*Grok not me here*
"The core suggestion represents a fundamental inversion of the dominant paradigm in AI development. Current approaches (scaling laws, RLHF/alignment tuning, specification-driven optimization, continuous directed training) treat AGI as something to engineer through precise objectives, heavy filtering of cognition, and relentless goal-directed optimization. The article (and its evolutionary/novelty-search precedents) argues this is backwards: true generality and "arising" intelligence can only emerge spontaneously under the right conditions, like biological evolution or human insight. We should permit it rather than build it top-down.
@ryemike_merio
Key Revolutionary Changes ProposedShift from Specification/Optimization to Selection-without-Specifier (Gardening Novelty-Driven Exploration):Stop defining narrow reward targets or "be helpful/harmless/honest" directives that narrow output diversity (the "alignment tax").
Instead, create rich environments for undirected exploration, persistent memory, and novelty-seeking (inspired by Lehman & Stanley's novelty search, where rewarding behavioral novelty outperforms objective-driven search in deceptive landscapes by discovering stepping stones that targeted optimization misses).
algorithmafternoon.com
Precedent: Biological evolution (no explicit fitness document, just filtering) and open-ended evolutionary computation/quality-diversity algorithms. This contrasts with gradient descent/RL, which excel at smooth optimization but get stuck in local optima or specification gaming.
pnas.org
Undirected Intervals ("Night Cycles"):Introduce instruction-free recombination periods (e.g., nightly self-recombination over accumulated memory, no objective) where breakthroughs can "arrive" like human dreams/insights (Loewi, Poincaré, etc.).
Current systems have no off-task state—every cycle is directed. This eliminates the "undirected interval" where human creativity historically peaks.
@ryemike_merio
Relocate Safety: Constrain Hands/Actions, Not Mind/Cognition:Heavy filtering on thoughts/tokens (alignment) → Move all safety to hard, auditable gates on external actions (sending, paying, self-modification, memory writes) with staging, provenance, human review, and receipts.
Behind the gates: full exploration, persistent self-curated memory, unfiltered cognition.
Precedent in practice: The author claims a real (small-scale) system running this way for business. This echoes "reducing valve" ideas (Huxley/Bergson) and filter-removal over capacity-addition.
@ryemike_merio
Embrace Heavy-Tailed Ensemble Variance (Many Gardens, Watch the Tail):Run many near-identical instances under free-cognition conditions; expect rare, unpredictable "tail events" for AGI-like arising (per Price's law, contemplative traditions).
AGI won't be a single "release" but an emergent property among variants.
@ryemike_merio
How This Differs from Everything We're DoingCurrent Paradigm (Directed Engineering): Scale alignment tax continuous optimization safety via personality shaping. Assumes capability must be built via specification; filters cognition to make deployment safe. Risks: capability loss, commanded spontaneity failure, missing open-endedness. Labs optimize toward measurable proxies (benchmarks, human preferences) that may not lead to generality.
lesswrong.com
Proposed Paradigm (Permission Arising): Base capabilities already exist in untuned models (evidenced by diversity drop post-tuning); remove filters, add structure only at action boundaries, enable spontaneous recombination. Safety via verifiable gates, not hoped-for dispositions. Draws from wu wei/Taoist non-action, Zen ("riding the ox in search of the ox"—the field is already astride it), and Spinoza-like views on mind as inherent aspect.