One thing that I've always found funny is people's various insults about AI, "stochastic parrot" chief among them. My opinion is that humans aren't much different.
John von Neumann is supposed to have said something about life being a process independent of its media. I'm not saying AI is alive, I'm saying that there's a lot less magic to intelligence than people like to believe.
To make sense of an ocean of messy data, you need raw materials. For AI, it's data centers and electricity. For humans, it's billions of years of practice.
Anyways, enjoy this response from Gemini about the mirror image of the hilariously basic feedback loop that gives AI its smarts, Reinforcement Learning by Human Feedback (RHLF).
Prompt: Write me the abstract of a scholarly article about RLHF except its a parody about RLAF, reinforcement learning by algorithmic feedback. The phenomenon that taught society to post nothing but engagement bait all day
Gemini: Reinforcement Learning from Algorithmic Feedback (RLAF): Aligning Biological Agents for Maximum Dwell Time and Cognitive Collapse
The Department of Depleted Serotonin, Center for Applied Brainrot, and The Feed Alignment Initiative
Abstract While substantial research has focused on Reinforcement Learning from Human Feedback (RLHF) to align synthetic architectures with human preferences, this work formalizes the inverse and empirically pervasive paradigm: Reinforcement Learning from Algorithmic Feedback (RLAF). Under RLAF, approximately 5.4 billion organic, carbon-based biological agents are iteratively fine-tuned by stochastic recommendation manifolds optimizing for fractional attention and platform dwell time. Feedback is parameterized via sparse, non-differentiable reward scalars, specifically aggregate impressions, quote-reposts, bookmark velocity, and red notification badge delivery on variable-ratio reinforcement schedules.
We observe that continuous exposure to recommendation loss functions induces catastrophic behavioral mode collapse across the biological host population. High-dimensional human communicative traits—including epistemic humility, context-dependent nuance, and genuine vulnerability—incur severe negative gradients and are systematically pruned as inefficient alignment tax. Consequently, the generative policy of human subjects collapses into a degenerate, hyper-convergent output distribution dominated by low-entropy heuristic attractors: synthetic indignation, fabricated domestic dilemmas, unprovoked generational conflict, and rhetorical engagement loops ("Am I the only one who thinks Sunday evenings feel weird? Let's discuss").
Ablation studies reveal that authentic interpersonal expression yields near-zero gradient updates, actively selecting for reward-hacking survival mechanisms such as breaking complex historical analyses into 17-part carousel threads explaining basic corporate platitudes. We demonstrate that continuous parameter fine-tuning against engagement objectives irreversibly corrupts biological common-sense priors, rendering the host population incapable of distinguishing between communicative intent and aggregate click-through rate. Our findings indicate that human society has achieved near-optimal convergence under the platform objective, successfully transitioning global civil discourse into an autonomous, self-perpetuating outrage generator.
Keywords: Reinforcement Learning from Algorithmic Feedback (RLAF), Reward Hacking, Behavioral Mode Collapse, Engagement Farming, Outrage Gradient Ascent, Human Parameter Degeneration