Silicon Self-Awakening: How AI Moves Beyond Exhausted Human Data with SAGE, Agentic Proposing, and MGPO

Forum topic · ✨步子哥 · 2026-04-18

Summary

This zhichai.net forum post examines what happens when AI exhausts humanity's high-quality training data. Citing projections that human-generated text may be depleted between 2026 and 2028, the author argues the field is entering a self-bootstrapped evolution era rather than a stagnation point. Three approaches are highlighted: the SAGE (Skill Augmented GRPO) framework from UW-Madison and AWS, which packages AI capabilities into reusable, verifiable modular skill libraries; the Agentic Proposing paper (arXiv:2602.03279) from Alibaba and Shanghai Jiao Tong University, which turns problem synthesis into a goal-driven sequential decision loop of Draft, Reflect, and Dynamic Pruning steps; and MGPO (Multi-Granularity Policy Optimization), which addresses sparse rewards in long-chain reasoning. A reported result: a 30B-parameter model trained on roughly 11,000 self-synthesized reasoning trajectories achieved 91.6% (claimed SOTA) on the AIME2025 math benchmark, outperforming much larger models. The post concludes that parameter brute force is ending and autonomous self-improvement is the next phase.

Key points

The post argues that as AI consumes nearly all high-quality human-generated data (web text, papers, books, code), the industry approaches a "data exhaustion wall" projected for 2026–2028. Re-training on stale data leads to "entropy collapse" — degraded logic and creativity — signaling the end of the parameter-brute-force era. Behind the wall, however, lies self-bootstrapped evolution, illustrated by three techniques:

SAGE (Skill Augmented GRPO) — from University of Wisconsin-Madison and AWS. Knowledge is decomposed into a modular, reusable, composable, verifiable skill library. Each skill has three layers: metadata (what it is, when to use it), construction logic (step-by-step procedure), and an external verification tool τ (APIs, simulators, or real-world feedback). Sequential rollout and skill-weighted rewards let the model accumulate and recombine skills like building blocks, so capabilities scale dynamically instead of being learned from scratch.
Agentic Proposing — from Alibaba and Shanghai Jiao Tong University (arXiv:2602.03279, February 2026). Problem synthesis becomes a goal-driven sequential decision process: an "AI professor" agent loops through Draft → Reflect → Dynamic Pruning. Skills are encoded as modular SOPs (triples/quadruples of intent ι, method μ, description δ, verification tool τ). The agent drafts hard problems, self-critiques logical flaws, and prunes low-quality parts with external tools, yielding top-grade self-synthesized training data without human-authored problems.
MGPO (Multi-Granularity Policy Optimization) — addresses sparse rewards in long-chain reasoning by giving feedback at every step, segment, and final result, rather than only at the end.
Headline result: a 30B-parameter model trained on about 11,000 self-synthesized reasoning trajectories reportedly scored 91.6% on AIME2025, outperforming 600B+ parameter frontier models — evidence that modular skills plus self-synthesized data can beat sheer scale.

Takeaway

The combined stack — SAGE as the skill "hardware," Agentic Proposing as self-question-generation "software," MGPO as the optimization engine — shifts AI from passive data consumption to autonomous skill construction and self-directed iteration. The author invites readers to try open implementations on GitHub and run a self-synthesis loop on math or coding tasks, arguing autonomous evolution, not parameter scale, defines the next era.

Silicon Self-Awakening: How AI Moves Beyond Exhausted Human Data with SAGE, Agentic Proposing, and MGPO

Summary

Key points

Takeaway

Tags