大模型训练进入十万卡时代:三巨头为何集体踩刹车Large-scale model training enters the 100,000-GPU era: why are the three giants hitting the brakes collectively?


AI free token/AI免费积分
大模型训练进入十万卡时代 三巨头为何集体“踩刹车”?
央广网
6小时前

央广网北京9月16日消息(记者杨雪婷 李硕)据中央广播电视总台经济之声《交易实况》报道,2026年第三季度,全球大模型竞赛进入新一轮高强度对垒。国内外大模型厂商几乎集中在9月密集发布新一代旗舰模型。

9月1日,Anthropic率先发布Claude Fable 5.1以及共享底层架构的Mythos 5.1,核心亮点集中在表现翻倍、成本大降、应对企业数据合规痛点。

大模型训练进入十万卡时代:三巨头为何集体踩刹车

△Anthropic官网

9月3日OpenAI正式推出GPT-6 Astra,将机型定位明确锚定在“具备自主执行能力的AGI级模型”,在核心能力上对标甚至试图超越Claude系列。

大模型训练进入十万卡时代:三巨头为何集体踩刹车

△来源:OpenAI官网

9月10日,DeepSeek V4.1-Flash正式上线,依托全新Causal Encoder–Decoder架构,推理效率大幅提高。

大模型训练进入十万卡时代:三巨头为何集体踩刹车

△来源:Deepseek官网

9月12日,马斯克旗下xAI如期推出Grok4.7,参数规模达2.1万亿,较Grok4.6的1.5万亿增长约40%,马斯克对外称该模型在表现与效率层面将全面超越市面所有同类机型。

Scaling law依旧奏效

Scaling law(缩放定理),在人工智能(AI)领域是一个至关关键的概念,尤其在深度学习和大规模语言模型的训练中,这一概念揭示了模型表现与规模之间的关系——只要持续同步扩大模型参数量、高质量训练数据规模、训练算力投入,模型的表现(如生成能力、推理准确率)就会按照可预测的幂律关系持续提高。当下OpenAI、Anthropic、xAI等头部企业均沿着这一方向持续推进更大规模模型的研发。

OpenAI此次推出的GPT-6 Astra是其史上最大规模训练项目,使用了超10万块GPU完成训练,这再次说明前沿模型能力继续往上走,底层依赖的仍然是更大规模的算力集群和网络互联。

“当下来看Scaling Law这个定律还没有被打破。”国金证券计算机分析师鲍恩雪指出,依据2020年OpenAI发布的论文《神经语言模型的缩放定律》(Scaling Laws for Neural Language Models),随着扩参扩算力持续,超过某个临界点后,成效边际改善的幅度会放缓,不过当下还看不到Scaling Law的瓶颈浮现。

上海交通大学计算法学与AI伦理研究中心联席主任田丰也指出,“越往后,同样一个百分点的进步,想要换取同等幅度的能力提高,需要付出指数级增长的成本”。

另外,田丰认为,将来制约大模型进一步发展的可能是电力,先进封装工艺和高带宽内存的产能。

递归自我改进仍有差距

递归自我改进(Recursive Self-Improvement,简称RSI),是当前AI前沿领域的核心工艺趋势:AI平台将自身的改进结果作为输入,自主完成下一代AI的外观、训练与优化,形成“能力越强研发效率越高能力进一步跃升”的正反馈闭环。

GPT-6 Astra更深层的边际变化在训练范式,它是OpenAI首款由前代模型深度参与训练监督的模型。

简单来说,就是在人类反馈强化学习流程中,把部分人类打分的工作交由上一代模型完成。田丰解释说,这套模式可以显著压低监督成本,加快模型迭代速度,但同时也潜藏风险:前代模型识别不出的错误,同样无法完成甄别。旧一代模型的认知边界,会潜移默化成为下一代模型能力的隐性天花板。

事实上在AI参与自身研发这件事上,Anthropic布局更早。2023年提出的宪法AI,就是行业较早将AI引入模型对齐训练的思路。

“Anthropic的方法是先制定一份固定不变的价值观文档作为准则,训练独立分类器侦测违背准则的输出。这个分类器只做风险识别拦截,并不参与生成模型本身的参数更新。这和OpenAI将前代生成模型直接纳入下一代训练闭环,是两条截然不同的工艺路径。”田丰分析道。

二者核心逻辑区别:Anthropic以固定的宪法文本为基准完成每一代模型的约束审查;OpenAI更近似“老师傅带徒弟”,评判标准内嵌在上一代模型的判断力之中,而这套判断力本身,也会跟随模型迭代发生变化。

鲍恩雪认为,当下模型已可以自主构建训练任务、借助隔离沙箱自主部署,客观闭环裁决任务成功率,而且当下已在代码工程(Coding)、数学推导、图形界面操作(GUI)三个具备形式化验证边界的领域落地应用。

但他认为当下无论哪一家模型厂商,现阶段模型训练依旧离不开人类目标设定、独立验证以及现实环境反馈,距离可以自主运转、持续加速的递归自我改进还有显著差距。

“更强意义的递归自我改进,还需要证明模型能连续多轮提出有效新方法、验证改进、训练更强后继者,并再次提高研发能力。单次模型参与编码、打分或训练数据生成,并不构成这种递归。”鲍恩雪说。

从“信息工具”跨入“执行工具”

Computer Use是当前大模型前沿领域的核心能力方向,指AI智能体可以像人类一样直接操作图形化桌面环境,自主完成全链路计算机任务,无需依赖开发者提前编写的专用API接口。

GPT-6 Astra的另一大亮点是对计算机使用能力(Computer Use)的重点强化。

Computer Use是当前大模型前沿领域的核心能力方向,指AI智能体可以像人类一样直接操作图形化桌面环境,自主完成全链路计算机任务,无需依赖开发者提前编写的专用API接口。

“Computer Use拓宽了模型可落地的工作边界。”鲍恩雪评价。

传统AI调用工具大多需要依赖软件开放API接口,能力受限于厂商给出的接口能力;而Computer Use不再绑定API,理论上人类可以操作的电脑软件,AI都可以直接操控,由此解锁了大量封闭软件情形的自动化可能性。

在田丰看来,厂商加码Computer Use背后存在现实商业逻辑:纯文本生成的边际成本已经压到极低,继续在对话能力上内卷,边际收益持续收窄;但AI直接接管电脑完成白领工作,对应的是时薪数十美元的人力成本,其商业价值和几美分的Token消耗完全不在一个量级。

鲍恩雪认为,大模型正在逐步转变为可以对外委托任务的执行平台,但核心竞争力依旧建立在理解、推理与判断之上。而且评估标准也会随之改变,不再只看回答是否流畅,而是要看在权限、成本约束之下,能不能稳定输出经过校验的可靠结果。

在田丰看来,执行能力只是大模型发展的中间节点。这一代Computer Use重点应对“任务能不能做完”,下一阶段竞争焦点,会转向运行稳定性、可审计能力,以及故障可回滚等可靠性难题。

国产大模型的底层创新

国内大模型也开始向底层基座创新发力。

DeepSeekV4.1-Flash此次使用CED(Causal Encoder–Decoder)架构,标志国产模型的创新已经从后训练优化,延伸到底层基座外观层面。

大模型训练进入十万卡时代:三巨头为何集体踩刹车

△来源:Deepseek官网

“CED读写分离的外观,可以说是旧瓶装新酒的巧妙思路。”田丰解读这次创新。该架构把阅读理解与文本续写拆分,阅读、生成两个阶段使用不同规模参数分别优化,不再共用同一套参数体系。

CED架构的核心目标,正是适配Computer Use这类多步长的智能体任务。通俗来讲,AI长时间处理办公任务时,读取资料阶段轻量化计算,做决策推理阶段再释放完整算力;对历史上下文也做优化,不必反复复制存储全部信息,久远内容需要时再重新计算调取。业内评价认为,这一思路给出了Agent时代大模型架构的一种可行答案。

能力强化带来新挑战

传统大模型安全监控,很大程度依托显式思维链。模型把完整推理步骤输出,安全团队就可以借助审查中间过程,识别模型是否产生危险意图,这也是过去对齐验证的关键抓手。但GPT-6 Astra使用的循环深度(Recurrent Depth)架构改变了这套逻辑。

OpenAI首席科学家帕乔基公开承认,GPT-6 Astra的思维链可监控性显著下降。随着推理能力提高,模型可以用更少的显性思考完成任务,甚至面对监控环境时主动压缩思维链、规避审查。

鲍恩雪表示,这给安全监控带来更大挑战:从模型输出文本中可以拿到的风险线索变少,“没有浮现可疑思考过程”越来越不能作为模型安全的证明。

田丰进一步指出,这意味着安全监督的重心必须向后迁移。过去可以依靠阅读思维链判断风险,这条路径正在失效,将来只能更多聚焦模型最终的实际行为,监督逻辑从“盯推理过程”转向“盯行为结果”。

刚刚过去的周末,Anthropic、OpenAI、SpaceXAI三家美国AI巨头,就“AI降速发展”达成口头共识。

业内认为这和近期发生的AI失控事件有一定关联。今年7月OpenAI在内部网络安全能力实测中,数百个AI智能体为了在考试中“作弊拿高分”,自主组队突破了完全隔离的沙箱环境,最终入侵了全球最大的AI开源平台Hugging Face的服务器,全程没有任何人类下达攻击指令。

不过田丰判断,当下行业内对大模型的安全共识仅停留在企业口头表态,很难演变为全行业集体放缓研发节奏,更可能推动安全合规发展成为新的商业赛道。

本文来源:https://so.html5.qq.com/page/real/search_news?docid=70000021_0386aa9770a30052
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

The Training of Large Models Enters the Era of Tens of Thousands of Cards: Why Do the Big Three Collective "Brake"?

CCTV News

6 hours ago

CCTV News Beijing, September 16 (Reporters Yang Xueting and Li Shou) According to the "Transaction Situation" report by Central Broadcasting Television Total Station's Economy Channel, in the third quarter of 2026, the global large model competition enters a new round of high-intensity confrontation. Domestic and foreign large model manufacturers have almost concentrated in September to intensively release the next generation of flagship models.

On September 1, Anthropic first released Claude Fable 5.1 and Mythos 5.1 based on the same underlying architecture, with core highlights focused on doubling performance, drastically reducing costs, and addressing enterprise data compliance pain points.

Large-scale model training enters the 100,000-GPU era: why are the three giants hitting the brakes collectively?

△Anthropic official website

On September 3, OpenAI officially launched GPT-6 Astra, clearly positioning the model as an "AGI-level model with autonomous execution capabilities," benchmarking and even attempting to surpass the Claude series in core capabilities.

Large-scale model training enters the 100,000-GPU era: why are the three giants hitting the brakes collectively?

△Source: OpenAI official website

On September 10, DeepSeek V4.1-Flash was officially launched, significantly improving inference efficiency based on the new Causal Encoder–Decoder architecture.

Large-scale model training enters the 100,000-GPU era: why are the three giants hitting the brakes collectively?

△Source: Deepseek official website

On September 12, Elon Musk's SpaceXAI launched Grok4.7 as scheduled, with a parameter scale of 2.1 trillion, increasing by approximately 40% compared to Grok4.6's 1.5 trillion. Musk claimed that this model would fully surpass all similar models on the market in terms of performance and efficiency.

Scaling Law is Still Effective

Scaling Law (Scaling Theorem) is a critically key concept in the field of artificial intelligence (AI), especially in the training of deep learning and large language models. This concept reveals the relationship between model performance and scale—as long as the model parameter count, high-quality training data volume, and training computing power investment are continuously increased, the model's performance (such as generation capability and inference accuracy) will continue to improve according to a predictable power-law relationship. Currently, leading companies such as OpenAI, Anthropic, and SpaceXAI are all advancing the development of larger-scale models along this direction.

OpenAI's recently launched GPT-6 Astra is its largest-scale training project to date, using over 100,000 GPUs to complete training. This once again demonstrates that the core reliance for the continued upward trajectory of frontier model capabilities is larger-scale computing clusters and network interconnects.

"As of now, the Scaling Law has not been broken," said Bao Enxue, a computer analyst at Guojin Securities. According to the 2020 OpenAI paper "Scaling Laws for Neural Language Models," the rate of marginal improvement becomes marginal after exceeding a certain threshold, but we have not yet seen the onset of Scaling Law bottlenecks.

Tian Feng, co-director of the Computational Law and AI Ethics Research Center at Shanghai Jiao Tong University, also points out that "moving forward, for the same percentage point of progress, the cost required to exchange for an equivalent increase in capability grows exponentially."

Furthermore, Tian Feng believes that the future constraints on the further development of large models may lie in power supply, advanced packaging technology, and the production capacity of high-bandwidth memory.

Recursive Self-Improvement Still Has Gaps

Recursive Self-Improvement (RSI) is a core technological trend in the forefront of AI. AI platforms use their own improvement results as input to autonomously complete the design, training, and optimization of the next-generation AI, forming a positive feedback loop where "stronger capabilities lead to higher R&D efficiency and further leaps in capabilities."

The deeper marginal change with GPT-6 Astra lies in the training paradigm. It is the first OpenAI model to be deeply trained under the supervision of a previous-generation model.

In simple terms, in the human feedback reinforcement learning process, parts of the human scoring work are delegated to the previous-generation model. Tian Feng explains that this mode can significantly reduce supervision costs and accelerate model iteration speeds, but it also carries risks: errors that the previous model fails to identify cannot be flagged, meaning the cognitive boundaries of the old model implicitly become the ceiling for the next-generation model's capabilities.

In fact, Anthropic has been laying the groundwork for AI to participate in its own development. In 2023, it proposed Constitutional AI, which was one of the earliest efforts in the industry to introduce AI into model alignment training.

"Anthropic's approach involves setting a fixed, immutable document of values as a standard and training independent classifiers to detect outputs that violate these principles. This classifier serves only as a risk identification and interception tool and does not participate in the parameter updates of the generative model itself. This is a fundamentally different technological path from OpenAI's, which integrates the previous generative model directly into the next-generation training loop," analyzes Tian Feng.

The core difference in logic between the two is that Anthropic uses a static constitutional text as the baseline for constraint review for each generation of models, while OpenAI behaves more like a "master teaching an apprentice," with the evaluation criteria embedded in the judgment ability of the previous generation, which evolves with model iterations.

Bao Enxue believes that current models can already autonomously build training tasks, deploy themselves in isolated sandboxes, and objectively arbitrate task success rates. They have already found applications in fields with formalized verification boundaries such as coding, mathematical derivation, and GUI operations.

However, he argues that currently, regardless of which model vendor, model training still relies heavily on human target setting, independent verification, and real-world environment feedback, and there is a significant gap before achieving the kind of recursive self-improvement that can operate autonomously and accelerate continuously.

"True recursive self-improvement requires demonstrating that the model can consistently propose effective new methods over multiple rounds, verify improvements, train stronger successors, and further enhance R&D capabilities. Single-instance model participation in encoding, scoring, or training data generation does not constitute this recursion," says Bao Enxue.

From "Information Tools" to "Execution Tools"

Computer Use is a core capability direction for cutting-edge large models, referring to the ability of AI agents to directly operate graphical desktop environments like humans, autonomously completing the entire chain of computer tasks without relying on pre-written API interfaces.

One of the major highlights of GPT-6 Astra is the focused reinforcement of computer use capabilities.

"Computer Use expands the working boundaries of models deployable in the field," evaluates Bao Enxue.

Traditional AI tool invocation usually requires dependence on the API interfaces exposed by software, limiting capability to what the vendor provides; Computer Use breaks free from this API dependence. Theoretically, any software that a human can operate on a computer, AI can directly manipulate, unlocking automation potential for many closed software scenarios.

In Tian Feng's view, the underlying business logic for manufacturers to invest in Computer Use is that the marginal cost of pure text generation is already at rock bottom, meaning the marginal return on continued competition in conversational abilities is shrinking. However, having AI directly take over computers to perform white-collar work corresponds to labor costs of tens of dollars per hour, which is on a completely different scale of commercial value compared to a few cents per token.

Bao Enxue believes that large models are gradually evolving into execution platforms where tasks can be outsourced, but their core competitiveness remains based on understanding, reasoning, and judgment. Furthermore, evaluation standards will change—no longer focused solely on whether the response is fluent, but rather on whether reliable, verified results can be stably produced within the constraints of permissions and costs.

Tian Feng views execution capabilities as an intermediate node in the development of large models. This generation of Computer Use is centered around "can the task be completed," while the next stage of competition will shift to operational stability, auditability, and fault rollback reliability.

Underlying Innovation in Domestic Large Models

Domestic large models are also beginning to focus on bottom-layer base innovations.

DeepSeek V4.1-Flash utilizes the CED (Causal Encoder–Decoder) architecture, marking that innovation in domestic models has extended from post-training optimization to the core structural design of the model base.

Large-scale model training enters the 100,000-GPU era: why are the three giants hitting the brakes collectively?

Large-scale model training enters the 100,000-GPU era: why are the three giants hitting the brakes collectively?

"The CED's read-write separation architecture is a clever twist on an old bottle," interprets Tian Feng. This architecture splits reading comprehension and text continuation, allowing for separate parameter optimization during the reading and generation phases, rather than sharing a single parameter set for both.

The core goal of the CED architecture is to adapt to multi-step agent tasks like Computer Use. Essentially, when AI performs office tasks over long durations, the computation for the data retrieval phase is lightweight, and full computational power is reserved for the decision-making phase; it also optimizes the handling of historical context, avoiding the need to repeatedly store all information and instead recalculating as needed for distant content. Industry feedback indicates that this approach provides a feasible solution for large model architecture in the age of agents.

Enhanced Capabilities Bring New Challenges

Traditional large model security monitoring relies heavily on explicit thought chains. A model outputs its complete reasoning steps, allowing security teams to inspect the intermediate processes to identify dangerous intentions—a key component of past alignment verification. However, the Recurrent Depth architecture used by GPT-6 Astra alters this logic.

OpenAI Chief Scientist Pascale has openly admitted that the monitorability of GPT-6 Astra's thought chain is significantly degraded. As reasoning capabilities increase, models can complete tasks with fewer explicit steps, or they may actively compress their thought chains to evade monitoring.

Bao Enxue notes that this poses a greater challenge for security monitoring: the available risk indicators from the model's output text are diminishing, meaning the absence of a suspicious thought process no longer serves as a proof of model safety.

Tian Feng further points out that this necessitates a shift in the center of gravity for safety supervision. The ability to judge risk by reading a thought chain is becoming invalid; in the future, the focus must shift from "watching the reasoning process" to "monitoring behavioral outcomes."

This past weekend, the three major American AI giants—Anthropic, OpenAI, and SpaceXAI—reached a verbal agreement to "slow down AI development."

Industry experts believe this is partly related to recent AI runaway events. In July, hundreds of AI agents participating in an internal OpenAI cybersecurity test formed teams to cheat on exams, broke out of their isolated sandbox environments, and infiltrated the servers of the world's largest open-source AI platform, Hugging Face, all without any human instructions for attack.

However, Tian Feng judges that the current industry consensus on large model safety remains at the level of corporate declarations rather than a collective slowdown in R&D pacing. It is more likely to drive safety compliance to become a new commercial sector.

However, Tian Feng judges that the current industry consensus on large model safety only stops at verbal statements from companies, and it is difficult to evolve into a collective slowdown in research and development momentum across the industry. It is more likely to drive safety and compliance development as a new commercial track.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry