Do transformers lose object-like 'seeing' of definitions through training?
MoonL88537 · x · 2026-09-09
- doomslide argues mathematicians learn to see definitions the way babies learn to recognize objects; this dimension of intelligence is highly deficient in LLMs and improving slowly
- MoonL88537 amplifies: this single sentence captures the entire field of deep learning with transformers — that's exactly how they work, and training is blinding them
- A follow-up claims heavy RL, personas and steering break true learning and cause the weird behaviors we observe
Related event: Heavy RL May Be Eroding Models' Genuine Understanding(2 posts)→
More from Research
- Math benchmark success shows log-linear diminishing returns with test-time compute, says ramez — sebkrier · 2026-09-09
- Must-Read Papers of the Week: MoE Scaling Laws, World Model Physics Benchmarks, VLA Training — TheTuringPost · 2026-09-09
- Fresh Memory, Stale Plans: PLANFENCE Blocks All Invalid Agent Actions Across 30 Workflows — rohanpaul_ai · 2026-09-09
- 0.54M-parameter robot policy hits 95% on LIBERO, 7700x smaller than π0.5 — ChongZzZhang · 2026-09-09
- Shanghai AI Lab proposes first systematic taxonomy of cognition-induced AI agent risks — jiqizhixin · 2026-09-09
- Stu-tU: RANSAC scoring that survives 128x scale miscalibration — ducha_aiki · 2026-09-09