Reflections on AI at the end of 2025
Overview
Section titled “Overview”Salvatore Sanfilippo, the creator of Redis, provides a profound analysis of the state of AI at the end of 2025. His core viewpoints can be summarized into the following five aspects:
1. The "Verification and Dismissal" of Cognitive Capabilities
Section titled “1. The "Verification and Dismissal" of Cognitive Capabilities”- Shedding the "Stochastic Parrot" Label: By 2025, the academic community finally reached a consensus, acknowledging that LLMs are not merely probabilistic prediction machines, but systems possessing internal representations of the meaning of prompts and their output content.
- The Essence of Chain of Thought (CoT): CoT is viewed as a form of "internal search." By sampling within the representation space combined with Reinforcement Learning (RL), models can purposefully converge to useful answers by altering their own states.
2. Evolutionary Drivers: From "Scale" to "Reinforcement Learning"
Section titled “2. Evolutionary Drivers: From "Scale" to "Reinforcement Learning"”- Breaking the Data Bottleneck: Relying on Reinforcement Learning with "verifiable rewards," improvements in AI are no longer strictly limited by the quantity of human corpus data.
- The Next Big Thing: By continuously evolving in domains with clear reward signals, such as programming optimization, LLMs combined with RL will become the core driving force of AI development.
3. The Reshaping of Programming Paradigms
Section titled “3. The Reshaping of Programming Paradigms”- Conversion of Skeptics: Due to a significant increase in Return on Investment (ROI), even the most conservative programmers have begun to accept AI assistance.
- Divergence in Collaboration Models: The programming world has split into two camps: those who view AI as a "conversational colleague" and those who view it as an "independent coding agent."
4. The Debate on the Path to AGI
Section titled “4. The Debate on the Path to AGI”- Architectural Pluralism: While some are searching for alternatives to Transformers (such as world models), the author believes that existing LLMs, acting as "differentiable reasoning machines," could potentially achieve AGI even without a paradigm revolution.
- The Reversal of the ARC Test: The ARC reasoning test, once thought to be insurmountable for LLMs, has now been conquered by optimized large models, validating the potential of existing architectures.
5. Conclusion and the Ultimate Challenge
Section titled “5. Conclusion and the Ultimate Challenge”- Architecture Unchanged, Perception Shifted: The author emphasizes that the underlying architecture of LLMs has not changed because of CoT; what has changed is our perception of their capabilities.
- Existential Crisis: For the next 20 years, the most fundamental challenge in the field of AI is not technological breakthrough, but how to avoid human extinction.