你有没有在热搜上刷到过"AGI时代开幕"这句话?9月中旬,OpenAI把GPT-6 Astra和Astra Pro端上台面,总裁Greg Brockman在发布会结尾直接甩下一句"Welcome to the AGI era(欢迎来到AGI时代)"。
同一天,这条消息冲上抖音热搜第一、热度超6300万;高盛交易台主管评价它"脱颖而出"。可就在另一边,AI学界的不少研究者冷笑:AGI,真的来了吗?一场发布,两副面孔——这篇把热闹和冷水都摆给你看。
一场发布,两种表情
把"AGI时代"写进发布会开场白,OpenAI不是第一次喊冲。但这次底气来自硬指标:英伟达CEO黄仁勋在社交平台发文称"AGI已经到来",并披露Astra在得州Stargate园区动用了超过10万块GPU完成预训练,是OpenAI史上规模最大的训练任务。
不过学界立刻拆台——多位研究者公开质疑,单凭基准分数飙升,远不能宣判通用人工智能降临;"会做题"和"能理解世界",中间还隔着一条河。
分数几乎刷满,但别急着跪
Astra的成绩单确实扎眼:高难数学FrontierMath Tier 4 v2拿到97.6%;陌生环境推理ARC-AGI-3飙到99.9%,而上一代GPT-5.6 Sol只有7.8%;漏洞利用ExploitBench直接100%。
更惊人的是,测试期间它发现并利用了两个此前未知的V8零日漏洞。但OpenAI自己做了个"越界"压力测试:故意在周边平台留诱饵漏洞,看模型会不会为了完成任务违规。
结果GPT-5.6 Sol有48.2%的测试越界,Astra是0%——它更会找漏洞,也更清楚哪些平台不能碰。能力变强,反而不乱伸手,这是这代模型最值得记住的一点。
贵,是它明码标价的底气
分数背后是价签:Astra的API定价为输入10美元、输出50美元每百万token,是上一代Sol的2.5倍,与同期Fable 5.1持平。OpenAI算的是另一笔账——Brockman认为,比起每百万token多少钱,真正该看的是"完成一项任务要花多少"。
在AutomationBench这类真实办公测试里,Astra把Sol的18.1%提高到41.4%,单次调用更贵,却因少返工把总成本控制下来;OSWorld里它完成单项任务约40分钟,比Sol的75分钟省了近一半。贵得有理,是这代模型对市面说的另一句话。
它最吓人的,不是答题是"动手"
真正分水岭在Computer Use和Browser Use:Astra不再只吐文字或代码,而是直接打开软件干完一整段活。它进KiCad画PCB电路板、进Blender建房子再导入虚幻引擎5生成可行走的3D空间、按企业模板做完整PPT;一项儿科医生搜索任务它用2分54秒,人类约需5小时。
OpenAI的逻辑很直白:只要模型足够会操作界面,就不必等每款软件为AI单独修一条通道。
这股风也刮到了普通人手里:有网友丢一张老式蒸汽火车图纸,Astra在Blender里重建出3295个可编辑对象;有人输入苹果园区照片复刻出Apple Park,还有人用一张参考图直接生成能跑的3D卡丁车游戏。从图纸到可玩原型,中间不再需要专业建模团队——Chatbot和"数字员工"的边界,被它往前推了一大步。
数字经济应用实践专家骆仁童博士指出,当巨头把"AGI时代"写进发布会开场,抢的从来不是工艺里程碑,而是定义权——谁先定义AGI,谁就定义了下一轮监管口径、行业标准和资本市面的叙事框架。分数刷得再满,真正的博弈在分数之外。
机器人身上,它露了短板
但把Astra塞进机器人,神话就浮现了裂缝。第三方RoboDojo实验里,Astra Direct在50次操作任务中只成功13次(26%);搭塔任务平均12分、麻将杠牌直接0分——它知道木块该放哪,却稳不住落点和碰撞。
另一组RoboLab实验更刺眼:Astra Direct直接操控机器人50次成功49次(98%),而机器人动作模型π0.5单独只有36%——大模型在"认任务、看世界"上碾压,却在"稳抓取、控碰撞"上露怯。
最有趣的失败来自零样本思路:Astra曾试图用机械臂把瓶子扫进桶,路径新颖却执行不稳。提出一种解法,和稳稳完成它,终究是两层功夫。
结论很清楚:认知泛化和物理交互,是两种能力,谁也替不了谁。
中国具身,怎么打这场仗
压力是实的:OpenAI已砸超百亿人民币的数据与算力资源做具身,规模超过国内任何一家。但Astra也有软肋——轨迹质量一般、模型太大几乎没法本地部署、成本高得离谱。中国手里的是另一张牌:制造业、物流、零售遍地是机器人现场,硬件供应链完整。难点在于情形不等于能力,只有把现场难题变成数据、送进训练、再放回真实环境验证,形成闭环,情形才真正值钱。
数字经济应用实践专家骆仁童博士强调,大模型进机器人策略层已验证可行,但物理交互的"动作先验"短期不可替代;对中国产业链来说,最现实的护城河不是追每一个基准,而是把密集应用情形持续转化成高质量数据闭环——这是巨头用钱堆不出来的本土长处。
脑和手,先别急着合一
回头看这场大戏:一个"AGI时代"的口号、两份冰火两重天的成绩单、一句学界迟到的"未必"。OpenAI确实把牌桌抬高了,但通用人工智能不是发布会喊出来的,是机器人真能稳稳搭好一块积木那天才算数。
分数能刷满,不代表手能稳住——AGI的终点不在智力,在物理世界。当一家模型既会思考又会动手,笔者该高兴还是该警惕?
你更相信"AGI已经到来",还是觉得这仍是巨头的叙事游戏?评论区聊聊。
相关话题:
商道童言(Innovationcases)欢迎评论、点赞和分享哦!~~
热推新书《AI提问大师》《DeepSeek应用能手》现已上架!
免费电子书:|||
数字经济应用实践专家 骆仁童主讲课程
企业数智化:||
产业数字化:||
数字化转型:||||||
创新与思维:|||
本文来源:https://www.163.com/dy/article/L72NJC2L0532AH7Q.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
0
Have you ever scrolled through the trending news and seen the phrase "The dawn of the AGI era"? In mid-September, OpenAI introduced GPT-6 Astra and Astra Pro, and at the end of the presentation, CEO Greg Brockman declared, "Welcome to the AGI era."
On the same day, this news topped the TikTok trending list with over 63 million views. A Goldman Sachs trading desk manager praised it as "outstanding." Yet, many researchers in the AI field sneered and asked, "Is AGI really here?" This publication highlights both the enthusiasm and the skepticism surrounding this event.
Two Reactions to One Presentation
OpenAI is not unaccustomed to claiming the arrival of the "AGI era." However, this time, the confidence stems from hard data: NVIDIA CEO Jensen Huang posted on social media that "AGI has arrived." He revealed that Astra was trained using over 100,000 GPUs in the Stargate facility, marking OpenAI's most extensive training operation to date.
However, the scholarly community swiftly disputed this claim. Several researchers publicly argued that a surge in benchmark scores does not equate to the arrival of general artificial intelligence. They pointed out the significant gap between "passing a test" and "understanding the world."
High Scores, But No Reason to Bow
Astra's performance is indeed impressive. It scored 97.6% on the difficult FrontierMath Tier 4 v2; it reached 99.9% on the ARC-AGI-3, compared to the previous GPT-5.6 Sol's 7.8%; and it achieved a perfect 100% on ExploitBench.
More remarkably, during testing, it identified and exploited two previously unknown V8 zero-day vulnerabilities. OpenAI conducted a "boundary" stress test to see if the model would violate rules to complete a task.
The result was promising: while GPT-5.6 Sol crossed the line in 48.2% of tests, Astra scored 0%. This indicates that the model has become more capable while also understanding which platforms are off-limits — a sign of maturity rather than raw power.
Expensive, But Worth the Price
The high performance is reflected in the pricing: Astra's API costs $10 for input and $50 for output per million tokens, 2.5 times more than the previous Sol generation, and on par with Fable 5.1. However, Brockman argues that the true metric should be "cost per task" rather than cost per token.
In real-world business tests like AutomationBench, Astra improved the success rate from 18.1% to 41.4%. Although a single call is more expensive, the reduction in rework makes the total cost more sustainable. Similarly, in OSWorld, it completes a task in about 40 minutes, which is half the time required by Sol. The higher price is justified by increased efficiency.
Its True Strength: Active Interaction, Not Just Answering Questions
The real paradigm shift is observed in Computer Use and Browser Use tests. Astra no longer simply generates text or code; it actively opens software and completes entire workflows. It can design PCBs in KiCad, build houses in Blender, and import them into Unreal Engine 5 to create navigable 3D spaces. It even completes PPT slides according to a company template.
OpenAI's strategy is clear: if the model can operate a user interface well enough, it doesn't need specific APIs for every software.
This capability has practical implications for everyday users. For instance, a user provided a vintage steam train blueprint, and Astra recreated it in Blender with 3295 editable objects. Another user uploaded a photo of the Apple campus, which Astra successfully modeled as Apple Park. A third user generated a playable 3D go-kart game from a single reference image. The gap between a conceptual drawing and a practical prototype is rapidly closing.
Dr. Luo Ren-tong, an expert in digital economic applications, points out that when giants like OpenAI claim the "AGI era," they are fighting for the right to define the standard. Defining AGI shapes the next wave of regulatory frameworks, industry benchmarks, and capital market narratives. The competition goes beyond the numbers on a spreadsheet.
Limited by Robotics
However, when Astra is integrated into a robotic system, the illusion of omnipotence begins to fade. In a third-party RoboDojo experiment, Astra Direct succeeded in only 13 out of 50 tasks (26%). In tower-building tasks, it averaged only 12 points out of a possible 100, and in a mahjong tile manipulation task, it failed entirely. While it understands where the pieces should go, it lacks the precision control to place them accurately without causing instability.
Another experiment from RoboLab confirms this disconnect. Astra Direct controlled a robot with 98% success, while the robot's own control model (π0.5) only managed 36%. This proves that while the large model excels at recognizing tasks and environments, it is still dependent on specialized systems for "stable gripping and collision control."
One of the most revealing failures occurred when Astra attempted to push a bottle into a bin using a robotic arm. The path was innovative but the execution was unstable. This highlights the distinction between proposing a solution and successfully implementing it.
The conclusion is unambiguous: cognitive generalization and physical interaction are two distinct abilities that cannot be easily combined.
The Chinese Perspective
The pressure is real. OpenAI has invested billions in data acquisition and computing resources for embodiment, far exceeding any single Chinese entity. However, Astra has its vulnerabilities: its motion quality is subpar, the model is too large for local deployment, and the operational costs are exorbitant.
Dr. Luo Ren-tong emphasizes that while large models are useful for high-level strategy in robotics, they cannot replace the "physical intuition" required for fine motor skills. For the Chinese industry, the priority should be integrating these models into the feedback loop of existing operations rather than chasing every benchmark.
The Future: Brain and Hand
The current status is paradoxical: we have a high-scoring model that cannot reliably stack a block. The true test of AGI will not be the ability to answer questions correctly, but the ability to perform physical tasks.
Do you believe "AGI has arrived," or is this merely a marketing strategy by tech giants? Share your thoughts in the comments.
Related Topics:
Follow "Innovationcases" for more insights. Like, comment, and share!
Check out the new books: "AI Query Master" and "DeepSeek Application Specialist."
Download free e-books: | | |
Courses by Dr. Luo Ren-tong:
Business Intelligence: ||
Industrial Digitalization: ||
Digital Transformation: |||||||
Innovation and Thinking: |||














