<antirez>
5 hours ago
- By 2025, most AI researchers stopped claiming LLMs are stochastic parrots due to accumulating functional evidence.
- Chain-of-thought (CoT) improves LLM output by enabling internal sampling and reinforcement learning for token sequences.
- Reinforcement learning with verifiable rewards overcomes scaling limits on token count, with potential for continued improvement.
- Programmer resistance to AI-assisted coding has decreased as LLMs provide useful code and hints with acceptable return on investment.
- Some researchers explore alternatives to Transformers, but LLMs may still reach AGI without new paradigms.
- Critics who now credit CoT for fundamentally changing LLMs are misleading; the architecture and token prediction remain the same.
- The ARC-AGI test has become solvable by LLMs with extensive CoT, transitioning from an anti-LLM test to validation.
- The fundamental AI challenge for the next 20 years is avoiding extinction.