东京 —— 日本人工智能初创公司 Sakana AI 周一宣布，将从零开始开发自有的国内模型，旨在降低在国防和网络安全等领域依赖美国和中国模型所带来的风险。
TOKYO -- Japanese artificial intelligence startup Sakana AI will develop its own domestic models from scratch, the company announced on Monday, aiming to reduce the risks associated with relying on U.S. and Chinese models in such areas as defense and cybersecurity.
首席执行官 David Ha 在此间举行的一个活动上表示，公司即将启动一项从头预训练模型的工作。
The company is about to launch an effort to pretrain a model from scratch, CEO David Ha said at an event here.
这款暂定名为 “Sakana Sawara” 的新模型将完全在内部开发，包括向模型输入海量数据的预训练阶段。开发工作计划于 11 月启动，规模较小的模型预计于 2027 年春季推出，随后将着手开发更大规模的版本。
Provisionally named Sakana Sawara, the new model will be developed entirely in-house, including the pretraining stage in which the model is fed massive amounts of data. Development is scheduled to begin in November, with a smaller-scale model due out in the spring of 2027, followed by work on a larger version.
预计该大规模模型将包含数千亿个以上参数，这是衡量模型规模和能力的指标。即便如此，它的规模仍可能小于美国和中国开发的最先进前沿模型。
The larger model is expected to contain more than several hundred billion parameters, a measure of model size and capability. Even so, it may be smaller than the most advanced frontier models developed in the U.S. and China.
6 月面向公众发布的 “Sakana Fugu” 通过组合多个 AI 模型实现了强劲性能。它在国内运行，被视为一种主权 AI —— 即其基础设施在本国境内建设和维护的 AI。但其性能仍可能受到其架构中外国开发模型的影响。
Sakana Fugu, released for general use in June, achieves strong performance by combining multiple AI models. It is operated domestically and is considered a form of sovereign AI -- that is, AI whose infrastructure is built and maintained within a country. But its performance can still be influenced by the foreign-developed models in its architecture.
Sakana AI 首席执行官 David Ha 于 10 月 26 日在东京发言。
Sakana AI CEO David Ha speaks in Tokyo on Oct. 26.
该公司专注于国防和金融领域的 AI 应用。7 月，它与防卫省签署了一份概念验证合同，涵盖 AI 驱动的情报分析。完全由国内自主开发的 Sawara 将更易于部署在国防和网络安全等敏感领域。
The company has focused on AI applications in defense and finance. In July, it and the Ministry of Defense signed a proof-of-concept contract covering AI-powered intelligence analysis. The all-domestically developed Sawara would be easier to deploy in such sensitive fields as defense and cybersecurity.
AI 系统在识别软件漏洞方面的能力已大幅增强。它们能够快速检测并协助修复安全缺陷，以及分析网络攻击，正成为网络防御不可或缺的力量。与此同时，攻击者越来越多地利用 AI 增强自身能力，使得防御技术的进步更加迫切。
AI systems have become dramatically more capable of identifying software vulnerabilities. Their ability to rapidly detect and help fix security flaws, as well as analyze cyberattacks, is becoming indispensable for cyberdefense. At the same time, attackers are increasingly using AI to enhance their own capabilities, making defensive advances more urgent.
尽管Sakana AI是日本最大的独角兽企业，但它仍远小于美国和中国的领先科技公司，其中一些公司正在开发拥有数万亿参数的AI模型。
Although Sakana AI is Japan's largest unicorn, it remains much smaller than the leading American and Chinese technology companies, some of which are developing AI models with trillions of parameters.