself values by following a "constitution" created by an elite few.