文章以时间线将“AI agents”叙事量化为延期：行业把 2025 宣传为“agents 之年”，结果变成“谈论之年”，关键承诺被推迟到 2026 或更晚。与此相对，一篇数月前低调发表的论文《Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models》声称可用数学证明：基于 transformer 的 LLM 在复杂度达到某个阈值后“无法完成计算与 agentic 任务”，且即使加入超越纯词预测的推理模型也无法根治可靠性问题。作者之一 Vishal Sikka 直接给出二元结论：这些系统“没有办法可靠”，因此不应让其运行核电站，只能接受用于低风险事务并伴随错误。

反方用“可验证性”与进展指标对冲这一论断：在 Davos，Google AI 负责人 Demis Hassabis 声称在减少幻觉上有突破；创业公司 Harmonic 也宣称其产品 Aristotle 在编码可靠性基准上“领先”。其方法依赖形式化数学验证：把输出编码为 Lean 编程语言，以利用 Lean 的可证明性来验证代码，从而尝试“保证”系统可信度。该路径的边界同样被明确为范围限制：Harmonic 当前聚焦“mathematical superintelligence”，历史类文章等不可数学验证任务不在能力圈内。公司联合创始人 Tudor Achim 还提出一项任务尺度判断：多数模型已具备足以推理“预订旅行行程”的“纯智能水平”。

双方共识集中在一个上限数字：OpenAI 研究者在去年 9 月的论文中承认幻觉仍“困扰”领域，并用一个三模型实验佐证：让包含 ChatGPT 在内的 3 个模型给出首席作者的博士论文题目，3 个都编造了虚假题目且都报错年份；OpenAI 随后在博客中直言准确率“永远不可能达到 100%”。现实影响是企业级 agents 因工作流被幻觉打断而难以普及。折中路径是“护栏”：用外围组件过滤错误输出；Sikka 也承认可用组件绕过纯 LLM 的内在限制，而 Achim 则把幻觉视为学习机制的一部分，认为“永远存在”但可通过验证与约束缩小风险。

The article quantifies the “AI agents” narrative as a schedule slip: 2025 was marketed as “the year of agents,” but became the year of talking about them, with the transformative moment pushed to 2026 or later. A low-profile paper published months earlier, “Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models,” claims a mathematical proof that transformer-based LLMs cannot reliably perform computational and agentic tasks beyond some complexity threshold, and that even reasoning models won’t fix the reliability gap. Coauthor Vishal Sikka summarizes the implication as binary: they “cannot be reliable,” so agents should not run nuclear power plants, and even mundane automation must tolerate mistakes.

The industry counters with progress and verification claims. At Davos, Google AI chief Demis Hassabis reported breakthroughs in reducing hallucinations, and startup Harmonic says its product Aristotle tops coding-reliability benchmarks by using formal methods: it encodes outputs in the Lean programming language so proofs can verify code correctness. The scope limit is explicit—Harmonic’s mission is “mathematical superintelligence,” and non-verifiable tasks like history essays sit outside the method’s boundary. Cofounder Tudor Achim argues many models already have enough “pure intelligence” to reason through tasks like booking a travel itinerary.

A shared constraint is expressed as a ceiling on accuracy. OpenAI scientists wrote in a September paper that hallucinations persist even in the latest models, demonstrating it by asking three models (including ChatGPT) for a lead author’s dissertation title: all three fabricated titles and all misreported the year; OpenAI later stated accuracy will never reach 100%. The practical trend is that corporate adoption of agents is slowed because hallucinations can disrupt entire workflows. The proposed equilibrium is guardrails—external components to filter errors—accepting that hallucinations remain intrinsic while narrowing the delta between verification and failure.