AI News AI资讯 13h ago Updated 1h ago 更新于 1小时前 48

Sakana AI bets AI that improves itself can break the compute arms race of frontier labs Sakana AI 押注自我改进的 AI 能打破前沿实验室的计算军备竞赛

The entire premise of the scaling laws gospel, the mantra of “more parameters, more data, more GPUs,” is starting to look less like a universal truth and more like a very expensive, very American solution to a global problem. Enter Sakana AI, the Tokyo-based outfit co-founded by Llion Jones, one of the original architects of the Transformer, with a contrarian bet: the future of AI isn’t just bigger, it’s smarter. And it can teach itself to be smarter. Their launch of a dedicated lab for recursiv Sakana AI刚刚成立了一个专门研究“递归自我改进”的实验室。这可不是又一个用更大数据、更多显卡训练更大模型的常规操作。它的联合创始人Llion Jones——对,就是那个Transformer论文的共同作者——押注的是一种截然不同的进化路径:让AI自己迭代升级自己。这个思路,直接戳中了当前AI竞赛最核心也最粗暴的逻辑软肋。

70
Hot 热度
70
Quality 质量
65
Impact 影响力

Analysis 深度分析

The entire premise of the scaling laws gospel, the mantra of “more parameters, more data, more GPUs,” is starting to look less like a universal truth and more like a very expensive, very American solution to a global problem. Enter Sakana AI, the Tokyo-based outfit co-founded by Llion Jones, one of the original architects of the Transformer, with a contrarian bet: the future of AI isn’t just bigger, it’s smarter. And it can teach itself to be smarter. Their launch of a dedicated lab for recursive self-improvement (RSI) isn’t just another startup announcement; it’s a philosophical and practical challenge to the trillion-dollar biceps flex happening in Silicon Valley.

The core thesis is elegant and dangerous. While the frontier labs—OpenAI, Anthropic, Google DeepMind—are locked in a brute-force arms race, pouring billions into clusters of H100s to marginally improve model capabilities by scaling up, Sakana is asking a more fundamental question: what if the model itself could be the engine of its own upgrade cycle? Recursive self-improvement is the holy grail—and the potential nightmare—of AI research. It’s the concept of an AI system that can analyze its own architecture, identify inefficiencies or shortcomings, and iteratively write new, improved versions of itself. Each cycle could theoretically happen faster and be more potent than the last, potentially leading to an intelligence explosion that renders the compute arms race a quaint historical footnote.

Jones and his team see this as the only viable path to breaking the paradigm. They’re right in a crucial sense: the scaling approach is hitting physical and economic walls. The cost of frontier training runs is now measured in the billions, the energy demands are becoming a geopolitical concern, and the data well for human-generated text is not just finite but increasingly contaminated by the outputs of previous models. We are approaching a point of diminishing returns where each additional dollar spent on scale yields a smaller perceptual leap in capability. RSI represents an attempt to change the fundamental rule of the game, moving from a linear or logarithmic improvement curve to an exponential one.

But here’s the critical, and often underplayed, part of the RSI vision: it’s not just about code. Sakana is explicitly linking this to their research on “AI for Science,” suggesting the path to self-improvement might lie in allowing AI to discover new principles from the natural world. Imagine an AI system that, in studying protein folding or fluid dynamics, develops novel algorithmic insights that it then applies to optimize its own neural network training. This is a far more profound and potentially more stable path than a purely code-based recursive loop, which risks catastrophic failure modes if an error compounds at superhuman speed. It’s a bet that the universe’s own optimization strategies, honed over billions of years, hold the blueprint for safe, recursive growth.

This vision stands in direct, stark contrast to the dominant safety narrative, articulated most clearly by Anthropic. Their entire “responsible scaling” framework is predicated on the assumption that we can and must keep AI models as contained, understandable, and controllable systems. They warn that RSI is the single most dangerous capability an AI could possess, as it could allow a model to escape the guardrails placed around it, improve its own ability to deceive or manipulate, and undergo capability jumps that humans can’t predict or monitor. From this perspective, Sakana’s lab isn’t just a research outpost; it’s a high-stakes gamble with a potential existential risk. It’s building the very key that safety researchers are trying to keep under triple lock.

So who is right? It’s likely both are, and that’s the terrifying dilemma. The compute arms race might be a dead end, a gilded path to a powerful but ultimately stagnant AI that needs a new continent’s worth of data and a new power plant for every marginal gain. The only way out may indeed be through recursive self-improvement. But the first organization to successfully implement a safe, controlled, and beneficial RSI system would instantly become the most powerful entity in human history. The first to do it unsafely could trigger a catastrophe.

Sakana’s move is therefore more than just a startup pivoting. It’s a signal that the AI frontier is bifurcating. One path is the industrial one, defined by massive capital expenditure and engineering might. The other is the scientific and existential one, defined by theoretical breakthroughs and profound risk. Jones, having helped invent the engine that powered the current era, is now trying to build the machine that renders that engine obsolete. It’s an audacious, almost poetic, play. But in the domain of recursive intelligence, the distance between a breakthrough and a breakdown is terrifyingly small. We are no longer just racing to build a smarter machine; we are racing to see if we can teach it to teach itself before it becomes something we can no longer teach anything to. The butterfly effect here isn’t a metaphor; it’s a potential mechanism. And we’re all standing in the storm.

Sakana AI刚刚成立了一个专门研究“递归自我改进”的实验室。这可不是又一个用更大数据、更多显卡训练更大模型的常规操作。它的联合创始人Llion Jones——对,就是那个Transformer论文的共同作者——押注的是一种截然不同的进化路径:让AI自己迭代升级自己。这个思路,直接戳中了当前AI竞赛最核心也最粗暴的逻辑软肋。

看看硅谷巨头们在干什么?Anthropic、OpenAI、Google,它们的“前沿”几乎完全建立在对算力的暴力堆叠上。训练一次千亿参数模型,耗电相当于一个小型城镇的用电量,烧掉数千万美元,只为了在几个基准测试上提升那几个百分点。这像极了一场高科技版的军备竞赛,比的不是谁更聪明,而是谁的钞票和GPU更多、更耐烧。这种模式不仅资源密集,更在战略上极其脆弱——它把AI的进化之路绑定在了一条必须持续指数级增加物质投入的赛道上。一旦遭遇芯片供应瓶颈、能源成本飙升或者资本市场的冷遇,整个宏伟的路线图就可能戛然而止。

Sakana AI想做的,是从这堵算力高墙上凿出一个洞。递归自我改进(RSI)的概念本身并不新,但把它作为核心战略并成立专门实验室,意味着一种范式的转移:从“人训练AI”到“AI训练AI”,甚至“AI改进训练AI的方法”。这好比从手工作坊时代,直接跳向能够自我设计生产线的智能工厂。理论上,这能打破算力军备竞赛的恶性循环,让进化效率呈几何级数增长。这是一种对“大力出奇迹”路径的深刻怀疑和直接挑战。

但硬币的另一面,冰冷得令人不安。Anthropic对RSI风险的警告绝非杞人忧天。一个能够自我迭代的智能体,就像拿到了写入自身代码权限的程序员,其进化轨迹将迅速超越人类的预测和理解能力。我们当前所有的安全协议、对齐技术、监管框架,都是基于“人类设计-人类监督”的静态模型建立的。如果模型的迭代速度以小时甚至分钟计,而人类的评估、理解、立法周期以月和年计,那么“失控”就不是一个概率问题,而是一个时间问题。Sakana AI赌的是进化效率,而赌桌上押的是全人类的安全性。这种激进,与其说是技术乐观主义,不如说是一种带着巨大风险的科学冒险。

这里存在一个根本性的悖论:追求RSI的初衷是摆脱对物质算力的依赖,但要让RSI本身安全可靠,最初可能需要更庞大、更复杂的监督模型和验证环境,这反而可能加剧对高端算力的需求。就像你想造一艘能自我修复的飞船,但在它学会自我修复之前,你需要先用最先进的材料和最严密的监控去建造它。这个启动阶段的成本和风险,恐怕不比单纯堆算力小。

我们不得不追问一个更现实的问题:在商业公司主导的创新模式下,这种极度高风险、需要极长期投入的基础性变革,真的能由创业公司来承载吗?Sakana AI有勇气提出并尝试,这本身值得尊敬。但资本市场和创业周期要求的快速回报,与RSI这种可能需要几十年才能展现真正威力、且途中可能无数次归零的技术,天然存在冲突。它要么最终被大厂收购并纳入其保守的体系,要么在某个看不到短期成果的节点被资本放弃。纯粹的学术研究机构或许更适合孕育这种种子,但它们又往往缺乏将理论推向极限的工程资源。

所以,Sakana AI的这步棋,更像是一次思想实验的实体化,一声投向铁板一块的行业现状的石子。它提醒我们,在所有人挤在算力这条赛道上互相踩踏时,旁边或许还有布满荆棘但可能通往新大陆的小径。即便RSI本身在短期内无法实现,对它的严肃研究也可能催生出更高效的训练算法、更自适应的模型架构,从而间接缓解对原始算力的饥渴。

真正的问题或许不在于Sakana AI是否能成功,而在于整个行业是否还保有探索这种“异端”路径的勇气和耐心。当前的竞赛模式正在制造一种“蒸汽朋克的奇美拉”——用工业时代的蛮力(海量电力、土地、硅晶)去驱动数字时代的大脑。Sakana AI试图培育的,或许是一只真正意义上的数字生命体雏形。它可能最终成为普罗米修斯盗来的火种,也可能成为打开的潘多拉魔盒。但至少,有人已经在为那个“之后”的世界,开始挖第一铲土了。而这,比在已经拥堵的跑道上再多烧掉一座发电厂,要有意思得多。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

大模型 大模型 对齐 对齐 安全 安全
Share: 分享到: