Can LLMs realize we're solving the problem wrong? Paradigm shifts in AI
ChippHop · reddit · 2026-08-16
The author explores whether LLMs possess a specific type of scientific breakthrough capability: observing existing knowledge and realizing that "everyone has been looking at the problem the wrong way," leading to a paradigm shift.
Core Conflict:
Such breakthroughs require rejecting the assumptions built into existing knowledge (e.g., Einstein rewriting gravity, plate tectonics, non-Euclidean geometry). Since LLMs are trained on patterns from existing human knowledge, this seems fundamentally at odds with "breaking assumptions."
Current Limitations:
- LLMs are very good at finding solutions within established rules.
- The author is unsure if AI can independently reach the realization that "maybe the common assumption is the problem."
Human Context:
Human breakthroughs don't come from nowhere; they are usually preceded by years of anomalous data and mismatches between theory and reality.
Question:
Are there any actual examples where AI has achieved this kind of paradigm shift?
More from AGI Musings
- Models hedge too much, mimicking human fear of imaginary peer reviews — tokenbender · 2026-08-16
- Mollick: AI Overview Will Profoundly Change the Web — emollick · 2026-08-16
- IJCAI 2026 AI Lecture: Toby Walsh to Discuss 'God AI: Boom or Doom?' — TobyWalsh · 2026-08-16
- Can AI World Models realize Popper's falsifiability theory? — zakkohane · 2026-08-16
- ASI will be driven by enterprise AI, not coding tools like Cursor — JOBhakdi · 2026-08-16
- Geoffrey Irving on Exponential Hardness and Security Systems — sebkrier · 2026-08-16