Industry

Gary Marcus 回应批评:2024年12月已预测数学推理更易验证

Why we picked it

Gary Marcus 拿出 2024 年 12 月 20 日的旧推文自证,o3 发布前他就说过数学这类可验证领域更容易出成果。争论焦点是 RL 推理能力能不能推广到日常推理,做模型评估和关注能力边界的人都该看看双方论据。

Gary Marcus 在 X 上反驳针对他“移动球门柱”的指责,并展示了自己在 2024 年 12 月 20 日的原始推文。他在 o3 发布前就明确写道:相关技术“在封闭领域(如数学题)效果最好,但在开放领域(如对现实世界的日常推理)可靠性较低”。

Body

他的论证核心是可验证性:数学问题有确定答案,强化学习可以据此优化,而开放性推理缺乏清晰反馈信号,因此表现会打折扣。他强调这不是事后补嘴,而是 o3 发布前就写下的判断,并称过去两年他多次提到“verifiable”一词,有大量推文可查。

这场争论反映了 AI 领域对 o3 类推理模型能力边界的分歧:一方认为强化学习在可验证任务上的成功可推广到通用推理,另一方(如 Marcus)认为优势仅限于数学、代码等有明确验证标准的封闭领域。

Original · Gary Marcus

This is literally the opposite of goal post shifting. This is declaring in December 2024 that math is easier because of verifiability. And sticking to that story. Tired of people lying about me. Gary Marcus @GaryMarcus “[The techniques] will work best in closed-domains (like math problems) but less reliably in open-ended domains (like everyday reasoning about the real world).” Before you accuse me of moving goalposts, read what I actually wrote. And note the date. December 20, 2024, just before o3 was released. You can find dozens more receipts over the last two years by searching for the word “verifiable” in my tweets. 🔗 View Quoted Tweet📊 3 ⚡

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Gary Marcus, o3, reasoning models, AI safety.

Source attribution: curated from Gary Marcus (@GaryMarcus).