我的征尘是星辰大海。。。

The dirt and dust from my pilgrimage forms oceans of stars...

-------当记忆的篇章变得零碎，当追忆的图片变得模糊，我们只能求助于数字存储的永恒的回忆

作者:黄教授

手机视频列表

AGI大转向大厂集体放弃数学成神的道路

视频

音频

原始脚本

AGI 大转向，大厂集体放弃数学成神路线，AI底层逻辑彻底变了，AI圈正在发生一场低调但颠覆性的战略改造。
最近两年，谷歌 DeepMind 国内外头部大厂都在持续收缩，甚至裁撤 AI 数学、 AI 基础科学专项团队。
大量深耕数理推理、科学证明的顶尖人才流失，长线基础科研预算大幅削减。
放在几年前，这几乎无法理解。
过去整个行业有一条近乎神圣的共识。
数学是 AGI 的唯一必由之路。
想要造出真正的通用人工智能，必须先彻底走通数学与形式逻辑。
如今这条曾经支配行业的底层信仰。
正在被硅谷彻底推翻。
本文讲透这场路线巨变，曾经全员押注的数学成神为什么凉了？大厂是放弃逻辑了？还是看懂了 AGI 的终极真相？一、曾经的行业铁律，为什么数学被定位 AGI 终点？在2022年之前，DeepMind 定义了全球 AGI 的研发范式。
其核心逻辑清晰且极具说服力。
行业把 AI 能力分为两层，第一层是表层语言，智能聊天、写作、翻译、总结全部建立在自然语言之上。
而自然语言的本质是模糊、多义、依赖语境的，没有绝对标准答案。
模型的所谓智能，本质是海量数据的概率匹配与模式复刻，不需要真正理解，更谈不上严谨推理，只是高级仿形。
第二层是底层逻辑智能，对应的就是数学、形式化证明、多部演绎推理。
数学是人类唯一无歧义、自洽、可严格核验、绝对二元对错的符号体系，没有例外，没有模糊，没有侥幸。
当时研究者的核心判断非常笃定。
语言只能堆砌 AI 的外壳，数学才是检验真思维的内核。
一个连高阶数学证明、严谨逻辑推演都做不到的 AI 无论聊天多流畅，都只是统计拟合的产物。
绝非通用智能。
由此诞生了统治行业多年的核心路线。
AGI 的本质是极致逻辑自洽，逻辑的终极形态是数学。
所以功课数学就是打通 AGI 这就是当年谷歌重金砸向 AlphaProof 几何推理、形式化验证的底层信仰，也是国内大厂扎堆冲刺奥数解题。
数学猜想证明专项数理团队的根本原因。
彼时全行业共识，谁先跑通 AI 数学，谁就拿到 AGI 入场券。
二。
信仰崩塌，大厂不是卷不动，是路线被政委如今的团队裁撤，预算收缩，人才流失。
外界普遍解读为基础研究烧钱，回报太慢，大厂熬不住了。
这是典型的表层误解。
真实原因是数年大规模试错后，行业彻底证伪了旧 AGI 公理。
旧信仰已淘汰，数学严谨推理是 AGI 的前置门槛，必须先精通完整形式化数学体系。
才能诞生通用推理能力。
新共识，当前行业主流，数学是顶级智能的能力标尺与上限证明，但绝非 AGI 的必经起点。
一字之差，路线天翻地覆。
通俗直白翻译，以前认为不懂高阶数学，绝对成不了 AGI 现在证明，AGI可以先拥有通用世界推理能力。
再通过工具迭代后填补其数学能力。
谷歌砍掉专项数学团队，不是放弃逻辑推理，而是放弃了从零硬啃数学来孵化 AGI 的低效路线。
这是一次彻底的范式降级与工程现实主义回归。
三、核心拆解。
数学成神路线天然走不通，所有顶级大厂集体改道。
不是偶然，是三条底层硬核规律决定的。
一、 LLMs 统计智能与数学演绎智能本质互斥，当前大模型的核心能力。
来自文本数据的统计泛化，学习人类的解题步骤、推理范式、逻辑结构，复刻思维流程。
而纯数学的核心是公理演绎与绝对必然性。
每一步推导溯源至底层公理，0概率、0容错、纯逻辑内生。
这就带来一个关键真相，AI可以满分拿下 IMO，完成复杂数学证明。
但它只是复刻了人类的推理模板，没有建立底层公理认知，耗费海量算力和人力。
训练模型模仿数学家，只能得到仿真能力，无法内生通用智能，投入产出比极低，没有 AGI 孵化价值。
二。
数学只是智能子集，不是智能母体。
过去最大的认知误区是把数理逻辑等同于通用智能。
真正的 AGI 需要适配复杂的真实世界。
常识因果判断、物理世界感知、动态任务规划、模糊信息决策、社会逻辑交互、突发场景纠错，严谨数学推理仅仅是其中极小的确定性分支。
一个能破解顶级数学猜想的 AI 依然可能无法处理日常模糊场景，无法完成现实复杂规划。
反过来，一个具备完整世界建模、通用因果推理的 AGI。
随时可以通过外挂工具补齐数学短板。
专攻封闭数学领域，根本无法孵化开放世界通用智能，属于典型的舍本逐末。
三、科研周期与商业周期彻底错配。
AI 数学前沿形式化证明属于十年以上长线基础科学研究，需要长期无商业化压力的纯探索环境。
但当下 AI 行业是季度迭代、高速内卷、强商业化导向的战场。
资本市场、企业营收、产品落地的硬性要求，不允许团队长期投入无短期回报的基础探索。
独立专项数学团队的解散，不是技术失败，是工程性价比与商业现实的必然选择。
四、关键澄清。
不放弃数学，只是不再单独造神，全网大量误读。
大厂放弃数学，放弃严谨逻辑，AI彻底走向玄学统计。
完全错误，真实的行业调整逻辑非常清晰，淘汰独立专用团队，保留并内化核心能力。
一、谷歌并未停止数学推理迭代。
Gemini 的符号计算、逻辑推演、数理证明能力持续升级，只是不再设立独立长线攻坚团队。
二、 OpenAI  Anthropic 反而持续加码高阶推理。
数学逻辑难题依旧是评判顶级模型上线的核心硬指标。
三、所有头部模型的核心竞争力依然是推理深度、逻辑自洽。
结论严谨度。
行业完成了两条路线的新旧交替，旧路线已淘汰，封闭专攻数学逻辑孵化通用 AGI。
新路线当前唯一主流，搭建通用世界建模与通用推理能力，工具外挂加模型迭代，补齐数理严谨性。
简单总结。
数学从 AGI 的前置门票变成了顶级智能的上限加分项。
五、终极大局观。
这场转向定义了 AI 未来十年形态，这次低调的战略改造是 AI 从理想主义科研幻想走向现实主义工程落地的里程碑，直接敲定未来 AGI 的三大特征。
一、 AGI 诞生形态是世界通用决策者，而非专职数学家。
未来通用 AI 的核心优势是适配模糊、复杂、不确定的真实物理世界。
而非极致严谨的抽象数学世界。
落地能力、决策能力、交互能力取代纯数理能力，成为核心评判标准。
二。
未来 AI 的严谨性来自模型推理加工具校验的混合架构，不再追求模型先天内置完美公理体系。
由大模型负责思维拆解、因果推理、任务规划。
由计算器、符号库、证明器、代码脚本负责精准纠错、数值验算、逻辑兜底，以最低算力成本平衡智能泛化与逻辑严谨。
三、数理基础研究回归学术，产业全力聚焦落地纯理论、纯数学的 AI 探索，彻底退出产业主流，交由高校和实验室长线研究。
产业端所有资源全面倾斜， Agent 智能体，真实世界交互，场景落地，工程效率优化。
曾经全行业笃信极致自洽的数学逻辑。
是 AI 通往超级智能的唯一神路。
数年试错终得真相，数学是智能最高精度的检验标尺。
却不是通用智能诞生的土壤。
AGI 的真谛从来不是精通公式与证明，而是在纷繁复杂的现实世界中拥有稳定、通用、可迭代的自主思考与决策能力。
大厂集体放弃数学成神的幻想，不是倒退，是清醒。
AI不再追求用极致理论完美定义智能，而是开始用真实世界的能力。
真正实现智能。

修正脚本

AGI 大转向，大厂集体放弃数学成神路线，AI底层逻辑彻底变了，AI圈正在发生一场低调但颠覆性的战略改造。
最近两年，谷歌、DeepMind、国内外头部大厂都在持续收缩，甚至裁撤 AI 数学、 AI 基础科学专项团队。
大量深耕数理推理、科学证明的顶尖人才流失，长线基础科研预算大幅削减。
放在几年前，这几乎无法理解。
过去整个行业有一条近乎神圣的共识。
数学是 AGI 的唯一必由之路。
想要造出真正的通用人工智能，必须先彻底走通数学与形式逻辑。
如今这条曾经支配行业的底层信仰，正在被硅谷彻底推翻。
本文讲透这场路线巨变，曾经全员押注的数学成神为什么凉了？大厂是放弃逻辑了？还是看懂了 AGI 的终极真相？一、曾经的行业铁律，为什么数学被定位 AGI 终点？在2022年之前，DeepMind 定义了全球 AGI 的研发范式。
其核心逻辑清晰且极具说服力。
行业把 AI 能力分为两层，第一层是表层语言，智能聊天、写作、翻译、总结全部建立在自然语言之上。
而自然语言的本质是模糊、多义、依赖语境的，没有绝对标准答案。
模型的所谓智能，本质是海量数据的概率匹配与模式复刻，不需要真正理解，更谈不上严谨推理，只是高级仿形。
第二层是底层逻辑智能，对应的就是数学、形式化证明、多步演绎推理。
数学是人类唯一无歧义、自洽、可严格核验、绝对二元对错的符号体系，没有例外，没有模糊，没有侥幸。
当时研究者的核心判断非常笃定。
语言只能堆砌 AI 的外壳，数学才是检验真思维的内核。
一个连高阶数学证明、严谨逻辑推演都做不到的 AI，无论聊天多流畅，都只是统计拟合的产物。
绝非通用智能。
由此诞生了统治行业多年的核心路线。
AGI 的本质是极致逻辑自洽，逻辑的终极形态是数学。
所以攻克数学就是打通 AGI，这就是当年谷歌重金砸向 AlphaProof 几何推理、形式化验证的底层信仰，也是国内大厂扎堆冲刺奥数解题、数学猜想证明专项数理团队的根本原因。
彼时全行业共识，谁先跑通 AI 数学，谁就拿到 AGI 入场券。
二、信仰崩塌，大厂不是卷不动，是路线被证伪。如今的团队裁撤，预算收缩，人才流失。
外界普遍解读为基础研究烧钱，回报太慢，大厂熬不住了。
这是典型的表层误解。
真实原因是数年大规模试错后，行业彻底证伪了旧 AGI 公理。
旧公理：数学严谨推理是 AGI 的前置门槛，必须先精通完整形式化数学体系，才能诞生通用推理能力。
新共识：当前行业主流认为，数学是顶级智能的能力标尺与上限证明，但绝非 AGI 的必经起点。
一字之差，路线天翻地覆。
通俗直白翻译：以前认为不懂高阶数学，绝对成不了 AGI，现在证明，AGI可以先拥有通用世界推理能力，再通过工具迭代填补其数学能力。
谷歌砍掉专项数学团队，不是放弃逻辑推理，而是放弃了从零硬啃数学来孵化 AGI 的低效路线。
这是一次彻底的范式降级与工程现实主义回归。
三、核心拆解。
数学成神路线天然走不通，所有顶级大厂集体改道，不是偶然，是三条底层硬核规律决定的。
一、 LLMs 统计智能与数学演绎智能本质互斥：当前大模型的核心能力来自文本数据的统计泛化，学习人类的解题步骤、推理范式、逻辑结构，复刻思维流程。
而纯数学的核心是公理演绎与绝对必然性，每一步推导溯源至底层公理，0概率、0容错、纯逻辑内生。
这就带来一个关键真相，AI可以满分拿下 IMO，完成复杂数学证明，但它只是复刻了人类的推理模板，没有建立底层公理认知，耗费海量算力和人力，训练模型模仿数学家，只能得到仿真能力，无法内生通用智能，投入产出比极低，没有 AGI 孵化价值。
二、数学只是智能子集，不是智能母体。
过去最大的认知误区是把数理逻辑等同于通用智能。
真正的 AGI 需要适配复杂的真实世界，常识因果判断、物理世界感知、动态任务规划、模糊信息决策、社会逻辑交互、突发场景纠错，严谨数学推理仅仅是其中极小的确定性分支。
一个能破解顶级数学猜想的 AI 依然可能无法处理日常模糊场景，无法完成现实复杂规划。
反过来，一个具备完整世界建模、通用因果推理的 AGI，随时可以通过外挂工具补齐数学短板。
专攻封闭数学领域，根本无法孵化开放世界通用智能，属于典型的舍本逐末。
三、科研周期与商业周期彻底错配。
AI 数学前沿形式化证明属于十年以上长线基础科学研究，需要长期无商业化压力的纯探索环境。
但当下 AI 行业是季度迭代、高速内卷、强商业化导向的战场，资本市场、企业营收、产品落地的硬性要求，不允许团队长期投入无短期回报的基础探索。
独立专项数学团队的解散，不是技术失败，是工程性价比与商业现实的必然选择。
四、关键澄清。
不放弃数学，只是不再单独造神，全网大量误读：大厂放弃数学，放弃严谨逻辑，AI彻底走向玄学统计。
完全错误，真实的行业调整逻辑非常清晰：淘汰独立专用团队，保留并内化核心能力。
一、谷歌并未停止数学推理迭代，Gemini 的符号计算、逻辑推演、数理证明能力持续升级，只是不再设立独立长线攻坚团队。
二、 OpenAI、Anthropic 反而持续加码高阶推理，数学逻辑难题依旧是评判顶级模型上线的核心硬指标。
三、所有头部模型的核心竞争力依然是推理深度、逻辑自洽、结论严谨度。
行业完成了两条路线的新旧交替：旧路线是封闭专攻数学逻辑孵化通用 AGI，已淘汰。新路线是当前唯一主流，搭建通用世界建模与通用推理能力，工具外挂加模型迭代，补齐数理严谨性。
简单总结：数学从 AGI 的前置门票变成了顶级智能的上限加分项。
五、终极大局观。
这场转向定义了 AI 未来十年形态，这次低调的战略改造是 AI 从理想主义科研幻想走向现实主义工程落地的里程碑，直接敲定未来 AGI 的三大特征。
一、 AGI 诞生形态是世界通用决策者，而非专职数学家。
未来通用 AI 的核心优势是适配模糊、复杂、不确定的真实物理世界，而非极致严谨的抽象数学世界，落地能力、决策能力、交互能力取代纯数理能力，成为核心评判标准。
二、未来 AI 的严谨性来自模型推理加工具校验的混合架构，不再追求模型先天内置完美公理体系。
由大模型负责思维拆解、因果推理、任务规划，由计算器、符号库、证明器、代码脚本负责精准纠错、数值验算、逻辑兜底，以最低算力成本平衡智能泛化与逻辑严谨。
三、数理基础研究回归学术，产业全力聚焦落地：纯理论、纯数学的 AI 探索，彻底退出产业主流，交由高校和实验室长线研究。
产业端所有资源全面倾斜向 Agent 智能体、真实世界交互、场景落地、工程效率优化。
曾经全行业笃信极致自洽的数学逻辑，是 AI 通往超级智能的唯一神路。
数年试错终得真相：数学是智能最高精度的检验标尺，却不是通用智能诞生的土壤。
AGI 的真谛从来不是精通公式与证明，而是在纷繁复杂的现实世界中拥有稳定、通用、可迭代的自主思考与决策能力。
大厂集体放弃数学成神的幻想，不是倒退，是清醒。
AI不再追求用极致理论完美定义智能，而是开始用真实世界的能力真正实现智能。

英文翻译

AGI Great Shift: Big Tech Collectively Abandons the "Mathematical Ascension" Route, The Underlying Logic of AI Has Changed Completely, A low-key but disruptive strategic transformation is taking place in the AI industry.
Over the past two years, Google, DeepMind, and leading tech giants at home and abroad have continuously scaled back, and even disbanded specialized teams for AI mathematics and AI basic science.
A large number of top talents who have long been engaged in mathematical reasoning and scientific proof have left, and the budget for long-term basic scientific research has been significantly cut.
A few years ago, this was almost unimaginable.
In the past, the entire industry had an almost sacred consensus.
Mathematics is the only path to AGI.
To create real artificial general intelligence, we must first fully work through mathematics and formal logic.
Today, this underlying belief that once dominated the industry is being completely overturned by Silicon Valley.
This article thoroughly explains this great route change: why did the once industry-wide bet on "mathematical ascension" lose momentum? Are big tech giving up logic? Or have they seen through the ultimate truth of AGI?
I. The once iron law of the industry: Why was mathematics positioned as the end point of AGI?
Before 2022, DeepMind defined the global R&D paradigm of AGI.
Its core logic is clear and extremely convincing.
The industry divides AI capabilities into two layers. The first layer is surface language: intelligent chat, writing, translation and summarization are all built on natural language.
The essence of natural language is vague, polysemous, context-dependent, and has no absolute standard answer.
The so-called intelligence of a model is essentially probabilistic matching and pattern replication of massive data. It does not require real understanding, let alone rigorous reasoning. It is just advanced imitation of form.
The second layer is underlying logical intelligence, which corresponds to mathematics, formal proof, and multi-step deductive reasoning.
Mathematics is the only unambiguous, self-consistent, strictly verifiable, absolutely binary right-wrong symbolic system for human beings, with no exceptions, no ambiguity, no luck.
Back then, researchers' core judgment was very firm.
Language can only pile up the shell of AI, while mathematics is the core that tests true thinking.
An AI that cannot even complete high-level mathematical proof and rigorous logical deduction, no matter how smooth its conversation is, is just a product of statistical fitting.
It is by no means general intelligence.
This gave birth to the core route that ruled the industry for many years.
The essence of AGI is extreme logical self-consistency, and the ultimate form of logic is mathematics.
Therefore, conquering mathematics means unlocking AGI. This was the underlying belief behind Google's heavy investment in AlphaProof's geometric reasoning and formal verification, and also the fundamental reason why domestic big tech rushed to set up specialized mathematical teams for Olympiad problem solving and mathematical conjecture proof.
At that time, the whole industry reached a consensus that whoever cracked AI mathematics first would get the admission ticket to AGI.
II. Belief collapses: Big tech are not unable to keep up, the route has been falsified.
Today's team disbandment, budget cuts and brain drain,
are generally interpreted by the outside world as basic research burns too much money and returns too slowly, so big tech can't hold on anymore.
This is a typical superficial misunderstanding.
The real reason is that after years of large-scale trial and error, the industry has completely falsified the old AGI axiom.
Old axiom: Rigorous mathematical reasoning is the prerequisite threshold of AGI. Only after mastering the complete formal mathematical system can general reasoning ability be born.
New consensus: The current mainstream view in the industry is that mathematics is the capability benchmark and upper limit proof of top-level intelligence, but it is by no means the necessary starting point of AGI.
A difference in this core concept leads to a completely changed route.
Plain explanation: It used to be believed that AGI could never be achieved without understanding advanced mathematics. Now it has been proven that AGI can first have general world reasoning ability, and then fill its mathematical ability through tool iteration.
Google cutting its specialized mathematics team is not giving up logical reasoning, but giving up the inefficient route of forcing through mathematics from scratch to incubate AGI.
This is a complete paradigm downgrade and a return to engineering realism.
III. Core Analysis.
The mathematical ascension route is inherently unworkable. All top big tech changing routes collectively is not an accident, it is determined by three underlying hard rules.
1. The statistical intelligence of LLMs and deductive mathematical intelligence are essentially mutually exclusive: The core capability of current large models comes from statistical generalization of text data, learning human problem-solving steps, reasoning paradigms and logical structures, and replicating thinking processes.
The core of pure mathematics is axiomatic deduction and absolute necessity. Every step of derivation can be traced back to the underlying axioms, with 0 probability of error, 0 fault tolerance, and pure endogenous logic.
This leads to a key truth: AI can get full marks in IMO and complete complex mathematical proofs, but it just replicates human reasoning templates and does not establish cognition of underlying axioms. It consumes massive computing power and manpower to train the model to imitate mathematicians, and can only get simulation ability, cannot generate endogenous general intelligence, with extremely low input-output ratio and no AGI incubation value.
2. Mathematics is only a subset of intelligence, not the mother of intelligence.
The biggest cognitive mistake in the past was equating mathematical logic with general intelligence.
Real AGI needs to adapt to the complex real world: common sense causal judgment, physical world perception, dynamic task planning, fuzzy information decision-making, social logic interaction, error correction in unexpected scenarios. Rigorous mathematical reasoning is just a very small deterministic branch among all these capabilities.
An AI that can crack top mathematical conjectures may still be unable to handle daily fuzzy scenarios and complete complex real-world planning.
Conversely, an AGI with complete world modeling and general causal reasoning can make up for its mathematical shortcomings at any time through external tools.
Focusing solely on the closed mathematical field can never incubate open-world general intelligence, which is a typical case of putting the cart before the horse.
3. The scientific research cycle and business cycle are completely mismatched.
Formal proof at the frontier of AI mathematics belongs to long-term basic scientific research of more than ten years, which requires a long-term pure exploration environment without commercial pressure.
But today's AI industry is a battlefield of quarterly iteration, high-speed involution, and strong commercial orientation. The hard requirements of capital market, corporate revenue and product landing do not allow teams to invest in long-term basic exploration with no short-term returns.
The disbandment of independent specialized mathematics teams is not a technical failure, but an inevitable choice of engineering cost performance and commercial reality.
IV. Key Clarification.
We are not giving up mathematics, we just no longer deify it alone. There is a lot of misreading online: big tech give up mathematics, give up rigorous logic, and AI completely goes to metaphysical statistics.
This is completely wrong. The logic of the industry's actual adjustment is very clear: eliminate independent specialized teams, retain and internalize core capabilities.
1. Google has not stopped the iteration of mathematical reasoning. Gemini's symbolic computing, logical deduction and mathematical proof capabilities are continuously upgraded, it just no longer sets up an independent long-term research team.
2. OpenAI and Anthropic are actually continuing to increase investment in high-order reasoning, and mathematical logic problems are still the core hard indicators for evaluating the launch of top models.
3. The core competitiveness of all leading models is still reasoning depth, logical self-consistency and conclusion rigor.
The industry has completed the old-new replacement of two routes: the old route of incubating general AGI by focusing on closed mathematical logic has been eliminated. The new route, the only current mainstream, is to build general world modeling and general reasoning capabilities, and fill in mathematical rigor through external tools and model iteration.
Simple summary: Mathematics has changed from a prerequisite ticket for AGI to an upper-limit bonus item for top intelligence.
V. Ultimate Big Picture.
This shift defines the form of AI for the next decade. This low-key strategic transformation is a milestone for AI from idealistic scientific research fantasy to realistic engineering landing, and directly confirms the three major characteristics of future AGI.
1. The birth form of AGI is a general decision-maker for the world, not a full-time mathematician.
The core advantage of future general AI is adapting to the vague, complex and uncertain real physical world, rather than the extremely rigorous abstract mathematical world. Landing capability, decision-making capability and interaction capability have replaced pure mathematical capability as the core evaluation standard.
2. The rigor of future AI comes from the hybrid architecture of model reasoning plus tool verification, and it no longer pursues the model's innate built-in perfect axiom system.
Large models are responsible for thinking decomposition, causal reasoning and task planning, while calculators, symbol libraries, provers and code scripts are responsible for accurate error correction, numerical calculation and logical support, balancing intelligent generalization and logical rigor at the lowest computing power cost.
3. Basic mathematical research returns to academia, and the industry fully focuses on landing: pure theoretical, pure mathematical AI exploration has completely withdrawn from the mainstream of the industry, and is handed over to universities and laboratories for long-term research.
All resources on the industrial side are fully tilted to Agent, real-world interaction, scenario landing and engineering efficiency optimization.
Years ago, the whole industry believed that extremely self-consistent mathematical logic was the only divine path for AI to superintelligence.
After years of trial and error, the truth finally comes: Mathematics is the highest-precision measuring ruler of intelligence, but it is not the soil where general intelligence is born.
The true meaning of AGI has never been to master formulas and proofs, but to have stable, general, iterable independent thinking and decision-making capabilities in the complicated real world.
Big tech collectively giving up the fantasy of mathematical ascension is not regression, but sobriety.
AI is no longer pursuing defining intelligence with extreme theoretical perfection, but starting to truly realize intelligence with capabilities in the real world.