“把算术外包给计算器，把做饭外包给外卖，都没问题。但当我们一遇到问题就习惯性直接问 AI 时，最可怕的事情正在发生：人类正在自愿停止思考。”
作为前国际计算语言学协会（ACL）顶会主席、经典 AI 教材作者，以及培养了 Transformer 奠基论文第一作者 Ashish Vaswani 和前 OpenAI 研究副总裁 Barret Zoph 的学界泰斗，Kevin Knight 在语言智能领域深耕了三十余年。在近日钛动科技首届科技增长大会（TGC 2026）的现场，这位亲手参与缔造并见证了现代自然语言处理（NLP）兴衰的技术开拓者，没有给出通稿式的乐观预言，而是用一种罕见的坦率与自嘲，撕开了当代 AI 繁荣背后的残酷切片与深层隐忧。
在这场横跨语言学、纯数学、具身智能与认知哲学的开场演讲中，他给出的不仅是一套关于 AI 进化的底层坐标，更是一份关于技术如何反噬研究者与整个人类思维的清醒诊断：
- 一次无可奈何的技术败退：大模型的暴力美学如何在一夜之间将老一派科学家毕生的经验技巧扫入历史垃圾堆，迫使顶尖学者在承认“语言已被攻克”后，仓皇转行逃向纯数学；
- 衡量智能演进的唯一硬尺度：抛开虚无的 Benchmark 刷榜，AI 质变的核心在于“无需人类干预的长程认知时长（Cognitive Runtime）”，模型正从秒级的工具调用，跃迁至跨越数十小时乃至“百年级”的自主攻坚；
- “卧室假说”与物理世界的壁垒：围棋和数学之所以被率先攻克，是因为它们拥有近乎零成本的自我反馈机制，机器可以在“卧室”里闭关成神；而具身智能、生物医药与真实商业之所以进展迟缓，是因为物理世界的试错成本无法被算法凭空压缩；
- 圈内人的荒诞两极与认知警钟：面对技术奇点，身处一线的学者们走向了极端——有人开始抽烟（笃信 AI 必将攻克癌症），有人散尽家财不再存退休金（认定毁灭即将来临）；但真正的危机既非乌托邦也非世界末日，而是在机器长考能力爆发的今天，人类正在将自己最核心的“困惑与思考”全盘让渡。
以下为编译后 Kevin Knight 的分享整理。
一、从每天早晨与机器下棋说起
大约十年前，我在家里做了一个国际象棋机器人。它在一个真实的实体棋盘上下棋，直到今天，我每天早晨的第一件事依然是给自己冲一杯咖啡，然后和这个机器人下一盘棋。这纯粹是我的个人爱好。
差不多同一时期，DeepMind 做出了轰动世界的 AlphaGo 和 AlphaZero。它们起初学习人类大师的棋谱，但随后进化为纯粹通过自我博弈来学习——通过自己的对弈经验，探索获胜的策略。你可以形象地理解为：AlphaGo 走进自己的卧室闭关下棋，一周之后从楼上走下来，它就成了世界冠军。当时这并没有让我感到焦虑，因为棋盘游戏不是我的研究领域，我依然享受每天早晨的那盘棋。
我真正深耕几十年的领域是自然语言处理（NLP）——教机器做英汉翻译、回答问题、理解文本。在早期，这些都是极度困难的命题。读研究生时，我妻子甚至会和朋友感叹，觉得我很可怜，因为我选了一个“不可能被解决的问题”作为毕业课题。
几十年来，我们所做的核心就是构建“语言模型”。这项探索可以追溯到 1950 年克劳德·香农（Claude Shannon）。我们给机器设定了一个看似奇怪的任务，比如这句话：“Mary saw toys in the window and wanted [it / them]”（玛丽看到橱窗里的玩具，想要它/它们）。机器该选哪个？人类凭常识知道她大概率想要的是玩具（them），而不是橱窗（it）。为了做出正确预测，机器必须理解现实世界、理解人类社会运作的逻辑。正是通过预测下一个词这个看似简单的任务，我们逐步解决了机器翻译等一系列宏大问题。
到了 20 世纪 80 年代，IBM 取得了里程碑式的突破：仅通过前两个词来预测下一个词。我们沿着统计路径构建了第一代系统，机器开始通过阅读文本进行学习并与人交流。后来 Google 基于这套传统思路构建了更大的模型。当时整个学术界的共识是：这条路会按部就班地走下去，我们在自然语言处理上的进展将会是线性的、缓慢改良的。
但到了 2016 年前后，质变发生了。
二、当一辈子的经验归零，我逃向了数学
循环神经网络（RNN）、Transformer、再到 ChatGPT，机器预测下一个词的能力呈断崖式跃迁。当我们走到这一步时，整个学术界突然发现：我们居然把语言问题彻底解决了。机器不仅能理解、翻译，还能流畅地生成语言。
它是如何做到的？最底层的动力首先是训练规模：全网量级的文本被灌入模型，人类无需介入，机器自己从中掌握统计规律；其次是后训练（Post-training），引入人类反馈强化学习（RLHF），由人类对模型的回答给出优劣判断。
如果把 1950 年到 2030 年拉一条时间线，不画准确率或困惑度（Perplexity），而是画我们这群 NLP 科学家的“心理情绪曲线”，那会非常戏剧化：
- 起初我们很快乐：虽然进度停滞不前，但因为不知道未来在哪，大家探索得津津有味；
- 统计学浪潮到来：一部分人很兴奋，因为机器能从文本中学习了；另一部分热爱传统语言学、厌恶统计学的人则极其痛苦；
- 平稳推进期：英汉翻译这类技术在几十年前行得很慢但很稳，甚至可以成立机器翻译初创公司；
- 神经网络与大模型革命：整整一代研究者陷入了巨大的焦虑与混乱。这意味着我们过去几十年积累的技巧、算法和经验，在一夜之间几乎被扫进了垃圾堆，必须强行切换到全新的体系。
到了 GPT-4 时代，传统的 NLP 研究体系甚至陷入了一种迷茫：作为核心命题的“语言理解”已经被基本攻克，研究者们甚至不知道该去往何处。
我当时的反应是：既然这一局机器赢了，那我就换一个更难的战场。
过去几十年，我一直认为“语言”是人类最狡黠、最深奥的产物——词语充满多义性，排列组合变幻莫测；相比之下，“数字”按部就班，乘法口诀表清晰规整。但当我真正跳进纯数学领域研究了几个星期后，我才发现数学的混沌程度远超想象，全世界最顶尖的数学家面对最基础的数学猜想也束手无策。
于是几年前我转行做数学，在学术期刊发表论文，希望借助 AI 去加速数学研究的进程。然而就在前段时间，AI 浪潮同样席卷了数学界：系统不仅攻克了价值百万美元的千禧年数学难题之一，OpenAI 更是集中发布了 300 多篇论文，用全新的证明思路解决了 300 多个极为重要的数学问题。数学界同样感受到了当年 NLP 学者所经历的震颤。
为什么大模型能在数学领域实现这一跃迁？一方面，互联网上沉淀了古往今来顶尖数学家的大量论文，模型通读后掌握了跨领域的超强关联能力；但更关键的是推理与自校验机制。就像 AlphaGo 一样，在形式化数学系统里，证明对不对，机器能够自我验证。机器再次走进了它的“卧室”，闭关演算一周，走下楼时，它已经是顶尖的数学家。
三、机器能离开人类“自主思考”多久？
很多人凭直觉感受到 AI 进化的步伐正在加快，我的切身体会也是如此。在语言和数学领域，几乎每隔几个月就会出现彻底颠覆旧认知的突破。
但该如何量化这种“加速”？我的观察维度是：机器在脱离人类干预的情况下，能够独立、自主地为你解决问题多长时间？
- 最开始，它只能自主运转 15 秒，回答一个单轮问题，然后停下来等你输入下一句；
- 接着，它能自主运转 10 分钟，在网络上收集材料、整理事实，交给你一份结构完整的调研报告；
- 进一步，它可以在 1 小时内完全按照你的产品需求，自主写出一个小型软件；
- 而到了近期，它已经可以连续工作 10 个小时，自主构建一个庞大的软件系统，期间完全不需要向人类求助。
如果把这个逻辑推到极致：近期被攻克的那道百万美元级数学难题，背后是相当于机器自主思考了 100 年的算力消耗——具体而言，是 10,000 个智能体在无人类介入的情况下，高度协同、完全自主地连续推演了 88 个小时。
人类现实中面临的绝大多数商业和实际业务问题，复杂度大致就落在这个区间：攻克某种癌症也许需要超过 10 小时的连续深度推演，而解决企业内部复杂的增长与运营问题，同样需要突破单轮对话的局限。AI 的核心跃迁，正是这种“长程自主思考时间”的持续拉长。
四、下一轮飞跃的试金石
既然语言和纯形式化数学已被相继攻破，接下来的突破口会在哪里？答案集中在几个方向：机器人（具身智能）、生物医药，以及由 AI 完全自主运行一家企业。
决定这些领域推进快慢的本质瓶颈，在于反馈机制（Feedback Loop）从何而来。
AlphaGo 和纯数学模型之所以能以惊人速度进化，是因为它们拥有近乎零边际成本的自闭环环境——机器可以在数字世界里自我试错、自我判定；但现实物理世界没有现成的“卧室”：
- 在机器人领域：机械臂完成任务就是成功，失败了就要反思重来。但实体硬件不能光靠脑内模拟，哪怕仿真环境做得再逼真，它在物理世界的受力、摩擦与损耗依然不可避免；
- 在生物医药领域：大模型可以读完生物化学的全部文献，提出精妙的分子构型，但这些假说是否成立，必须经过现实中的湿实验，以及漫长的人体临床试验周期；
- 在自主商业实体领域：未来一定会出现由机器完全接管运转的公司，与人类企业同台竞争。商业世界的反馈极其残酷直接——就是现金流与利润。但机器公司可能还没来得及跑通正确的商业模型，手头的资金就已经消耗殆尽。
物理世界的规律不会被算法压缩，凡是需要依赖物理世界提供真实反馈的领域，AI 的演进节奏都会受到现实客观规律的拉扯。
五、最大的危机不是 AI 毁灭世界，而是人类停止思考
伴随这种加速度，整个产业弥漫着极端的撕裂感。
在我身边的 AI 研究同行中，乐观主义者和悲观主义者走向了两个极致：极端的乐观者甚至开始抽烟，因为他们坚信 AI 很快就能攻克肺癌，无需再关注身体健康；极端的悲观者则彻底放弃为退休存钱，因为他们认为 AI 很快就会带来不可抗的灾难，不如及时行乐。
我既不抽烟，也会继续存钱。我并不担心科幻小说里那些邪恶 AI 的毁灭性设定，但我确实对另一种正在悄然发生的变化感到深切的担忧。
我们把算术外包给计算器，没有任何问题；我们把粮食生产外包给农业机械，把做饭外包给外卖配送，这为我们腾出了宝贵的时间，人类的核心主体性并没有受损。
但当人类开始把“思考本身”外包给机器时，问题就完全变了。
即使像我这样的人，在日常生活中也开始频繁出现一种惯性：原本遇到一个复杂问题需要自己推演、琢磨、攻关，但现在的本能反应是“直接问 AI 吧”。当我们把认知推演、逻辑碰撞与深层困惑统统让渡给系统时，人类作为个体和整体社会，都将面临认知能力的系统性退化。
AI 不仅仅是在思考，它正在学会自主行动、理解整个数字世界，并在极其复杂的领域持续建立优势。对于每一个身处其中的人来说，我们唯一的应对方式，是在快速适应技术转向的同时，始终把住独立判断的缰绳：
让机器去做推演，但永远不要让机器替你思考。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
“The outsourcing of arithmetic to calculators and the outsourcing of cooking to out-sale is not problematic. But when we get into trouble and ask AI directly, the worst thing is happening: humans are voluntarily stopping thinking."
As former Chair of the International Association of Computing Linguistics (ACL), author of the classic AI teaching materials, as well as the first author of the Transformer foundation paper, Ashish Vaswani, and former Vice-President of Open AI Research, Barret Zoph, Kevin Knight, spent more than 30 years in the field of linguistic intelligence. At the scene of the first TGC 2026 conference on science and technology growth in recent days, this technology pioneer, who was personally involved in shaping and witnessing the emergence and decline of modern natural language processing (NLP), did not give a general optimism, but, instead, with a rare sense of frankness and self-scribbling, tore the brutal slices and deep-seated concerns behind the current AI’s prosperity.
In this opening speech on cross-linguistics, pure mathematics, intelligence and cognitive philosophy, he gave not only a set of bottom coordinates on the evolution of AI, but also a sobering diagnosis of how technology counters the thinking of researchers and humans as a whole:
- A failed technology: how the violent aesthetics of LLM swept the life-long experience of the old scientists into the wastebin of history overnight, forcing the top scholars to flee to pure mathematics after acknowledging that “language has been taken over”;
- The only hard measure of intelligent evolution: to move away from the non-existent Benchmark chart, the core of the quality change in AI lies in the long-range cognition of human intervention, which is moving from a second-class tool call to an autonomous attack that spans dozens of hours or even 100th grade;
- Barriers to the “Bedroom Hypothesis” and the physical world: chess and mathematics were the first to be attacked because they had a self-responding mechanism at close to zero cost, and machines could be shut down in the “Bedroom”; and smart, biomedicine and real business were slow because the trial costs of the physical world could not be reduced by algorithms;
- Bipolar and cognitive alarms: in the face of technological wonders, scholars in the front line are extreme — people start smoking (and believe that AI is bound to fight cancer), and people spend their money off their pensions (conceive that destruction is imminent); but the real crisis is not utopia or the end of the world, but rather today, when machine capacity breaks out, humanity is taking over its core of “disturbing and thinking”.
Below is a compilation of Kevin Knight's share.
I. Starting with every morning playing chess with machines
About 10 years ago, I made a chess robot at home. It plays chess on a real physical board, and until today, the first thing I do every morning is make myself a cup of coffee, and then play with this robot. It's just my personal hobby.
Almost the same period, DeepMind made AlphaGo and AlphaZero, the world's sensational. They first learned the recipes of the masters of humanity, but then evolved to learn purely through self-playing -- - To explore winning strategies through your own game experience. You can visualize it: AlphaGo went into his bedroom and closed the game, and came down from upstairs a week later it became the world champion. This did not cause me anxiety, because board games are not my field of study, and I still enjoy the game every morning.
The area where I've really worked for decades is natural language processing — teaching machines to translate, answer questions, understand texts. In the early days, these were extremely difficult propositions. My wife even lamented me when she went to graduate school, because I chose a topic that “can't be solved”.
For decades, the core of what we have done has been to build “language models”. This exploration dates back to Claude Shannon in 1950. We set the machine a strange task, like, "Mary saw toys in the window and wanted [it]/ them." Which one should the machine choose? Humans know with common sense that she probably wants toys, not windows. In order to make the right predictions, the machine must understand the real world and understand the logic of human society. It is by predicting the next word, a seemingly simple task, that we have gradually resolved a number of ambitious issues, such as machine translation.
By the 1980s, IBM had made a landmark breakthrough: predicting the next word only through the first two words. We built the first generation of systems along the statistical path, and machines began learning and communicating with people by reading text. Then Google built a larger model based on this traditional approach. There was a consensus throughout the academic community that the road would continue in a systematic fashion and that our progress in the processing of natural languages would be linear and slowly improved.
But around 2016, the quality changes happened.
ii. When my life's experience is zero, I escape to math.
RNN, Transformer, then ChatGPT, the machine's ability to predict the next word is off the cliff. When we reached this point, the entire academic community suddenly found that we had solved the language problem completely. The machine not only understands, translates, but also fluently produces language.
How does it work? The bottom-up dynamic is, first, the size of the training: a web-based text is injected into the model, without human intervention, from which the machine masters its own statistical patterns; and second, post-training (Post-training), which introduces human feedback and enhanced learning (RLHF), where human responses to models are judged to be excellent and poor.
It would be very dramatic to pull a time line between 1950 and 2030, not drawing accuracy or confusion, but drawing the "psychological curve" of our NLP scientists:
- At first, we were happy: although the pace was stagnating, there was a great deal of interest in exploring the future;
- The advent of a wave of statistics: one part of the population is excited because the machine can learn from the text; another part of the population that loves traditional linguistics and dislikes statistics is extremely painful;
- Steady progression: English translation was slow but steady decades ago, and even a machine translation start-up company could be established;
- Neural network and LLM revolution: an entire generation of researchers has been caught up in great anxiety and chaos. This means that the skills, algorithms and experiences that we have accumulated over the past few decades have almost been swept into the dump overnight and have had to be replaced by a completely new system.
In the GPT-4 era, the traditional NLP research system was even confused: “linguistic understanding” as the central proposition has been basically attacked, and researchers don't even know where to go.
My reaction was that, since this machine won, I would have changed to a more difficult battlefield.
Over the past few decades, I have considered “languages” to be the most sophisticated and profound product of humankind — the multiplicity of terms and the vagaries of rankings; by contrast, the “numbers” are organized in a sequence and are clearly structured by the algorithm. But when I actually jumped into the field of pure mathematics for a few weeks, I realized that the confusion of mathematics was far beyond imagination and that the world's best mathematicians were helpless in the face of the most basic mathematical conjectures.
So a few years ago, I switched to mathematics, published papers in academic journals, and wanted to use AI to speed up math research. But just a while back, the AI wave also swept the mathematics world: not only did the system capture one of the million-dollar millennium math challenges, but Open AI concentrated more than 300 papers on publish and solved more than 300 extremely important mathematical problems with a whole new provenance. The math community also felt the tremors that NLP scholars had experienced.
Why can LLM make this leap in mathematics? On the one hand, the Internet has sank a large number of papers from ancient top mathematicians, and models have been read with a strong cross-cutting connection; but more crucial is the mechanisms of reasoning and self-certification. Like AlphaGo, in the formal mathematical system, proof is right, the machine can self-validate. The machine went into its bedroom again, closed for a week, and it was already the top mathematician when it walked down the stairs.
How long can machines leave human “self-thinking”?
A lot of people feel intuitively that the pace of evolution of AI is accelerating, and so will my body. In the area of language and mathematics, breakthroughs that completely subvert old perceptions occur almost every few months.
But how can quantization be such an “accelerator”? My observation dimension is: How long can the machine be able to solve your problems independently and autonomously, without human intervention?
- At first, it can only operate on its own for 15 seconds, answer a single round of questions and stop until you enter the next sentence;
- It then operates on its own for 10 minutes, gathers material on the Internet, collates facts and gives you a structured study;
- Further, it can write a small software within one hour, entirely in line with your product needs;
- In the near future, it has been able to work for 10 hours on its own to build a large software system without any need for human help.
If you push this logic to the extreme: the millions of-dollar mathematical puzzle that was recently taken over, behind you is a 100-year calculation of the machine's power consumption — specifically, 10,000 AI agents, with no human intervention, have been running for 88 hours in a highly coordinated and completely autonomous fashion.
The vast majority of the commercial and operational problems faced in human reality fall largely within this zone: Tackling a cancer may require more than 10 hours of continuous depth, and solving complex growth and operation problems within enterprises will also require breaking the limits of a single round of dialogue. The core leap of AI is the constant increase in this “long-range autonomous reflection time”.
IV. Next round of leaps and leaps
Now that language and pure form mathematics have been breached, where will the next breakthrough be? The answers focused on several directions: robotics (with intelligence), biomedicine, and a company run by AI.
The underlying bottleneck in determining the pace of progress in these areas lies in where the Feedback Loop comes from.
AlphaGo and Pure Mathematics have evolved at an alarming rate because they have a closed ring environment at near zero marginal cost - machines can test themselves and judge themselves in the digital world; but there is no “bedroom” available in the real physical world:
- In the robotic field: a mechanical arm is a success, and if it fails, it has to rethink. However, physical hardware cannot be simulated in the brain alone, even if it is made in a realistic environment, and its stress, friction and loss in the physical world remain inevitable;
- Biomedicine: LLM can read all the biochemistry literature and propose fine molecular configurations, but these hypotheses are valid and must go through actual wet experiments and a long clinical human trial cycle;
- In the area of autonomous business entities: in the future, there will always be companies that are fully controlled by machines, competing with human enterprises. Feedback from the business world is extremely cruel and direct — cash flows and profits. But before the machine company could run through the right business model, the money on hand was depleted.
The laws of the physical world will not be compressed by algorithms, and the pace of evolution of AI will be drawn by realistic and objective patterns in all areas that depend on the physical world for real feedback.
The biggest crisis is not AI to destroy the world, but to stop thinking.
This acceleration is accompanied by an extreme sense of tearing across the industry.
Among my AI research colleagues, optimists and pessimists went to two extremes: extreme optimists even started smoking because they were convinced that AI would soon be able to tackle lung cancer without needing to pay attention to health; extreme pessimists gave up saving for retirement because they thought that AI would soon bring irresistible disaster, and that it would be better to have fun.
I don't smoke, I keep saving. I'm not worried about the destructive designs of the evil AI in science fiction, but I am deeply concerned about another change that is happening.
We outsource arithmetic to calculators without any problems; we outsource food production to agricultural machinery, and we outsource cooking to out-sale distributions, which gives us valuable time, without compromising the core of humanity.
But when humans start outsourcing “thinking themselves” to machines, the problem is completely different.
Even people like me, in everyday life, are beginning to have a habit: it was a complicated problem that they had to develop, think about, attack, but now the instinct is to ask AI directly. Human beings, as individuals and as a whole, face systemic degradation of cognitive capacity as we transfer cognitive evolution, logical collisions and deep confusion to systems.
AI is not just thinking, it is learning to act on its own, understand the digital world as a whole and build its strengths in extremely complex areas. For everyone in it, the only way to deal with it is to keep the reins of independent judgement while rapidly adapting to technology:
Let the machine do it, but never let it think for you.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.