LLM-AGI Debate Lacks Hard Evidence

The honest take is that there is substantial empirical evidence that LLMs can keep unlocking broader capabilities, but no fundamental result that proves next-token prediction alone is sufficient for AGI. The debate is mostly about definitions, missing capabilities, and whether agentic systems built around LLMs count as AGI.

• –Scaling-law and emergence papers show capabilities can improve nonlinearly with size and data, which is evidence for a plausible path, not a theorem.
• –Work on zero-shot reasoning and ReAct-style tool use shows LLMs can do more when prompting, planning, and external actions are added.
• –Skeptical papers argue many “emergent” gains are partly artifacts of prompting, in-context learning, or evaluation design, so raw benchmark jumps are not decisive.
• –The biggest open questions are grounded world modeling, causal understanding, continual learning, and autonomous goal pursuit rather than language generation alone.
• –Practically, the most defensible position is that LLMs are probably a core substrate for future general systems, but “LLM alone = AGI” remains unproven.