Salvatore Sanfilippo（Redis 之父）在 2025 年做年终 AI 反思，共 8 条观点。背景是：他 2009 年创造 Redis，2020 年从 Redis 退休，2024 年底回归；如今深度使用 AI 工具，把 Claude 当编码伙伴。作为“顶级工程师 + 普通 AI 用户”的组合视角，他认为讨论已从“能不能”转向“怎么解释与怎么用”。
他指出，2021 年“随机鹦鹉”比喻曾流行，但到 2025 年几乎没人再坚持：证据来自模型在法律、医学、数学等考试与竞赛中超越大多数人类，以及逆向工程显示模型内部存在概念表征；Hinton 的观点是要预测下一个词就必须在某种意义上理解句子。他也强调思维链（CoT）是被低估的突破：先把相关表征“采样进上下文”，再在强化学习帮助下用逐 token 的路径把推理推向更高质量答案。
关于扩张瓶颈，他认为“可验证奖励”的强化学习改变了只靠人类文本数据的上限：在代码优化、数学证明等能自检对错或优劣的任务上，训练信号近似可无限供给，可能带来类似 AlphaGo 第 37 手那样的“非直觉但有效”的策略。程序员态度也在一年内发生转折：因为投入产出比跨过临界点；使用方式分成“把 LLM 当同事聊天协作”和“把 LLM 当自治编码智能体”两派。他同时反驳“CoT 改变本质”的说法：架构仍是 Transformer、目标仍是预测下一个 token；ARC（2019）也从反 LLM 证据变成支持证据——2024 年底 o3 在 ARC-AGI-1 达到 75.7%，2025 年更难的 ARC-AGI-2 顶尖模型也能 50% 以上。最后，他用一句话概括未来 20 年的根本挑战：避免灭绝。
Salvatore Sanfilippo (Redis’s creator) offers an end-of-year AI reflection in 2025 with 8 points. Context: he created Redis in 2009, retired from Redis in 2020, and returned in late 2024; he now uses AI tools heavily, treating Claude as a coding partner. From a “legendary engineer + everyday user” vantage point, he argues the debate has shifted from whether LLMs work to how they work and how to use them.
He notes the “stochastic parrot” framing popularized in 2021 is, by 2025, rarely asserted: evidence includes LLM performance across law, medicine, and math-style evaluations, plus interpretability work suggesting internal concept representations. He highlights chain-of-thought as an underestimated breakthrough: it surfaces relevant internal representations into context before answering, and—paired with reinforcement learning—guides token-by-token trajectories toward better solutions.
On scaling limits, he argues “verifiable reward” RL changes the ceiling imposed by finite human text: tasks like faster code or correct proofs provide self-checkable signals, enabling near-unbounded improvement and potentially AlphaGo-like “move 37” moments. He also observes a one-year attitude flip among programmers because ROI crossed a threshold, splitting usage into “LLM as colleague” versus “LLM as autonomous coding agent.” He rejects claims that CoT changes LLM nature: the architecture remains Transformer and the objective remains next-token prediction. ARC (designed in 2019 to resist memorization) flipped from anti-LLM to pro-LLM evidence: o3 reached 75.7% on ARC-AGI-1 in late 2024, and top models exceed 50% on the harder ARC-AGI-2 in 2025. He ends with a single sentence about the next 20 years: avoid extinction.