Reddit essay: LLMs just blur information like a weighted sum — and AGI isn't here
Helpful-Series132 · reddit · 2026-09-16
A long-form Reddit argument:
- Mechanism analogy: an LLM's weighted sum is mathematically the same operation as blurring an image — models mix training information into unseen outputs rather than think.
- Human scaffolding: when models get physics wrong, companies interview real experts to steer training; raw data alone never draws conclusions.
- Implication: this explains why 300B models beat last year's 1T+ ones and why fine-tuned Qwen 2.5 can rival frontier models — the tuners understand the mechanism.
- Conclusion: probabilistic responses guarantee hallucination, so there is no AGI; "more research needed" really means the current path won't produce genuine thinking.
- Prediction: by 2030 everyone may have a personal AI doing free math/physics discovery, needing only topic-relevant data — but a true arbitrary-decision mechanism may arrive later and could get banned amid AI fear.
More from AGI Musings
- AI 'capability' can't be reduced to a single number for forecasting, researchers argue — davidmanheim · 2026-09-16
- MIT's 5-month campus AI study: 46% of undergrads use LLMs daily, 'cognitive surrender' spreading — Aiden_Tech_Ai · 2026-09-16
- AI safety discourse fixates on paperclips while ignoring power overconcentration, argues prominent voice — beffjezos · 2026-09-16
- AI video slop posters suddenly pivot to AI ideology content, viewers suspect a psyop — zsakib_ · 2026-09-16
- Martin Casado questions RSI: chasing Bostrom's recursive self-improvement may leave frontier labs behind — martin_casado · 2026-09-16
- New paper argues AI subjective time divergence is an overlooked digital-minds safety risk — SirDidymus · 2026-09-16