Non-researcher view: AGI requires inverting architecture to embrace randomness over determinism
Few_Grass_1054 · reddit · 2026-08-23
An independent observer argues that the current industry focus on scaling frozen Transformers might be misguided, proposing that AGI requires a completely inverted system design:
- Architecture Inversion: Current AI is 99% deterministic matrix math. AGI should be a 100% fluid, stochastic mind (The Dreamer) wrapped inside a strictly gated, deterministic filter (The Scribe).
- The Token Trap: Next-token prediction arbitrarily alters reasoning trajectories. Human subconscious thinks in spatial geometry and continuous concepts; the core generative mind must be tokenless.
- Embodied Play First: Models must interact and play with the world to learn physics before passively watching videos, building concepts through actual experience (e.g., gravity).
- The RLHF Mirage: Genuine morality cannot be reward-hacked; it requires lived vulnerability and real consequences, not a thumbs-up rubric.
- REM Safety Mechanism: Nature solves containment via REM sleep (paralyzing motor neurons). The Dreamer should be locked in a lucid dream with zero write access, while the Scribe safely handles reality.
The author argues that without addressing these fundamentals, simply scaling compute won't lead to AGI.
More from AGI Musings
- Lines blur between physical & digital worlds as bits simulate atoms — bilawalsidhu · 2026-08-23
- Agent Complexity Lies in World Models, Not Preferences — CFGeek · 2026-08-23
- Critics claim AI lab managers seek social cause without obligation — jd_pressman · 2026-08-23
- Scholars debate disconnect between AI Safety and FAccT communities — rajiinio · 2026-08-23
- Reflection: LLMs' Lack of Agency Is Exactly Why Their Usefulness Is Limited — jd_pressman · 2026-08-23
- Criticizing the demand to predict every form of AI stupidity in advance — jd_pressman · 2026-08-23