In the language of #exploration/exploitation trade-offs in knowledge ecosystems: GenAI has taken over the exploitation side—transfer of existing knowledge; But the exploration side, where expertise deepens, and new knowledge emerges, remains active. 4/n
papers.ssrn.com/sol3/papers....

High-reputation, high-skill users stay engaged, but their roles are shifting—less novice Q&A, more expert-to-expert interactions. A potential silver lining: while #quantity declines, #quality may evolve in ways that sustain AI training data. 3/n

Our preprint (w/ Negin Maddah; link below) confirms the scale of decline via user-level analysis but reveals a key nuance: lower-skill, low-rep users disengaged, while high-skill, high-rep users remain active. (Figs. show both raw data & causal estimates+standard errors). 2/n

Interesting findings regarding diversity. However, I am not sure whether the variables used here (such as retention rate or graduation rates) are sufficient measures to compare quality.

...is essential for collaborative decisions , esp. for complex problems.

This introduces a critical trade-off: we must balance the immediate costs of potential human errors in human-AI collaboration against the long-term, system-wide costs of eliminating human exploration from the process. 5/5

2) Tension between human & AI promotes beneficial systemic #exploration rather than pure #exploitation of existing LLM knowledge. As we demonstrate in our recent preprint for dynamics of human teams, such cognitive diversity, driven by more stubborn individuals..,4/5
papers.ssrn.com/sol3/papers....

More importantly, the apparent "friction" between human and AI judgment may actually benefit in two key ways:

1) Human divergence from AI recommendations generates valuable new training data for future LLMs, particularly important given concerns about reaching #Peak_Data in AI training. 3/5

While many rush to conclude that human doctors may impede diagnostic accuracy, this interpretation misses crucial nuances: First, the current gap between LLM-only and LLM-assisted diagnoses will likely narrow as medical professionals develop greater familiarity and trust with these AI tools. 2/5

To address this issue, we use VAE (a generative model), to condense the high-dimensional, discrete action space into a low-dimensional, continuous one. This transformation enables the managing RL agent to efficiently operate within the reduced space, optimizing the network structure continuously 3/n

So using a reinforcement learning (RL) governing agent, we can -- at least in theory -- dynamically adjust these interaction networks to optimize system-wide performance. However, as the number of agents grows, the dimension of action space increases exponentially, creating scalability problem 2/n

Ready for your visit :-)