I make things and explore places. Co-Founder @CitadelAI. Research at UTokyo. Product @GoogleBrain. Math @nyushanghai.
Tokyo
Joined October 2013
- Tweets198
- Following308
- Followers493
- Likes754
Kenny Song retweeted
高度なフロンティアAIやAIエージェントの登場は、AIの安全性・セキュリティを客観的に評価する体制整備と基盤構築が、喫緊の課題であることを示しています。Citadel AIは、日本における実効性ある評価環境の検討と、安全・安心なAIの社会実装を支援してまいります。
citadel-ai.com/ja/news/2026/…
#信頼できるAI
Kenny Song retweeted
生成AI導入に必須となる、総計４万種類の安全性検証用データセットの提供を開始します！
citadel-ai.com/ja/news/2025/…
Kenny Song retweeted
Citadel AI has joined Singapore's Global AI Assurance Sandbox to provide third-party testing of generative AI and agentic applications.
Learn more at: citadel-ai.com/news/2025/09/…
Kenny Song retweeted
経営課題化するAIガバナンス：
Citadel AI、Global AI Assurance Sandboxに参画。エージェント型AIや、データ漏洩やプロンプトインジェクションといったリスクにも対応。
citadel-ai.com/ja/news/2025/…
Kenny Song retweeted
AI Safety discussion panel in Google Shibuya with Kit Kitamura, @helloksong and @HiroTHamadaJP
It was great to make our first podcast appearance talking about AI safety, governance, and global adoption with @dwhitena and @chrisbenson!
Thanks for inviting us to the show!
🎤Our latest episode is out!
AI is moving fast — but not always safely.🚦
We sit down with @kobayashi_rick & @helloksong of @CitadelAI to discuss AI safety, Japan vs US adoption, real-world failures & strategies for reliable AI monitoring.
🎧 [link below]
#AISafety #GenAI
Kenny Song retweeted
Congratulations to Arseniy Aseev, Founding Engineer at Citadel AI, for passing the ISO 42001 Lead Auditor certification!
While Citadel AI isn't an auditing firm, we're actively integrating the latest AI governance practices from ISO 42001 into Citadel Lens to help our customers achieve their AI safety and compliance goals.
Very proud that we won Japan's national project on AI safety!
Our team has been working on AI evaluation and standards for a long time, and it's great to see this big investment from NEDO drive the entire ecosystem forward.
Citadel AI is pleased to announce it won the “R&D on Advancing AI Safety” project from @nedo_info (NEDO), together with @AIST_JP.
citadel-ai.com/news/2025/05/…
Kenny Song retweeted
生成AIのセーフティ基準の策定・普及と、評価・管理技術の開発を目的としたNEDOの事業において、産総研(@AIST_JP )と共に @CitadelAI が採択されました！
citadel-ai.com/ja/news/2025/…
Kenny Song retweeted
生成AIの有害な入出力を制御 シタデルAI、顧客対応に活用 - 日本経済新聞
「攻めと守り」を一元管理し、ユースケースに即したAIセーフティを実現する、次世代ファイアウォールです!
nikkei.com/article/DGXZQOUC2…
The new GPT-4.1 isn't in ChatGPT yet, but you can try it out in LangCheck Studio for free!
4.1 is our new default model here: langcheck-studio.citadel-ai.…
The Citadel AI team is excited to release LangCheck Studio – our free, public playground for experimenting with LLM-as-a-judge!
Try it here: langcheck-studio.citadel-ai.…
Learn more about LLM-as-a-judge: citadel-ai.com/blog/2025/04/…
Nice to see our launch featured in Nikkei! Eval Insight helps our enterprise customers manage and understand their AI reliability and safety.
nikkei.com/article/DGXZQOUC2… (Japanese-only)
Citadel AI launches Eval Insight: generate plain-English summaries of AI performance, safety, and security with our LLM-powered evaluation suite.
citadel-ai.com/news/2025/03/…
Kenny Song retweeted
"Eval-Centric AI" を自分以外の方が使われているのを見ると、やっぱり嬉しいですね
This talk highlights key approaches in LLMOps and Eval-centric AI behind successful projects that reached over 23 million users in Japan's largest C2C marketplace. #FOSSASIASummit2025 #FOSSASIA_mercari eventyay.com/e/4c0e0c27/sess…
Kenny Song retweeted
Fantastic talk from Mercari on how they apply Eval-Centric AI and design LLM-as-a-judge metrics in Lens!
This talk highlights key approaches in LLMOps and Eval-centric AI behind successful projects that reached over 23 million users in Japan's largest C2C marketplace. #FOSSASIASummit2025 #FOSSASIA_mercari eventyay.com/e/4c0e0c27/sess…
Excited to officially announce our work with @Deloitte to deploy our AI evaluation, monitoring, and governance solution to more organizations!
We're thrilled to announce our technology collaboration with @Deloitte to help organizations accelerate AI adoption and minimize AI risks.
citadel-ai.com/news/2025/03/…
Kenny Song retweeted
本協業においてデロイト トーマツ グループは、企業がAIのリスクを効率的にモニタリングしながら、幅広い領域でAI活用を進められるように支援し、Citadel AI @CitadelAI は、先進的なAI品質検証・モニタリング技術を提供します。
deloi.tt/4c1MlC1
Kenny Song retweeted
This week, the International Network of AI Safety Institutes published their new report on evaluating LLMs in 10 languages.
We're proud to have worked with the AI Safety Institutes to contribute Citadel AI's expertise in multilingual red teaming to this joint testing exercise.
See the report → aisi.go.jp/assets/pdf/Intern…
Using Deep Research is a pretty big mindset shift — now I can trust the accuracy of LLM answers for broad queries (not just for coding questions).
There are some UX oddities, though:
- Why is “GPT-4o with scheduled tasks” a model but Deep Research is a tool?
- Unlike other tools, Deep Research takes a minimum of 2 queries and 2 responses, but it’s not clear when you’re in Deep Research mode
- It always asks a clarification question but is off the mark 50% of the time
- Inline source links disappear if you ask it to rephrase anything
Still incredibly useful though! It’s like having super well-researched blog posts on-demand + live Q&A with the author.
Kenny Song retweeted
Introducing Citadel Lens for biometric systems!
Lens can now directly test your biometric identification and verification systems, and automatically generate reports based on ISO standards to prepare for the EU AI Act.
citadel-ai.com/news/2025/01/…
