技术发展速度飞快，转眼间，星辰延伸成星线，我们今天所处的位置与几天前相去甚远。越来越难以预测明天我们会身在何处。
有一点是明确的：我们正在进入通用人工智能(AGI) 领域，超级人工智能(ASI) 现在似乎触手可及。无论如何定义，AGI 不会突然出现；它会不断发展，我们已经看到了它逐渐展开的迹象。
来源：科技世代千高原，图片来源：网络
AGI的曙光
AGI 一直以来都是我们的终极目标——一种能够代替人类进行脑力劳动的技术，改变我们的工作、生活和思维方式。现在，随着我们步入 2025 年，AGI 的曙光已经初现，并有望随着时间的推移变得更加强大。
这是一个意义深远的转变，OpenAI 的 Sam Altman 和前 OpenAI 首席科学家、拥有自己的专注于 ASI 的初创公司的 Ilya Sutskever 等人相信，这将决定人类进步的方向。
2024 年 9 月，奥特曼发表了《智能时代》，这份宣言认为 AGI 不仅仅是一种工具，而是人类历史的一个新阶段。
从那时起，OpenAI 发布了越来越强大的推理模型——人工智能系统不仅能从涵盖全球大量书面文本的知识库中回答问题，还能思考和解决复杂问题。这一进步的影响尚未深入公众意识。但它们的影响是深远的。
例如，OpenAI 的 GPT-o1 模型在国际数学奥林匹克 (IMO) 资格考试中获得了 83% 的分数，该考试被广泛认为是世界上最困难的数学竞赛之一，需要创造力和深度推理能力才能在没有微积分等高级数学工具的情况下解决问题。
随后，GPT-o3 模型在 ARC-AGI 基准测试中取得了突破性的 87.5% 的成绩，该测试评估人工智能在不依赖预先训练的知识的情况下解决全新问题的能力。ARC-AGI 被认为是最严格的人工智能基准测试之一，因为它测试概念推理和自适应智能，而这些领域传统上由人类主导。
从狭义智能到通用能力
到目前为止，人工智能系统在专业领域表现优异——撰写文案、诊断疾病、优化物流——但仅限于狭义的范围内。AGI 承诺提供完全不同的能力：跨领域适应、推理和解决问题的能力。
大型语言模型 (LLM) 和多模态模型已经展现出 AGI 的原始特征，例如跨任务泛化、多模态推理和适应性。这些功能正在通过更好的架构、更大的数据集和更高效的训练方法不断改进。
与此同时，OpenAI 正在重新定义 AGI 的含义。它的公开定义仍然是“一个高度自主的系统，在大多数具有经济价值的工作上表现优于人类。”但这个终点已经变得如此模糊，据报道，微软和 OpenAI 将 AGI 与人工智能系统创造 1000 亿美元利润的能力联系起来。
AGI 挑战了我们对人类的理解。长期以来被视为人类特征的智力将不再是人类独有的。我们如何将 AGI 融入我们的生活（无论是作为工具、合作伙伴还是对手）将以前所未有的方式塑造我们的文化、价值观和身份。
超级智能
这也使我们走上了 ASI 的道路，自学习的 AGI 系统最终将超越人类的集体智能。
如今，特定领域的人工智能系统在科学、编程或医学等领域展现出超越人类的狭隘智能。例如，AlphaFold 通过以无与伦比的准确性预测蛋白质结构，彻底改变了结构生物学——这是一项超出人类能力的任务。
OpenAI 的推理模型包括一个递归循环，可在推理过程中优化其输出。虽然这种优化是暂时的，不会改变模型的底层参数，但它展示了更具动态性和适应性的 AI 系统的潜力。
研究人员正在努力探索增量学习和基于重放的方法等技术，以使人工智能系统能够在获取新知识的同时保留知识，从而使单个系统能够持续学习。
目标是雄心勃勃的：创造不仅能思考还能进化的机器。如果这些努力成功，其影响将是惊人的。
人机协作的新时代
去年 12 月，苏茨克维尔表示：“我们即将创造出一些工具，它们不仅仅是人类能力的延伸，而且在某些领域，它们的能力将超越我们自己。”他设想了一个人工智能能够带来科学突破、治愈疾病和解决以前被认为难以解决的问题的世界。他认为，这些进步可能预示着人类繁荣的新时代——这场复兴不仅由人类的智慧推动，还由与机器的合作推动。
由推理模型驱动的人工智能代理可以驾驭复杂的环境，整合不同的数据流，并解决曾经看似无法克服的问题。
在医疗保健领域，这可能意味着 AGI 系统不仅会标记潜在诊断，还会根据个人的基因构成设计整个治疗计划。在教育领域，虚拟导师可以实时适应学生的需求，不仅可以教授任何科目，还可以教授任何语言，以任何速度授课。这不是一个遥不可及的梦想——Altman 认为这种进步可以在“几千天”内实现。
如果有一天机器能够不断学习并无缝适应新挑战，那么它们走向超级智能就指日可待了。
目前，有一件事是肯定的：2025 年标志着一个新时代的开始。智能时代已经到来，随之而来的是人类可能面临的变革性——也是充满危机的——未来。
AGI 的出现不会是突然事件。随着人工智能系统从通用智能向 ASI 转变，它将逐渐展开。真正的问题不是 AGI 何时会出现，而是我们是否准备好引导其更好地发展。
Entering The Artificial General Intelligence Spectrum In 2025
Craig S. Smith Contributo
Craig S. Smith, Eye on AI host and former NYT writer, covers AI.
starlines at warp speed
getty
Technological development has hit warp speed – in a flash, stars have stretched into starlines and where we are today is far from where we were just days ago. It’s increasingly difficult to predict where we will be tomorrow.
One thing is clear: we are entering the Artificial General Intelligence (AGI) spectrum and Artificial Superintelligence (ASI) now seems clearly within reach. However it is defined, AGI will not appear suddenly; it will evolve and already we see signs of its incremental unfolding.
The Dawn of AGI
AGI has long been the ultimate goal—a technology capable of performing the mental work of humans, transforming how we work, live, think. Now, as we step into 2025, glimmers of AGI are already appearing and promise to grow stronger as the year moves along.
This is a shift so profound that some, like Sam Altman of OpenAI and Ilya Sutskever, formerly OpenAI’s chief scientist with his own startup focused on ASI, believe it will define the arc of human progress.
Google Starts Tracking All Your Devices In 6 Weeks—Forget Chrome And Android
California Wildfire Live Updates: Death Toll Rises To 10 As Police Arrest Arson Suspect Near Kenneth Fire
Samsung Updates Galaxy S24, S23, S22, S21—But S25 Will Be Different
In September 2024, Altman published The Intelligence Age, a manifesto arguing that AGI isn’t just a tool, it’s a new phase in human history.
Since then, OpenAI has released increasingly powerful reasoning models – AI systems that not only answer questions from a base of knowledge encompassing much of the world’s written text but can think and solve complex problems. The implications of this advance have not yet penetrated the public consciousness. But they are profound.
For example, OpenAI’s GPT-o1 model scored 83% on the International Mathematical Olympiad (IMO) qualifying exam, widely regarded as one of the most difficult math competitions in the world, requiring creativity and deep reasoning skills to solve problems without advanced mathematical tools like calculus.
Subsequently the GPT-o3 model achieved a groundbreaking score of 87.5% on the ARC-AGI benchmark, which evaluates an AI's ability to solve entirely novel problems without relying on pre-trained knowledge. ARC-AGI is considered one of the toughest AI benchmarks because it tests conceptual reasoning and adaptive intelligence, areas traditionally dominated by humans.
From Narrow Intelligence to General Capability
Until now, AI systems have excelled as specialists—writing copy, diagnosing diseases, optimizing logistics—but only within narrowly defined limits. AGI promises something fundamentally different: the ability to adapt, reason, and solve problems across domains.
Large language models (LLMs) and multimodal models are already demonstrating proto-AGI traits such as generalization across tasks, multimodal reasoning, and adaptability. These capabilities are improving iteratively through better architectures, larger datasets, and more efficient training methods.
Meanwhile, OpenAI is redefining what AGI means. Its public definition remains “a highly autonomous system that outperforms humans at most economically valuable work.” But that endpoint has grown so blurry, Microsoft and OpenAI are reportedly linking AGI to the ability of an AI system to generate $100 billion in profits.
AGI challenges our very understanding of what it means to be human. Intelligence, long regarded as humanity’s defining trait, will no longer be ours alone. How we integrate AGI into our lives—whether as tools, partners, or rivals—will shape our culture, values, and identity in ways no one has yet to grasp.
Superintelligence
It also puts us on the road to ASI, when self-learning AGI systems eventually surpass collective human intelligence.
Domain-specific AI systems exhibit superhuman narrow intelligence today within fields like science, programming, or medicine. AlphaFold, for example, has revolutionized structural biology by predicting protein structures with unparalleled accuracy – a task beyond human capability.
OpenAI’s reasoning models include a recursive loop that refines their outputs during inference. While this refinement is temporary and does not change the model’s underlying parameters, it demonstrates the potential for more dynamic and adaptive AI systems.
Researchers are diligently exploring techniques like incremental learning and replay-based approaches to enable AI systems to retain knowledge while acquiring new knowledge, allowing a single system to learn continuously.
The goal is ambitious: to create machines that not only think but evolve. If these efforts succeed, the implications are staggering.
A New Era of Human-Machine Collaboration
“We are on the cusp of creating tools that are not merely extensions of human ability but entities with capabilities that, in some domains, will exceed our own,” Sutskever said this past December. He envisions a world where AI can unlock scientific breakthroughs, cure diseases, and solve problems previously thought intractable. Such advancements, he argued, could herald a new era of human flourishing—a Renaissance driven not by human ingenuity alone but by a partnership with machines.
AI agents powered by reasoning models could navigate complex environments, integrate disparate data streams, and solve problems that once seemed insurmountable.
In healthcare, this could mean AGI systems that wouldn’t just flag potential diagnoses, but design entire treatment plans tailored to an individual’s genetic makeup. In education, virtual tutors could adapt in real time to a student’s needs, teaching not only any subject but in any language, at any pace. This isn’t a distant dream—it’s the kind of progress that Altman suggests could materialize within “a few thousand days.”
And if machines can one day learn continuously and adapt seamlessly to new challenges, their ascent to superintelligence cannot be far behind.
For now, one thing is certain: 2025 marks the beginning of a new epoch. The Intelligence Age is here, and with it comes the possibility of a future as transformative—and as fraught—as any humanity has ever faced.
The emergence of AGI won’t be a sudden event. It will be a gradual unfolding as AI systems move along a spectrum of general intelligence toward ASI. The real question isn't when AGI will emerge, but whether we are prepared to guide its development for the better.
来源：科技世代千高原
来源：科技世代千高原，图片来源：网络
