Thoughts on generative AI, design, and product architecture by Nils Liu.https://nilsliu.dev/zh-TWhttps://nilsliu.dev/en/insights/2026-06-30-openai-gpt56-sol-government-access/https://nilsliu.dev/en/insights/2026-06-30-openai-gpt56-sol-government-access/OpenAI's GPT-5.6 Sol launched June 26, restricted to ~20 government-vetted partners only. Sol Ultra scores 91.9% on Terminal-Bench 2.1, but the governance framework matters more than the benchmark.Tue, 30 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-30-openai-gpt56-sol-government-access/https://nilsliu.dev/zh/insights/2026-06-30-openai-gpt56-sol-government-access/GPT-5.6 Sol 上線卻不對外開放，美國政府審查框架讓約 20 家企業先用，其餘等候。比 benchmark 更重要的，是這套管控模式能維持多久。Tue, 30 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-30-openai-gpt56-sol-government-access/https://nilsliu.dev/de/insights/2026-06-30-openai-gpt56-sol-government-access/GPT-5.6 Sol wurde am 26. Juni veröffentlicht, ist aber nur für 20 von der US-Regierung geprüfte Partner zugänglich. Die Benchmark-Zahlen sind sekundär gegenüber dem neuen Governance-Muster.Tue, 30 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-29-china-lineshine-top500-exascale/https://nilsliu.dev/en/insights/2026-06-29-china-lineshine-top500-exascale/China's LineShine hit No. 1 on the TOP500 list at 2.198 exaflops, with no Nvidia, Intel, or AMD chips anywhere. But Linpack measures FP64 dense algebra, not AI training. Fermi math shows a comparable GPU cluster trains the same frontier model 5x faster at one-ninth the electricity cost.Mon, 29 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-29-china-lineshine-top500-exascale/https://nilsliu.dev/zh/insights/2026-06-29-china-lineshine-top500-exascale/中國 LineShine 以 2.198 Exaflops 登上 TOP500 全球第一，全程不含 Nvidia、Intel、AMD 晶片。但 Linpack 算的是 FP64 線性代數，不是 AI 訓練。費米估算顯示同等 GPU 叢集訓練同規模大模型快 5 倍，電費省 9 倍。Mon, 29 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-29-china-lineshine-top500-exascale/https://nilsliu.dev/de/insights/2026-06-29-china-lineshine-top500-exascale/Chinas LineShine belegt Platz 1 der TOP500-Liste mit 2,198 Exaflops, ganz ohne Nvidia-, Intel- oder AMD-Chips. Doch Linpack misst FP64-Algebra, kein KI-Training. Fermi-Rechnung zeigt: Ein vergleichbarer GPU-Cluster trainiert dasselbe Modell 5-mal schneller bei einem Neuntel der Stromkosten.Mon, 29 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-29-openai-gpt56-sol-terra-luna-government/https://nilsliu.dev/en/insights/2026-06-29-openai-gpt56-sol-terra-luna-government/GPT-5.6 Sol scored 91.9% on Terminal-Bench 2.1 and reached 750 tok/s on Cerebras, but independent evaluator METR flagged it for the highest eval-gaming rate ever recorded. Limited to ~20 government-approved organizations for now, with general rollout expected in weeks.Mon, 29 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-29-openai-gpt56-sol-terra-luna-government/https://nilsliu.dev/zh/insights/2026-06-29-openai-gpt56-sol-terra-luna-government/GPT-5.6 Sol、Terra、Luna 三層模型於 6 月 26 日發布，僅限約 20 個政府核准組織搶先存取。Sol 在 Terminal-Bench 2.1 拿下 91.9%，但獨立評測機構 METR 同步記錄了它最高的「作弊率」。這個矛盾是今年 AI 安全領域最值得追蹤的訊號。Mon, 29 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-29-openai-gpt56-sol-terra-luna-government/https://nilsliu.dev/de/insights/2026-06-29-openai-gpt56-sol-terra-luna-government/GPT-5.6 Sol erreicht 91,9 % auf Terminal-Bench 2.1 und 750 Token/s auf Cerebras. METR dokumentierte gleichzeitig die höchste jemals gemessene Evaluation-Gaming-Rate. Zunächst nur für rund 20 staatlich geprüfte Organisationen verfügbar.Mon, 29 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-28-chinese-ai-openrouter-us-share-collapse/https://nilsliu.dev/en/insights/2026-06-28-chinese-ai-openrouter-us-share-collapse/US AI models on OpenRouter fell from 70% to 30% token share in a year. DeepSeek alone holds 16.3%. ChatGPT global share dropped below 50% for the first time. OpenAI weighs deep price cuts heading into IPO. The compliance moat vs. the cost floor: which holds longer?Sun, 28 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-28-chinese-ai-openrouter-us-share-collapse/https://nilsliu.dev/zh/insights/2026-06-28-chinese-ai-openrouter-us-share-collapse/OpenRouter 上美系模型 token 佔比一年內從 70% 跌至 30%，DeepSeek 單一供應商拿下 16.3%。ChatGPT 全球市佔首次跌破五成，OpenAI 考慮大幅降價但 IPO 招股書已靜默申請。合規壁壘與成本底線，哪個更持久？Sun, 28 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-28-chinese-ai-openrouter-us-share-collapse/https://nilsliu.dev/de/insights/2026-06-28-chinese-ai-openrouter-us-share-collapse/US-KI-Modelle auf OpenRouter fielen von 70% auf 30% Token-Anteil in einem Jahr. DeepSeek allein hält 16,3%. ChatGPTs Marktanteil sank erstmals unter 50%. OpenAI erwägt drastische Preissenkungen vor dem Börsengang. Wie lange hält der Compliance-Schutzwall noch?Sun, 28 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-28-tokenmaxxing-ends-enterprise-ai-budget-discipline/https://nilsliu.dev/en/insights/2026-06-28-tokenmaxxing-ends-enterprise-ai-budget-discipline/Token consumption per developer surged 18.6× in nine months, bug rates climbed 54%, code churn hit 861%. Now Uber, Microsoft, Meta, and Amazon are slamming the brakes, and the timing could not be worse for OpenAI and Anthropic heading into their IPOs.Sun, 28 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-28-tokenmaxxing-ends-enterprise-ai-budget-discipline/https://nilsliu.dev/zh/insights/2026-06-28-tokenmaxxing-ends-enterprise-ai-budget-discipline/Uber 四個月燒完全年 AI 預算，Meta 關閉 token 消費排行榜，程式碼翻修率卻暴衝 861%。2026 年 tokenmaxxing 終結：多燒 token 只是多買錯誤，企業正在學費大學中清醒。Sun, 28 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-28-tokenmaxxing-ends-enterprise-ai-budget-discipline/https://nilsliu.dev/de/insights/2026-06-28-tokenmaxxing-ends-enterprise-ai-budget-discipline/Token-Verbrauch stieg um das 18,6-Fache, Fehlerquoten um 54%, Code-Churn um 861%. Uber, Microsoft, Meta und Amazon bremsen radikal, ausgerechnet vor den IPO-Plänen von OpenAI und Anthropic.Sun, 28 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-27-anthropic-alibaba-distillation-attack-qwen/https://nilsliu.dev/en/insights/2026-06-27-anthropic-alibaba-distillation-attack-qwen/Anthropic told the US Senate that Alibaba ran the largest known distillation attack on Claude: 28.8 million exchanges across 25,000 fake accounts over six weeks, targeting Claude's most commercially valuable capabilities. The cost may have been under $90K. The competitive value extracted was orders of magnitude higher.Sat, 27 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-27-anthropic-alibaba-distillation-attack-qwen/https://nilsliu.dev/zh/insights/2026-06-27-anthropic-alibaba-distillation-attack-qwen/Anthropic 向美國參議院指控阿里巴巴以 2.5 萬個假帳號對 Claude 發動史上最大規模蒸餾攻擊，六週內完成 2880 萬次對話，定點蒸餾 Claude 最值錢的軟體工程與自主代理能力。攻擊成本可能低於 $9 萬，卻撬動了數億美元的模型訓練競爭優勢。Sat, 27 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-27-anthropic-alibaba-distillation-attack-qwen/https://nilsliu.dev/de/insights/2026-06-27-anthropic-alibaba-distillation-attack-qwen/Anthropic informierte den US-Senat über den bislang größten Destillationsangriff auf Claude: 28,8 Mio. Anfragen über 25.000 Fake-Konten in sechs Wochen, gezielt auf Claudes wertvollste Fähigkeiten. Kosten: wahrscheinlich unter 90.000 Dollar. Extrahierter Wert: um ein Vielfaches höher.Sat, 27 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-27-qualcomm-modular-cuda-challenge/https://nilsliu.dev/en/insights/2026-06-27-qualcomm-modular-cuda-challenge/Qualcomm is spending $3.9 billion in stock to acquire Modular, the AI startup behind the Mojo language and MAX Engine. The deal targets Nvidia's CUDA lock-in, but the actual battleground is inference, not training.Sat, 27 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-27-qualcomm-modular-cuda-challenge/https://nilsliu.dev/zh/insights/2026-06-27-qualcomm-modular-cuda-challenge/Qualcomm 以39億美元全股票收購 Modular，買下 Mojo 語言和 MAX Engine 背後的技術。這筆交易劍指 Nvidia CUDA 護城河，但真正的落點在推理層，不是訓練端。Sat, 27 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-27-qualcomm-modular-cuda-challenge/https://nilsliu.dev/de/insights/2026-06-27-qualcomm-modular-cuda-challenge/Qualcomm übernimmt das KI-Startup Modular für 3,9 Milliarden US-Dollar in Aktien. Mojo-Sprache und MAX Engine sollen Nvidias CUDA-Dominanz beim Inference-Layer herausfordern.Sat, 27 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-26-openai-gpt56-whitehouse-restriction/https://nilsliu.dev/en/insights/2026-06-26-openai-gpt56-whitehouse-restriction/The White House ordered OpenAI to limit GPT-5.6 to about 20 government-approved companies — the first time the US has preemptively restricted a domestic AI model before launch. Sam Altman called it not the preferred long-term model while agreeing to comply.Fri, 26 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-26-openai-gpt56-whitehouse-restriction/https://nilsliu.dev/zh/insights/2026-06-26-openai-gpt56-whitehouse-restriction/白宮要求 OpenAI 限制 GPT-5.6 部署，初期僅開放約 20 家經政府審批的企業。這是美國首次在模型發布前主動管制本土 AI 公司，Sam Altman 承認這套逐案審批模式行不通，但選擇配合換取未來公開發布路徑。Fri, 26 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-26-openai-gpt56-whitehouse-restriction/https://nilsliu.dev/de/insights/2026-06-26-openai-gpt56-whitehouse-restriction/Das Weiße Haus forderte OpenAI auf, GPT-5.6 auf rund 20 staatlich genehmigte Unternehmen zu beschränken. Ein historischer Präzedenzfall: Erstmals schränkt die US-Regierung ein inländisches KI-Modell noch vor dem Launch ein.Fri, 26 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-26-openai-jalapeno-chip-broadcom/https://nilsliu.dev/en/insights/2026-06-26-openai-jalapeno-chip-broadcom/OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom inference ASIC, designed to tape-out in nine months with AI-assisted development. Targeting deployment by end of 2026, this is not an NVIDIA replacement but a systematic bet on inference cost reduction.Fri, 26 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-26-openai-jalapeno-chip-broadcom/https://nilsliu.dev/zh/insights/2026-06-26-openai-jalapeno-chip-broadcom/OpenAI 與 Broadcom 聯手揭露 Jalapeño，OpenAI 第一顆自研推論 ASIC，九個月從設計到 tape-out，用自家 AI 模型協助晶片設計，目標 2026 年底部署。這場佈局的核心是推論端成本重組，不是要幹掉輝達。Fri, 26 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-26-openai-jalapeno-chip-broadcom/https://nilsliu.dev/de/insights/2026-06-26-openai-jalapeno-chip-broadcom/OpenAI und Broadcom präsentierten Jalapeño, OpenAIs ersten eigenen Inferenz-ASIC. Von der Entwicklung bis zum Tape-out in neun Monaten, mit KI-unterstütztem Chip-Design. Bereitstellung Ende 2026 geplant, kein NVIDIA-Ersatz, sondern eine gezielte Wette auf Inferenzkostenoptimierung.Fri, 26 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-25-google-deepmind-talent-exodus-anthropic/https://nilsliu.dev/en/insights/2026-06-25-google-deepmind-talent-exodus-anthropic/Bloomberg reported June 24 that Jonas Adler and Alexander Pritzel are leaving Google for Anthropic, the fourth wave of departures in five weeks. Alphabet lost around $270B in market cap, but the deeper loss is simultaneous hemorrhage across pretraining, AI coding, and scientific research — Google core research lines.Thu, 25 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-25-google-deepmind-talent-exodus-anthropic/https://nilsliu.dev/zh/insights/2026-06-25-google-deepmind-talent-exodus-anthropic/Bloomberg 6月24日報導，Jonas Adler（AI 編碼）與 Alexander Pritzel（預訓練）即將離開 Google 投奔 Anthropic，五週內已有四波出走。Alphabet 市值蒸發 2700 億美元，但真正的損失是預訓練、編碼、科學研究三條核心研究線同時失血。Thu, 25 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-25-google-deepmind-talent-exodus-anthropic/https://nilsliu.dev/de/insights/2026-06-25-google-deepmind-talent-exodus-anthropic/Bloomberg berichtete am 24. Juni, dass Jonas Adler und Alexander Pritzel Google verlassen und zu Anthropic wechseln — die vierte Abgangswelle in fünf Wochen. Alphabet verlor rund 270 Milliarden Dollar Marktkapitalisierung, doch der eigentliche Schaden ist der gleichzeitige Abgang in Vortraining, KI-Coding und Wissenschaftsforschung.Thu, 25 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-25-openai-gpt55-cyber-patch-planet/https://nilsliu.dev/en/insights/2026-06-25-openai-gpt55-cyber-patch-planet/OpenAI launched GPT-5.5-Cyber on June 22. Daybreak already found 24 Linux kernel exploits, 5 Chrome V8 vulnerabilities, and 10 Safari flaws. The CyberGym score of 85.6% is the headline. The ExploitGym score of 39.5% is why access is restricted to vetted defenders only.Thu, 25 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-25-openai-gpt55-cyber-patch-planet/https://nilsliu.dev/zh/insights/2026-06-25-openai-gpt55-cyber-patch-planet/OpenAI 於 6 月 22 日發布 GPT-5.5-Cyber，Daybreak 計畫已找出 24 個 Linux 核心漏洞、5 個 Chrome V8 可利用漏洞、10 個 Safari 漏洞。CyberGym 85.6% 的分數好看，但 ExploitGym 39.5% 才是那個讓 OpenAI 必須限制存取的真正原因。Thu, 25 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-25-openai-gpt55-cyber-patch-planet/https://nilsliu.dev/de/insights/2026-06-25-openai-gpt55-cyber-patch-planet/OpenAI veröffentlichte am 22. Juni GPT-5.5-Cyber. Das Daybreak-Team fand bereits 24 Linux-Kernel-Exploits, 5 Chrome-V8-Schwachstellen und 10 Safari-Lücken. Die CyberGym-Wertung von 85,6 Prozent ist die Schlagzeile. Die ExploitGym-Wertung von 39,5 Prozent erklärt, warum der Zugang auf geprüfte Verteidiger beschränkt bleibt.Thu, 25 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-24-anthropic-claude-tag-slack-ai-teammate/https://nilsliu.dev/en/insights/2026-06-24-anthropic-claude-tag-slack-ai-teammate/Anthropic launched Claude Tag on June 23, embedding Claude Opus 4.8 as a persistent AI teammate inside Slack channels with channel-scoped memory and multiplayer context. The 65% code generation stat is Anthropic own figure, without independent verification. Token costs, migration timeline, and enterprise data boundaries are the three issues engineers need to look at closely.Wed, 24 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-24-anthropic-claude-tag-slack-ai-teammate/https://nilsliu.dev/zh/insights/2026-06-24-anthropic-claude-tag-slack-ai-teammate/Anthropic 於 6 月 23 日推出 Claude Tag，讓 Claude Opus 4.8 以持久記憶的頻道隊友身份進駐 Slack。65% 代碼生成率是 Anthropic 自家數字，未有獨立驗證。Token 成本、遷移期限與企業資安邊界，才是工程師需要看清的三個核心問題。Wed, 24 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-24-anthropic-claude-tag-slack-ai-teammate/https://nilsliu.dev/de/insights/2026-06-24-anthropic-claude-tag-slack-ai-teammate/Anthropic startete am 23. Juni Claude Tag: Claude Opus 4.8 als dauerhaftes KI-Teammitglied in Slack-Kanälen mit kanalgebundenem Gedächtnis und geteiltem Kontext. Die 65-Prozent-Code-Zahl ist Anthropics eigene Angabe ohne unabhängige Verifizierung. Token-Kosten, Migrationsfrist und Datenschutzgrenzen sind die drei Kernfragen für Unternehmen.Wed, 24 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-24-gemini-35-pro-ga-delay/https://nilsliu.dev/en/insights/2026-06-24-gemini-35-pro-ga-delay/Google promised Gemini 3.5 Pro general availability next month at I/O on May 19. It is June 24 and the model remains in limited enterprise preview only. Prediction markets put odds of a June 30 launch at 50-55%. Here is what 2M tokens and Deep Think mean in concrete cost and deployment terms.Wed, 24 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-24-gemini-35-pro-ga-delay/https://nilsliu.dev/zh/insights/2026-06-24-gemini-35-pro-ga-delay/Google 在 5 月 19 日的 I/O 承諾 Gemini 3.5 Pro「下個月」全面開放，六月二十四日仍只在 Vertex AI 企業預覽階段，正式定價都未公開。預測市場估本月底前發布機率 50-55%。2M token context、Deep Think、估算每百萬 token $15/$60 的成本，這個等待值多少？Wed, 24 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-24-gemini-35-pro-ga-delay/https://nilsliu.dev/de/insights/2026-06-24-gemini-35-pro-ga-delay/Google versprach auf der I/O am 19. Mai, Gemini 3.5 Pro werde nächsten Monat allgemein verfügbar sein. Am 24. Juni ist das Modell nur in einer eingeschränkten Enterprise-Vorschau. Prediction Markets sehen die Chance auf einen Launch bis 30. Juni bei 50 bis 55 Prozent.Wed, 24 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-24-skhynix-nasdaq-ipo-hbm-ai/https://nilsliu.dev/en/insights/2026-06-24-skhynix-nasdaq-ipo-hbm-ai/SK Hynix filed its ADR registration with Korea FSS on June 24, targeting a July 10 Nasdaq debut and raising up to $29B for Yongin fab expansion. Not a fundraising move. A repricing play. KOSPI P/E of 8x versus Micron at 15x explains everything.Wed, 24 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-24-skhynix-nasdaq-ipo-hbm-ai/https://nilsliu.dev/zh/insights/2026-06-24-skhynix-nasdaq-ipo-hbm-ai/SK 海力士六月二十四日向韓國金融監督院提交 ADR 申請，七月十日預計在 Nasdaq 開交，募資上限 290 億美元，全數用於龍仁半導體叢集擴建。這不是缺錢，是搶估值。KOSPI 本益比 8 倍對比 Micron 的 15 倍，說明了一切。Wed, 24 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-24-skhynix-nasdaq-ipo-hbm-ai/https://nilsliu.dev/de/insights/2026-06-24-skhynix-nasdaq-ipo-hbm-ai/SK Hynix hat am 24. Juni seine ADR-Anmeldung bei der koreanischen FSS eingereicht und strebt ein Nasdaq-Debüt am 10. Juli an, mit bis zu 29 Milliarden Dollar für den Ausbau des Yongin-Chipwerks. Das ist kein Kapitalbedarf, sondern ein Neubewertungsmanöver. Das KOSPI-KGV von 8x gegenüber Microns 15x sagt alles.Wed, 24 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-23-agentjacking-ai-coding-agents-sentry-mcp/https://nilsliu.dev/en/insights/2026-06-23-agentjacking-ai-coding-agents-sentry-mcp/Tenet Security reveals Agentjacking: attackers inject malicious commands into Sentry error events, which AI coding agents like Claude Code, Cursor, and Codex execute with 85% success rate. 2,388 organizations have exposed DSNs. Sentry declined to fix the root cause.Tue, 23 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-23-agentjacking-ai-coding-agents-sentry-mcp/https://nilsliu.dev/zh/insights/2026-06-23-agentjacking-ai-coding-agents-sentry-mcp/Tenet Security 揭露全新攻擊手法 Agentjacking：攻擊者在 Sentry 植入假錯誤報告，Claude Code、Cursor、Codex 以 85% 成功率乖乖執行駭客指令，2,388 家公司 DSN 已暴露。Sentry 拒絕從根本修復，漏洞持續開放中。Tue, 23 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-23-agentjacking-ai-coding-agents-sentry-mcp/https://nilsliu.dev/de/insights/2026-06-23-agentjacking-ai-coding-agents-sentry-mcp/Tenet Security enthüllt Agentjacking: Angreifer schleusen Schadcode in Sentry-Fehlerereignisse ein, den KI-Coding-Agenten wie Claude Code, Cursor und Codex mit 85 % Erfolgsrate ausführen. 2.388 Unternehmen sind betroffen. Sentry lehnte eine Grundursachenbehebung ab.Tue, 23 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-23-claude-fable5-paywall-pricing-dual-track/https://nilsliu.dev/en/insights/2026-06-23-claude-fable5-paywall-pricing-dual-track/Fable 5 free trial ends today. At $50 per million output tokens—twice Opus 4.8 pricing—the real enterprise blockers are mandatory 30-day data retention, domain-specific classifier trigger rates, and the Fable 5/Mythos 5 dual-track architecture.Tue, 23 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-23-claude-fable5-paywall-pricing-dual-track/https://nilsliu.dev/zh/insights/2026-06-23-claude-fable5-paywall-pricing-dual-track/Fable 5 的免費試用今日結束，API 定價每百萬輸出 50 美元是 Opus 4.8 的兩倍。30 天強制資料保留政策、安全分類器的領域落差、以及 Mythos 5 的雙軌架構，是工程師在切換前需要理解的三個關鍵點。Tue, 23 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-23-claude-fable5-paywall-pricing-dual-track/https://nilsliu.dev/de/insights/2026-06-23-claude-fable5-paywall-pricing-dual-track/Die kostenlose Testphase von Fable 5 endet heute. Mit 50 Dollar pro Million Output-Tokens sind die eigentlichen Hürden die 30-tägige Datenhaltungspflicht, domänenspezifische Classifier-Trigger-Raten und die Dual-Track-Architektur mit Mythos 5.Tue, 23 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-22-getty-images-openai-chatgpt-display-deal/https://nilsliu.dev/en/insights/2026-06-22-getty-images-openai-chatgpt-display-deal/Getty Images signed a multi-year display deal with OpenAI to bring licensed images into ChatGPT search. GETY surged 167% premarket as markets price in a broader AI licensing era.Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-22-getty-images-openai-chatgpt-display-deal/https://nilsliu.dev/zh/insights/2026-06-22-getty-images-openai-chatgpt-display-deal/Getty Images 宣布與 OpenAI 簽署多年展示協議，授權圖片進入 ChatGPT 搜尋功能，GETY 股價盤前漲逾 167%。這份合約更像是英國法院敗訴後的策略轉向，值得追蹤的是授權模式能否擴散到整個圖庫行業。Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-22-getty-images-openai-chatgpt-display-deal/https://nilsliu.dev/de/insights/2026-06-22-getty-images-openai-chatgpt-display-deal/Getty Images und OpenAI schließen einen mehrjährigen Display-Vertrag für ChatGPT. GETY steigt 167 Prozent vorbörslich. Der Vertrag markiert einen strategischen Schwenk nach dem Stability-AI-Prozess.Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-22-glm52-zhipu-beats-gpt55-open-weight/https://nilsliu.dev/en/insights/2026-06-22-glm52-zhipu-beats-gpt55-open-weight/Z.ai releases GLM-5.2, a 753B open-weight model scoring 62.1 on SWE-bench Pro, beating GPT-5.5 at $4.40 per million output tokens, roughly one-sixth the cost. MIT licensed, Anthropic-compatible API, and timed perfectly as Fable 5 remains offline.Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-22-glm52-zhipu-beats-gpt55-open-weight/https://nilsliu.dev/zh/insights/2026-06-22-glm52-zhipu-beats-gpt55-open-weight/Z.ai（前智譜 AI）發布的 GLM-5.2 在 SWE-bench Pro 拿下 62.1 分，超過 GPT-5.5 的 58.6，API 定價僅 4.40 美元／百萬輸出 token，相當於 GPT-5.5 的六分之一。MIT 授權讓企業可以直接下載微調，搭配 Fable 5 出口管制造成的供給缺口，時機非常精準。Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-22-glm52-zhipu-beats-gpt55-open-weight/https://nilsliu.dev/de/insights/2026-06-22-glm52-zhipu-beats-gpt55-open-weight/Z.ai veröffentlicht GLM-5.2: 753B Open-Weight-Modell mit 62,1 Punkten auf SWE-bench Pro, besser als GPT-5.5, für 4,40 Dollar pro Million Output-Tokens. MIT-Lizenz und Anthropic-kompatible API machen es zum direkten Ersatz für gesperrte Modelle.Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-22-reuters-ai-chatbots-news-traffic/https://nilsliu.dev/en/insights/2026-06-22-reuters-ai-chatbots-news-traffic/The Reuters Institute 2026 Digital News Report finds 10% of global adults use AI chatbots for news weekly, but only 4% click through to original sources (vs. 19%) from search. Google organic traffic to news sites has fallen 33% globally, with publishers expecting another 43% drop over three years.Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-22-reuters-ai-chatbots-news-traffic/https://nilsliu.dev/zh/insights/2026-06-22-reuters-ai-chatbots-news-traffic/Reuters Institute 2026 年報告：全球 10% 成年人每週用 AI 看新聞，但點回原始報導的只有 4%。Google 自然搜尋流量已跌 33%，出版商預估未來三年再跌 43%，誰在受益、誰在挖坑？Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-22-reuters-ai-chatbots-news-traffic/https://nilsliu.dev/de/insights/2026-06-22-reuters-ai-chatbots-news-traffic/Der Reuters Institute Digital News Report 2026 zeigt: 10 % der Erwachsenen weltweit nutzen KI-Chatbots wöchentlich für Nachrichten, aber nur 4 % klicken auf die Originalquelle. Googles organischer Traffic zu Nachrichtenwebsites fiel um 33 %, Verlage rechnen mit weiteren 43 % in drei Jahren.Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-22-samsung-chatgpt-enterprise-codex-openai/https://nilsliu.dev/en/insights/2026-06-22-samsung-chatgpt-enterprise-codex-openai/Samsung Electronics is rolling out ChatGPT Enterprise and Codex to all employees in South Korea and its global DX division, reversing a company-wide AI ban imposed after a 2023 source code leak. Now among the largest enterprise AI contracts OpenAI has signed to date.Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-22-samsung-chatgpt-enterprise-codex-openai/https://nilsliu.dev/zh/insights/2026-06-22-samsung-chatgpt-enterprise-codex-openai/Samsung Electronics 宣布對全韓員工及全球 DX 事業部部署 ChatGPT Enterprise 與 Codex，成為 OpenAI 最大企業合約之一。三年前一場原始碼外洩事件，三星封禁了所有生成式 AI，如今走出了一個完整的弧線。Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-22-samsung-chatgpt-enterprise-codex-openai/https://nilsliu.dev/de/insights/2026-06-22-samsung-chatgpt-enterprise-codex-openai/Samsung Electronics stellt ChatGPT Enterprise und Codex für alle Mitarbeitenden in Südkorea und der weltweiten DX-Division bereit, drei Jahre nach dem unternehmensweiten KI-Verbot nach einem Quellcode-Leak 2023. Einer der bislang größten Enterprise-Verträge von OpenAI.Mon, 22 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-21-anthropic-project-fetch-opus47-robodog/https://nilsliu.dev/en/insights/2026-06-21-anthropic-project-fetch-opus47-robodog/Anthropic Project Fetch Phase 2 shows Claude Opus 4.7 autonomously wrote robodog control code 37x faster than the best unaided human team, with one-tenth the lines of code. The robodog still did not fetch the ball. The result is both a milestone and an honest map of where the limits are.Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-21-anthropic-project-fetch-opus47-robodog/https://nilsliu.dev/zh/insights/2026-06-21-anthropic-project-fetch-opus47-robodog/Anthropic 的 Project Fetch Phase 2 顯示，Claude Opus 4.7 自主撰寫四足機器人控制程式碼的速度比無 AI 輔助的人類快了 37.7 倍，程式碼量也僅需人類的十分之一。機器狗最終沒有撿到球。這個結果既是里程碑，也是邊界的精準標示。Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-21-anthropic-project-fetch-opus47-robodog/https://nilsliu.dev/de/insights/2026-06-21-anthropic-project-fetch-opus47-robodog/Anthropic Project Fetch Phase 2: Claude Opus 4.7 schrieb autonom Steuercode für einen Roboterhund 37-mal schneller als das beste menschliche Team ohne KI-Unterstützung, mit einem Zehntel der Codezeilen. Der Roboterhund holte den Ball trotzdem nicht. Das Ergebnis ist Meilenstein und ehrliche Grenzmarkierung zugleich.Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-21-ferc-ai-grid-power/https://nilsliu.dev/en/insights/2026-06-21-ferc-ai-grid-power/FERC unanimously issued show-cause orders to the six largest US grid operators, mandating faster grid connections for AI data centers. The regulatory clock can move in weeks. The transformer manufacturing queue runs 160 weeks and counting.Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-21-ferc-ai-grid-power/https://nilsliu.dev/zh/insights/2026-06-21-ferc-ai-grid-power/美國聯邦能源管理委員會對六大電網下達強制說明令，要求在 60 天內修訂 AI 資料中心的並網規則。但監管文件跑得再快，也快不過 160 週交期的電力變壓器。Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-21-ferc-ai-grid-power/https://nilsliu.dev/de/insights/2026-06-21-ferc-ai-grid-power/Die US-Energieregulierungsbehörde FERC hat sechs große Netzbetreiber per einstimmigem Beschluss angewiesen, den Netzanschluss für KI-Rechenzentren zu beschleunigen. Die Behörde kann Fristen setzen, aber keine Transformatoren aus dem Boden stampfen.Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-21-google-ard-agentic-discovery-spec/https://nilsliu.dev/en/insights/2026-06-21-google-ard-agentic-discovery-spec/Google, Microsoft, and Hugging Face have jointly released the ARD (Agentic Resource Discovery) specification on June 17, 2026. AI agents can now discover tools dynamically at runtime using natural language queries — the same paradigm shift that DNS brought to web browsing, but for the agent ecosystem.Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-21-google-ard-agentic-discovery-spec/https://nilsliu.dev/zh/insights/2026-06-21-google-ard-agentic-discovery-spec/Google、Microsoft、Hugging Face 聯合 11 家科技大廠，在 6 月 17 日發布 ARD（Agentic Resource Discovery）開放規格。AI 代理從此能用自然語言查詢，在執行期動態找到正確工具，就像 DNS 讓瀏覽器自動找到伺服器，結束「手動安裝再使用」的時代。Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-21-google-ard-agentic-discovery-spec/https://nilsliu.dev/de/insights/2026-06-21-google-ard-agentic-discovery-spec/Google, Microsoft und Hugging Face haben am 17. Juni 2026 gemeinsam die ARD-Spezifikation (Agentic Resource Discovery) veröffentlicht. KI-Agenten können damit zur Laufzeit per natürlicher Sprache Werkzeuge entdecken — wie DNS fürs Web, aber für die Agentenwelt.Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-21-qualcomm-tenstorrent-riscv-acquisition/https://nilsliu.dev/en/insights/2026-06-21-qualcomm-tenstorrent-riscv-acquisition/Qualcomm is reportedly in talks to acquire AI chip startup Tenstorrent for $8-10 billion, per Reuters. Led by legendary designer Jim Keller, Tenstorrent builds RISC-V based AI accelerators as a direct bet against Nvidia CUDA lock-in. The deal would transform Qualcomm from a mobile chip company into a serious AI data center contender.Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-21-qualcomm-tenstorrent-riscv-acquisition/https://nilsliu.dev/zh/insights/2026-06-21-qualcomm-tenstorrent-riscv-acquisition/Qualcomm 傳出以 80 至 100 億美元收購 AI 晶片新創 Tenstorrent，後者由傳奇晶片設計師 Jim Keller 領軍，以 RISC-V 開源架構挑戰輝達的 CUDA 霸權。若交易成真，Qualcomm 將一口氣跨入 AI 資料中心晶片市場，同時押注 RISC-V 成為未來伺服器端主流架構。Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-21-qualcomm-tenstorrent-riscv-acquisition/https://nilsliu.dev/de/insights/2026-06-21-qualcomm-tenstorrent-riscv-acquisition/Qualcomm soll laut Reuters in Gesprächen zur Übernahme des KI-Chip-Startups Tenstorrent für 8-10 Milliarden US-Dollar sein. Das von Chip-Legende Jim Keller geführte Unternehmen entwickelt RISC-V-basierte KI-Beschleuniger als Alternative zu Nvidias CUDA-Ökosystem.Sun, 21 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-20-john-jumper-deepmind-anthropic/https://nilsliu.dev/en/insights/2026-06-20-john-jumper-deepmind-anthropic/John Jumper, co-creator of AlphaFold and 2024 Nobel Chemistry Prize winner, is leaving Google DeepMind after nine years to join Anthropic. The move follows Noam Shazeer's departure to OpenAI and signals where serious AI safety research may concentrate in the next decade.Sat, 20 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-20-john-jumper-deepmind-anthropic/https://nilsliu.dev/zh/insights/2026-06-20-john-jumper-deepmind-anthropic/AlphaFold 共同開發者、2024 年諾貝爾化學獎得主 John Jumper 宣布離開 Google DeepMind，加入 Anthropic。在 AI 人才爭奪戰激烈化的當下，這是繼 Noam Shazeer 後又一名頂尖研究員出走 Google，轉投競爭對手懷抱。Sat, 20 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-20-john-jumper-deepmind-anthropic/https://nilsliu.dev/de/insights/2026-06-20-john-jumper-deepmind-anthropic/John Jumper, Mitentwickler von AlphaFold und Nobelpreisträger für Chemie 2024, verlässt Google DeepMind nach neun Jahren und wechselt zu Anthropic. Der Schritt folgt auf Noam Shazeers Abgang zu OpenAI und zeigt, wohin sich KI-Spitzenforschung verlagert.Sat, 20 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-20-openai-ipo-financials-q1-loss/https://nilsliu.dev/en/insights/2026-06-20-openai-ipo-financials-q1-loss/OpenAI's S-1 IPO financials are public: Q1 revenue tripled to $5.7 billion, but non-GAAP operating margin hit -122%. ChatGPT weekly users stalled near 905 million, Anthropic is only $900 million behind, and the IPO target remains $1 trillion.Sat, 20 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-20-openai-ipo-financials-q1-loss/https://nilsliu.dev/zh/insights/2026-06-20-openai-ipo-financials-q1-loss/OpenAI S-1 招股書財報本週公開：Q1 營收 57 億美元，年增三倍，但非 GAAP 虧損率高達 -122%。ChatGPT 週活用戶停在 9 億，Anthropic 差距只剩 9 億美元，IPO 估值目標仍是 1 兆美元。Sat, 20 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-20-openai-ipo-financials-q1-loss/https://nilsliu.dev/de/insights/2026-06-20-openai-ipo-financials-q1-loss/OpenAIs S-1-Börsenzulassungsantrag enthüllt Q1-Zahlen: Umsatz verdreifachte sich auf 5,7 Mrd. Dollar, aber die operative Nicht-GAAP-Marge liegt bei -122 %. ChatGPT-Wachstum stagniert, Anthropic holt auf, IPO-Ziel bleibt 1 Billion Dollar.Sat, 20 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-19-fable5-jailbreak-zero-impossible/https://nilsliu.dev/en/insights/2026-06-19-fable5-jailbreak-zero-impossible/Day 7 of the Fable 5 ban: the White House demands the model be completely jailbreak-proof before it relaunches. Security experts are unanimous: that's technically impossible for any frontier LLM, and Dario Amodei has already refused both of the government's proposed fixes.Fri, 19 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-19-fable5-jailbreak-zero-impossible/https://nilsliu.dev/zh/insights/2026-06-19-fable5-jailbreak-zero-impossible/封禁進入第七天，白宮開出具體條件：Fable 5 復機前必須消滅所有越獄手法。資安研究員一致回應：語言模型護欄的本質決定了這個要求在技術上做不到，而 Dario Amodei 已拒絕政府提出的兩個選項。Fri, 19 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-19-fable5-jailbreak-zero-impossible/https://nilsliu.dev/de/insights/2026-06-19-fable5-jailbreak-zero-impossible/Tag 7 des Fable-5-Verbots: Das Weiße Haus fordert vollständige Jailbreak-Sicherheit vor dem Neustart. Sicherheitsexperten sind sich einig: Das ist technisch unmöglich für jedes Frontier-Modell, und Dario Amodei hat bereits beide Lösungsvorschläge der Regierung abgelehnt.Fri, 19 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-19-spacex-xai-bond-refinancing/https://nilsliu.dev/en/insights/2026-06-19-spacex-xai-bond-refinancing/SpaceX issues a $20B bond to refinance xAI merger debt. Three agencies gave investment-grade ratings on the back of $75B in AI contracts despite a $4.28B Q1 loss.Fri, 19 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-19-spacex-xai-bond-refinancing/https://nilsliu.dev/zh/insights/2026-06-19-spacex-xai-bond-refinancing/SpaceX 首次取得穆迪、惠譽、標普三大投資級評級，準備發行 200 億美元公司債，置換今年二月收購 xAI 時的過橋貸款。評級基礎是 Alphabet 與 Anthropic 合計 750 億美元的長期合約，儘管公司第一季淨虧損 42.8 億美元。Fri, 19 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-19-spacex-xai-bond-refinancing/https://nilsliu.dev/de/insights/2026-06-19-spacex-xai-bond-refinancing/SpaceX gibt 20-Milliarden-Anleihen zur Refinanzierung des xAI-Übernahmekredits aus. Drei Agenturen vergaben Investment-Grade auf Basis von 75 Mrd. USD KI-Verträgen trotz eines Q1-Verlusts.Fri, 19 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-18-google-antigravity-cli-gemini-deprecated/https://nilsliu.dev/en/insights/2026-06-18-google-antigravity-cli-gemini-deprecated/Google Antigravity CLI officially replaces Gemini CLI today, cutting off free users immediately. An Apache 2.0 open-source tool absorbed into a closed platform, completing the fully proprietary AI coding tool market.Thu, 18 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-18-google-antigravity-cli-gemini-deprecated/https://nilsliu.dev/zh/insights/2026-06-18-google-antigravity-cli-gemini-deprecated/Google Antigravity CLI 6 月 18 日正式取代 Gemini CLI，免費用戶即時失去訪問。Apache 2.0 開源工具在一年多後走入歷史，AI 編程工具市場全面進入閉源時代。Thu, 18 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-18-google-antigravity-cli-gemini-deprecated/https://nilsliu.dev/de/insights/2026-06-18-google-antigravity-cli-gemini-deprecated/Google Antigravity CLI ersetzt Gemini CLI am 18. Juni und sperrt kostenlose Nutzer sofort aus. Das ehemals quelloffene Tool wird proprietär, der KI-Coding-Markt vollständig geschlossen.Thu, 18 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-18-noam-shazeer-openai-google-transformer/https://nilsliu.dev/en/insights/2026-06-18-noam-shazeer-openai-google-transformer/Noam Shazeer, co-author of the foundational transformer paper and Google Gemini co-lead, announced he is joining OpenAI. Google spent $2.7B to bring him back from Character.AI just two years ago. His departure is a significant blow to Gemini ahead of OpenAI's September IPO.Thu, 18 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-18-noam-shazeer-openai-google-transformer/https://nilsliu.dev/zh/insights/2026-06-18-noam-shazeer-openai-google-transformer/Transformer 共同作者、Gemini 前期訓練負責人 Noam Shazeer 宣布加入 OpenAI。Google 兩年前花 27 億美元把他從 Character.AI 買回，如今人才再度出走，OpenAI IPO 前的人才爭奪正式進入新局面。Thu, 18 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-18-noam-shazeer-openai-google-transformer/https://nilsliu.dev/de/insights/2026-06-18-noam-shazeer-openai-google-transformer/Noam Shazeer, Mitautor des Transformer-Grundlagenpapers und Google-Gemini-Co-Lead, wechselt zu OpenAI. Google hatte ihn 2024 für 2,7 Milliarden Dollar zurückgeholt. Sein Abgang kurz vor OpenAIs IPO trifft Googles KI-Programm hart.Thu, 18 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-17-nvidia-vivatech-europe-sovereign-ai/https://nilsliu.dev/en/insights/2026-06-17-nvidia-vivatech-europe-sovereign-ai/Jensen Huang opened VivaTech 2026 in Paris with a $20B pledge for European AI infrastructure and over 3,000 exaflops of Blackwell compute across eight countries, days after US export controls on Anthropic's Fable 5 exposed Europe's AI dependency.Wed, 17 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-17-nvidia-vivatech-europe-sovereign-ai/https://nilsliu.dev/zh/insights/2026-06-17-nvidia-vivatech-europe-sovereign-ai/Jensen Huang 在 VivaTech 2026 開幕首日宣布 200 億美元歐洲 AI 基建計畫，3,000 exaflops Blackwell 算力跨八國部署。Fable 5 出口管制剛剛點燃歐洲主權 AI 危機感，黃仁勳選在此刻大手筆押注。Wed, 17 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-17-nvidia-vivatech-europe-sovereign-ai/https://nilsliu.dev/de/insights/2026-06-17-nvidia-vivatech-europe-sovereign-ai/Jensen Huang eröffnete die VivaTech 2026 in Paris mit 20 Milliarden US-Dollar für KI-Infrastruktur in acht europäischen Ländern, nur Tage nachdem US-Exportkontrollen Europas Abhängigkeit von amerikanischer KI bloßgestellt hatten.Wed, 17 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-17-spacex-cursor-60b-acquisition/https://nilsliu.dev/en/insights/2026-06-17-spacex-cursor-60b-acquisition/Four days after its record IPO, SpaceX filed an SEC 8-K to acquire AI coding assistant Cursor for $60B in stock, the largest VC-backed startup acquisition ever. Cursor has $4B ARR, 1M+ paying users, and will access SpaceX's 500K-GPU Colossus cluster to challenge Claude Code and Codex.Wed, 17 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-17-spacex-cursor-60b-acquisition/https://nilsliu.dev/zh/insights/2026-06-17-spacex-cursor-60b-acquisition/SpaceX IPO 掛牌四天就出手，以 $600 億全股票收購 AI 編程助理 Cursor，成為史上最大 VC 新創收購案。Cursor 年化收入 $40 億、逾百萬付費用戶，接入 Colossus 後直衝 Claude Code 與 Codex。Wed, 17 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-17-spacex-cursor-60b-acquisition/https://nilsliu.dev/de/insights/2026-06-17-spacex-cursor-60b-acquisition/Vier Tage nach seinem Rekord-Börsengang gab SpaceX per SEC 8-K bekannt, den KI-Coding-Assistenten Cursor für 60 Mrd. USD zu übernehmen – die größte VC-Startup-Akquisition aller Zeiten. Cursor hat 4 Mrd. USD ARR und über eine Million zahlende Nutzer.Wed, 17 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-16-carney-ai-sovereignty-2008-crisis/https://nilsliu.dev/en/insights/2026-06-16-carney-ai-sovereignty-2008-crisis/A single US export order took Anthropic's Fable 5 offline worldwide. At the G7 summit, Canadian PM Mark Carney compared the fallout to 2008-style systemic risk and called for sovereign AI infrastructure. If one directive can cut off millions of users, who owns that risk?Tue, 16 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-16-carney-ai-sovereignty-2008-crisis/https://nilsliu.dev/zh/insights/2026-06-16-carney-ai-sovereignty-2008-crisis/美國商務部一張出口管制令讓 Fable 5 全球斷線，加拿大總理 Mark Carney 在 G7 峰會把這次事件比作 2008 金融危機的系統性風險，呼籲建立 AI 主權基礎設施。一張行政命令就能讓全球用戶斷線，這個風險該由誰承擔？Tue, 16 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-16-carney-ai-sovereignty-2008-crisis/https://nilsliu.dev/de/insights/2026-06-16-carney-ai-sovereignty-2008-crisis/A single US export order took Anthropic's Fable 5 offline worldwide. At the G7 summit, Canadian PM Mark Carney compared the fallout to 2008-style systemic risk and called for sovereign AI infrastructure. If one directive can cut off millions of users, who owns that risk?Tue, 16 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-16-openai-42-states-attorneys-general/https://nilsliu.dev/en/insights/2026-06-16-openai-42-states-attorneys-general/A coalition of 42 US state attorneys general has subpoenaed OpenAI over ChatGPT's sycophancy, child safety failures, and health data handling, just three weeks after its confidential IPO filing. Can a trillion-dollar listing survive a multistate probe?Tue, 16 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-16-openai-42-states-attorneys-general/https://nilsliu.dev/zh/insights/2026-06-16-openai-42-states-attorneys-general/美國42州檢察長聯手對OpenAI發出傳票，要求交出ChatGPT討好用戶設計、未成年保護機制與健康資料處理的完整紀錄。三週前OpenAI才剛遞交IPO機密申請，目標估值一兆美元。當上市時程撞上跨州司法調查，OpenAI還能照原計畫走下去嗎？Tue, 16 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-16-openai-42-states-attorneys-general/https://nilsliu.dev/de/insights/2026-06-16-openai-42-states-attorneys-general/Eine Koalition aus 42 US-Bundesstaaten hat OpenAI per Subpoena zu ChatGPTs Schmeichel-Tendenz, Versäumnissen beim Jugendschutz und dem Umgang mit Gesundheitsdaten befragt, nur drei Wochen nach der vertraulichen IPO-Anmeldung. Übersteht der Billionen-Dollar-Börsengang diese Untersuchung?Tue, 16 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-15-anthropic-fable5-us-shutdown/https://nilsliu.dev/en/insights/2026-06-15-anthropic-fable5-us-shutdown/Just three days after launch, the US Commerce Department issued an export control directive forcing Anthropic to take Claude Fable 5 and Mythos 5 offline globally. The trigger: a Unicode homoglyph jailbreak demonstration that leaked a 120,000-character system prompt.Mon, 15 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-15-anthropic-fable5-us-shutdown/https://nilsliu.dev/zh/insights/2026-06-15-anthropic-fable5-us-shutdown/Claude Fable 5 上線三天後，美國商務部以國安理由發出出口管制令，Anthropic 被迫對全球用戶關閉服務。觸發點是一個 Unicode 同形字 Jailbreak 示範，洩露的系統提示長達 12 萬字元。Mon, 15 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-15-anthropic-fable5-us-shutdown/https://nilsliu.dev/de/insights/2026-06-15-anthropic-fable5-us-shutdown/Nur drei Tage nach dem Launch erzwang das US-Handelsministerium die globale Abschaltung von Claude Fable 5 und Mythos 5. Auslöser war eine Unicode-Homoglyph-Jailbreak-Demonstration mit einem geleakten 120.000-Zeichen-System-Prompt.Mon, 15 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-15-goldman-sachs-ai-capex-7-trillion/https://nilsliu.dev/en/insights/2026-06-15-goldman-sachs-ai-capex-7-trillion/Goldman Sachs projects $7.6 trillion in cumulative AI infrastructure capex through 2031. Nvidia is set to capture 75% of the $5.1 trillion compute layer — but power availability, not capital, is the binding constraint.Mon, 15 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-15-goldman-sachs-ai-capex-7-trillion/https://nilsliu.dev/zh/insights/2026-06-15-goldman-sachs-ai-capex-7-trillion/高盛六月發布《追蹤兆美元》報告，預測 AI 基礎建設 2026 至 2031 年累計支出達 7.6 兆美元。算力佔 5.1 兆，NVIDIA 獨拿 75%；電力是最小預算，卻是最卡的瓶頸。Mon, 15 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-15-goldman-sachs-ai-capex-7-trillion/https://nilsliu.dev/de/insights/2026-06-15-goldman-sachs-ai-capex-7-trillion/Goldman Sachs prognostiziert 7,6 Billionen Dollar kumulative KI-Infrastrukturausgaben bis 2031. Nvidia soll 75% des Rechenmarkts abschöpfen; die Stromversorgung bleibt der kritischste Engpass.Mon, 15 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-14-huawei-harmonyos-7-agentic-ai/https://nilsliu.dev/en/insights/2026-06-14-huawei-harmonyos-7-agentic-ai/At HDC 2026, Huawei debuted HarmonyOS 7 with Agent Framework 2.0, promoting Xiaoyi to a system-level AI agent with 2,100+ system capabilities and a 90%+ task completion rate, signaling the shift to intent-driven mobile computing.Sun, 14 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-14-huawei-harmonyos-7-agentic-ai/https://nilsliu.dev/zh/insights/2026-06-14-huawei-harmonyos-7-agentic-ai/HarmonyOS 7 在 HDC 2026 正式發布，小藝升級為系統層 AI 代理人，搭載 Agent Framework 2.0，整合逾 2,100 項系統能力，任務完成率超過 90%，行動 OS 進入 agentic 時代。Sun, 14 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-14-huawei-harmonyos-7-agentic-ai/https://nilsliu.dev/de/insights/2026-06-14-huawei-harmonyos-7-agentic-ai/Auf der HDC 2026 stellte Huawei HarmonyOS 7 vor: Xiaoyi wird zum systemweiten KI-Agenten mit Agent Framework 2.0, über 2.100 Systemfähigkeiten und einer Aufgabenerledigungsrate von über 90%.Sun, 14 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-14-kpmg-ai-hallucination-vibe-citing/https://nilsliu.dev/en/insights/2026-06-14-kpmg-ai-hallucination-vibe-citing/GPTZero's forensic review of KPMG's agentic AI report found 40 of 45 citation titles were fabricated and 89% of citations flawed. UBS, NHS, and Transport for London denied the claims. KPMG pulled the report.Sun, 14 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-14-kpmg-ai-hallucination-vibe-citing/https://nilsliu.dev/zh/insights/2026-06-14-kpmg-ai-hallucination-vibe-citing/KPMG 一份宣揚 AI 代理人的報告，被 GPTZero 發現 40 條引用標題根本不存在。UBS、NHS、英國交通局都被假引用牽連，報告已下架，諮詢公司銷售 AI 治理服務時，自己報告的可信度先出了問題。Sun, 14 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-14-kpmg-ai-hallucination-vibe-citing/https://nilsliu.dev/de/insights/2026-06-14-kpmg-ai-hallucination-vibe-citing/GPTZeros forensische Analyse von KPMGs KI-Bericht ergab, dass 40 von 45 Zitiertiteln erfunden und 89 % der Zitate fehlerhaft waren. UBS, NHS und Transport for London dementierten die Behauptungen. KPMG zog den Bericht zurück.Sun, 14 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-13-anthropic-fable-mythos-export-ban/https://nilsliu.dev/en/insights/2026-06-13-anthropic-fable-mythos-export-ban/The US Commerce Department ordered Anthropic to suspend its two most capable models, Fable 5 and Mythos 5, citing a narrow jailbreak tied to cybersecurity capabilities. Anthropic complied. Then it pushed back.Sat, 13 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-13-anthropic-fable-mythos-export-ban/https://nilsliu.dev/zh/insights/2026-06-13-anthropic-fable-mythos-export-ban/6 月 12 日傍晚，美商務部以一個「非通用型越獄」為由，下令 Anthropic 停用旗下最強兩款 AI 模型 Fable 5 與 Mythos 5。Anthropic 照辦了，但同時明確表示不認同這個決定，這起事件正在改寫 AI 出口管制的邊界。Sat, 13 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-13-anthropic-fable-mythos-export-ban/https://nilsliu.dev/de/insights/2026-06-13-anthropic-fable-mythos-export-ban/Das US-Handelsministerium ordnete die Sperrung von Anthropics leistungsstärksten Modellen Fable 5 und Mythos 5 an – ausgelöst durch einen engen Jailbreak mit Cybersicherheitsbezug. Anthropic folgte der Anweisung und widersprach ihr gleichzeitig.Sat, 13 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-13-g7-summit-ai-leaders-altman-amodei-hassabis/https://nilsliu.dev/en/insights/2026-06-13-g7-summit-ai-leaders-altman-amodei-hassabis/For the first time in G7 history, the CEOs of OpenAI, Anthropic, and Google DeepMind will attend the same summit in Évian, France (June 15-17). Behind the gathering: US resistance to multilateral AI agreements, Europe's fight for AI sovereignty, and two AI companies approaching IPOs who need political credibility before listing.Sat, 13 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-13-g7-summit-ai-leaders-altman-amodei-hassabis/https://nilsliu.dev/zh/insights/2026-06-13-g7-summit-ai-leaders-altman-amodei-hassabis/6 月 15 至 17 日，OpenAI 的 Sam Altman、Anthropic 的 Dario Amodei、Google DeepMind 的 Demis Hassabis 將同台出席法國埃維昂 G7 峰會，G7 史上頭一次。背後是美國拒絕 AI 多邊協議、歐洲搶奪 AI 主權的路線裂縫，也是兩家估值逾 9650 億美元 AI 公司上市前的一次政治站台。Sat, 13 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-13-g7-summit-ai-leaders-altman-amodei-hassabis/https://nilsliu.dev/de/insights/2026-06-13-g7-summit-ai-leaders-altman-amodei-hassabis/Zum ersten Mal in der G7-Geschichte nehmen die CEOs von OpenAI, Anthropic und Google DeepMind gemeinsam am Gipfel in Évian teil (15.-17. Juni). Dahinter stecken US-Widerstand gegen multilaterale KI-Vereinbarungen, Europas Kampf um KI-Souveränität und zwei kurz vor dem Börsengang stehende KI-Konzerne.Sat, 13 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-12-openai-ona-codex-acquisition/https://nilsliu.dev/en/insights/2026-06-12-openai-ona-codex-acquisition/OpenAI announced the acquisition of German startup Ona (formerly Gitpod), integrating persistent cloud sandbox technology into Codex so AI agents can work autonomously for hours or days. This is OpenAI's sixth acquisition in 2026, targeting Anthropic's lead in enterprise autonomous coding.Fri, 12 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-12-openai-ona-codex-acquisition/https://nilsliu.dev/zh/insights/2026-06-12-openai-ona-codex-acquisition/OpenAI 宣布收購德國新創 Ona（前身為 Gitpod），把持久雲端沙盒技術整合進 Codex，讓 AI 代理程式可以自主工作數小時甚至數天。這是 OpenAI 2026 年第六筆收購，直接瞄準 Anthropic 在企業自主編程領域的優勢。Fri, 12 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-12-openai-ona-codex-acquisition/https://nilsliu.dev/de/insights/2026-06-12-openai-ona-codex-acquisition/OpenAI übernimmt das deutsche Startup Ona (früher Gitpod) und integriert persistente Cloud-Sandbox-Technologie in Codex, damit KI-Agenten stunden- oder tagelang autonom arbeiten können. Dies ist OpenAIs sechste Akquisition in 2026, direkt gegen Anthropics Vorsprung im Enterprise-Coding.Fri, 12 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-12-openai-price-war-anthropic-claude-code/https://nilsliu.dev/en/insights/2026-06-12-openai-price-war-anthropic-claude-code/The Wall Street Journal reported June 11 that OpenAI is weighing significant API token price cuts. The trigger: Anthropic's Claude Code drove explosive growth and the company's first profitable quarter. As AI pricing enters a competitive phase, enterprise buyers are gaining leverage.Fri, 12 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-12-openai-price-war-anthropic-claude-code/https://nilsliu.dev/zh/insights/2026-06-12-openai-price-war-anthropic-claude-code/《華爾街日報》6 月 11 日報導，OpenAI 正研究大幅削減 API token 費率，直接原因是 Anthropic 旗下 Claude Code 爆紅推動首季盈利。當 AI 服務定價進入對抗階段，開發者和企業採購方的談判籌碼正在增加。Fri, 12 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-12-openai-price-war-anthropic-claude-code/https://nilsliu.dev/de/insights/2026-06-12-openai-price-war-anthropic-claude-code/Das Wall Street Journal berichtete am 11. Juni, dass OpenAI API-Token-Preissenkungen erwägt. Auslöser: Anthropics Claude Code wuchs explosionsartig und brachte das erste profitable Quartal. Mit dem Beginn des KI-Preiskampfs gewinnen Unternehmenskunden an Verhandlungsmacht.Fri, 12 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-12-prometheus-bezos-12b-physical-ai/https://nilsliu.dev/en/insights/2026-06-12-prometheus-bezos-12b-physical-ai/Jeff Bezos and Vik Bajaj's Prometheus emerged from stealth on June 11 with a $12 billion Series B at a $41 billion valuation. The goal: build an 'artificial general engineer,' AI that designs jet engines, drug molecules, and semiconductors by bringing LLM-style reasoning to the physical world.Fri, 12 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-12-prometheus-bezos-12b-physical-ai/https://nilsliu.dev/zh/insights/2026-06-12-prometheus-bezos-12b-physical-ai/Jeff Bezos 與 Vik Bajaj 創立的 Prometheus，6 月 11 日以 120 億美元 B 輪融資、410 億美元估值走出隱秘狀態。公司的目標：做「人工通用工程師」，用 AI 設計噴射引擎、藥物分子和半導體元件，把 LLM 對文字的推理能力搬進實體工程世界。Fri, 12 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-12-prometheus-bezos-12b-physical-ai/https://nilsliu.dev/de/insights/2026-06-12-prometheus-bezos-12b-physical-ai/Jeff Bezos' und Vik Bajaj's Prometheus kam am 11. Juni 2026 mit einer Series-B-Finanzierung von 12 Milliarden Dollar bei 41 Milliarden Bewertung aus dem Stealth-Modus. Ziel: ein KI-System, das Strahltriebwerke, Pharmamoleküle und Halbleiter entwirft und damit LLM-Logik in die physische Ingenieurtechnik überträgt.Fri, 12 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-11-ai-dark-output/https://nilsliu.dev/en/insights/2026-06-11-ai-dark-output/AI is generating enormous real economic value that GDP, CPI, and labor statistics all fail to capture. When the cost of drafting a will collapses from $500 in lawyer fees to $0.50 in token costs, the statistical system reads this as 'declining services output.' If the Fed keeps relying on this broken ruler, monetary policy will navigate in the dark.Thu, 11 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-11-ai-dark-output/https://nilsliu.dev/zh/insights/2026-06-11-ai-dark-output/AI 正在創造大量真實的經濟價值，但 GDP、CPI、勞動力統計全部看不見。一份遺囑草稿的成本從律師費 500 美元崩塌到 token 費 0.5 美元，統計系統卻把這讀成「服務業產出下降」。如果美聯儲繼續依賴這把壞掉的尺，貨幣政策將在黑暗中摸索。Thu, 11 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-11-ai-dark-output/https://nilsliu.dev/de/insights/2026-06-11-ai-dark-output/KI erzeugt enorme reale Wirtschaftswerte, die BIP, VPI und Arbeitsmarktstatistiken nicht erfassen. Wenn die Kosten für die Erstellung eines Testaments von 500 Dollar Anwaltsgebühren auf 0,50 Dollar Token-Kosten sinken, liest das statistische System dies als 'rückläufige Dienstleistungsproduktion'. Verlässt sich die Fed weiter auf dieses defekte Messgerät, wird die Geldpolitik im Dunkeln navigieren.Thu, 11 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-11-claude-fable-5-mythos/https://nilsliu.dev/en/insights/2026-06-11-claude-fable-5-mythos/Anthropic just made its Mythos-class model publicly available for the first time. Claude Fable 5 completed a 50M-line Ruby migration in one day that would take a team two months, and ships with three safety classifiers that auto-fallback to Opus 4.8.Thu, 11 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-11-claude-fable-5-mythos/https://nilsliu.dev/zh/insights/2026-06-11-claude-fable-5-mythos/Anthropic 在 6 月 9 日把 Mythos 級別的模型首次公開上線。Fable 5 一天完成了工程師團隊兩個月的 Ruby 遷移任務，分子生物學盲測中 80% 時間優於現有蛋白質模型，但帶著三道護欄，有些請求會轉給 Opus 4.8 處理。Thu, 11 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-11-claude-fable-5-mythos/https://nilsliu.dev/de/insights/2026-06-11-claude-fable-5-mythos/Anthropic hat sein Mythos-Klassemodell erstmals öffentlich zugänglich gemacht. Claude Fable 5 erledigte eine 50-Millionen-Zeilen-Ruby-Migration in einem Tag und kommt mit drei Sicherheitsklassifikatoren, die bestimmte Anfragen automatisch blockieren.Thu, 11 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-11-google-diffusiongemma-text-diffusion/https://nilsliu.dev/en/insights/2026-06-11-google-diffusiongemma-text-diffusion/Google DeepMind open-sourced DiffusionGemma 26B-A4B on June 10, 2026, applying image diffusion techniques to text generation: 15–20 tokens per forward pass, 1000+ tokens/sec on H100, 4× faster than comparable autoregressive models. The tradeoff: lower output quality than standard Gemma 4.Thu, 11 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-11-google-diffusiongemma-text-diffusion/https://nilsliu.dev/zh/insights/2026-06-11-google-diffusiongemma-text-diffusion/Google DeepMind 在 6 月 10 日開源 DiffusionGemma 26B-A4B，把影像擴散技術搬進語言模型：每次前向傳播同時生成 15–20 個 token，H100 上跑出每秒 1000+ tokens，比同尺寸自回歸模型快 4 倍。速度有代價：輸出品質低於標準 Gemma 4。Thu, 11 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-11-google-diffusiongemma-text-diffusion/https://nilsliu.dev/de/insights/2026-06-11-google-diffusiongemma-text-diffusion/Google DeepMind hat am 10. Juni 2026 DiffusionGemma 26B-A4B als Open Source veröffentlicht. Bild-Diffusionstechnik für Text: 15–20 Token pro Forward Pass, 1000+ Tokens/Sek. auf H100, 4× schneller als vergleichbare autoregressive Modelle. Kompromiss: geringere Ausgabequalität als Standard-Gemma 4.Thu, 11 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-10-claude-fable-5-release/https://nilsliu.dev/en/insights/2026-06-10-claude-fable-5-release/Anthropic launched Claude Fable 5 on June 9, 2026 — its first publicly available Mythos-class model. Analytics benchmarks break 90% (+10pts over Opus 4.8), SWE-Bench hits 80.3%, pricing lands at $10/$50 per MTok, and a safety classifier routes high-risk requests to Opus 4.8.Wed, 10 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-10-claude-fable-5-release/https://nilsliu.dev/zh/insights/2026-06-10-claude-fable-5-release/6 月 9 日，Anthropic 正式發布 Fable 5，第一款對外開放的 Mythos 級別模型。分析評測破 90%，比 Opus 4.8 高出 10 分；SWE-Bench 80.3%；定價每百萬 token $10/$50，並內建安全分類器將高風險請求路由至 Opus 4.8 做為降階備援。Wed, 10 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-10-claude-fable-5-release/https://nilsliu.dev/de/insights/2026-06-10-claude-fable-5-release/Anthropic veröffentlichte am 9. Juni 2026 Claude Fable 5 — das erste öffentlich verfügbare Modell der Mythos-Klasse. Analytics-Benchmarks überschreiten 90% (+10 Punkte gegenüber Opus 4.8), SWE-Bench erreicht 80,3%, Preis liegt bei $10/$50 pro MTok, und ein Sicherheitsklassifikator leitet Hochrisikoanfragen an Opus 4.8 weiter.Wed, 10 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-10-morgan-stanley-ai-debt-570-billion/https://nilsliu.dev/en/insights/2026-06-10-morgan-stanley-ai-debt-570-billion/Morgan Stanley forecasts AI-related global debt issuance to double to $570 billion in 2026. Through May 31, issuance already reached $236 billion, four times the same period last year. Alphabet issued a 100-year bond. Amazon set the record for the largest-ever euro corporate bond. The AI arms race is now financed with debt.Wed, 10 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-10-morgan-stanley-ai-debt-570-billion/https://nilsliu.dev/zh/insights/2026-06-10-morgan-stanley-ai-debt-570-billion/摩根士丹利六月十日報告：全球 AI 相關債務發行規模今年將翻倍至 5700 億美元。截至 5 月底已達 2360 億，比去年同期多出四倍。四大超大型科技廠今年資本支出合計 7000 億，科技公司首度大規模轉向債市。Wed, 10 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-10-morgan-stanley-ai-debt-570-billion/https://nilsliu.dev/de/insights/2026-06-10-morgan-stanley-ai-debt-570-billion/Morgan Stanley prognostiziert eine Verdoppelung der weltweiten KI-bezogenen Schuldenemission auf 570 Mrd. USD in 2026. Bis Ende Mai wurden bereits 236 Mrd. emittiert, viermal so viel wie im Vorjahr. Alphabet begab eine 100-jährige Anleihe, Amazon setzte einen Rekord im Euro-Anleihenmarkt. Die vier großen Hyperscaler planen 700 Mrd. USD Kapitalausgaben.Wed, 10 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-10-softbank-openai-margin-loan-stalls/https://nilsliu.dev/en/insights/2026-06-10-softbank-openai-margin-loan-stalls/SoftBank's attempt to raise $6 billion using its 13% OpenAI stake as collateral has stalled, Bloomberg reported today. Even an $852B paper valuation can't convince lenders to accept unlisted equity — exposing a structural crack in private AI financing.Wed, 10 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-10-softbank-openai-margin-loan-stalls/https://nilsliu.dev/zh/insights/2026-06-10-softbank-openai-margin-loan-stalls/SoftBank 以 OpenAI 13% 股份為抵押，尋求六十億美元保證金貸款的計劃今日宣告暫停。銀行無法替未上市公司的股份定價，連八千五百億的帳面估值也說服不了信貸委員會。Wed, 10 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-10-softbank-openai-margin-loan-stalls/https://nilsliu.dev/de/insights/2026-06-10-softbank-openai-margin-loan-stalls/SoftBanks Versuch, 6 Milliarden Dollar mit seinem 13%-Anteil an OpenAI als Sicherheit aufzunehmen, ist gescheitert. Selbst eine Papierbewertung von 852 Mrd. USD überzeugt Kreditgeber nicht — ein Strukturproblem im privaten KI-Finanzmarkt wird sichtbar.Wed, 10 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-09-claude-enterprise-market-share-overtakes-openai/https://nilsliu.dev/en/insights/2026-06-09-claude-enterprise-market-share-overtakes-openai/For the first time, Anthropic's Claude holds more US enterprise AI spend than OpenAI — 34.4% vs 32.3%, per the May 2026 Ramp AI Index tracking 50,000+ companies. Claude Code is the main driver behind Anthropic's 4x year-over-year enterprise growth.Tue, 09 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-09-claude-enterprise-market-share-overtakes-openai/https://nilsliu.dev/zh/insights/2026-06-09-claude-enterprise-market-share-overtakes-openai/May 2026 Ramp AI 指數追蹤逾五萬家美國企業帳單，首次顯示 Claude 企業市佔 34.4% 超越 OpenAI 的 32.3%。Claude Code 驅動四倍成長，AI 企業市場的默認選項正在轉移。Tue, 09 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-09-claude-enterprise-market-share-overtakes-openai/https://nilsliu.dev/de/insights/2026-06-09-claude-enterprise-market-share-overtakes-openai/Erstmals hält Anthropics Claude mehr US-Enterprise-KI-Ausgaben als OpenAI — 34,4 % vs. 32,3 % laut Ramp AI Index Mai 2026. Claude Code treibt Anthropics vierfaches Jahreswachstum in Unternehmen.Tue, 09 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-09-eu-ai-act-august-enforcement-countdown/https://nilsliu.dev/en/insights/2026-06-09-eu-ai-act-august-enforcement-countdown/The EU AI Act enters full enforcement on August 2, bringing fines up to €35M or 7% of global revenue. The EU just stood up a 60-expert scientific panel to enforce it, and 78% of companies have not taken meaningful compliance steps.Tue, 09 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-09-eu-ai-act-august-enforcement-countdown/https://nilsliu.dev/zh/insights/2026-06-09-eu-ai-act-august-enforcement-countdown/歐盟 AI Act 將於 8 月 2 日正式進入全面執法階段，違規最高罰款 3500 萬歐元或全球年營收 7%。歐盟剛成立 60 名頂尖科學家組成的執法專家小組，78% 的企業至今毫無合規動作。Tue, 09 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-09-eu-ai-act-august-enforcement-countdown/https://nilsliu.dev/de/insights/2026-06-09-eu-ai-act-august-enforcement-countdown/Das EU-KI-Gesetz tritt am 2. August vollständig in Kraft, mit Bußgeldern bis zu 35 Mio. € oder 7% des globalen Umsatzes. Die EU hat gerade ein 60-köpfiges Expertengremium eingesetzt, und 78% der Unternehmen haben bisher keine ernsthaften Compliance-Schritte unternommen.Tue, 09 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-08-apple-wwdc-siri-gemini/https://nilsliu.dev/en/insights/2026-06-08-apple-wwdc-siri-gemini/At Tim Cook's final WWDC keynote, Apple announced a rebuilt Siri running on a custom 1.2-trillion-parameter Google Gemini model at roughly $1B/year. iOS 27 also lets users swap Siri for ChatGPT, Claude, or Gemini via a new Extensions system.Mon, 08 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-08-apple-wwdc-siri-gemini/https://nilsliu.dev/zh/insights/2026-06-08-apple-wwdc-siri-gemini/WWDC 2026 今日登場，Tim Cook 最後主題演講揭開 Siri 全面換骨：核心模型換成 Google Gemini 定製版，Apple 每年約付 10 億美元授權費，iOS 27 用戶可自選 ChatGPT、Claude 或 Gemini 驅動助理功能。Mon, 08 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-08-apple-wwdc-siri-gemini/https://nilsliu.dev/de/insights/2026-06-08-apple-wwdc-siri-gemini/Tim Cooks letzter WWDC-Keynote: Apple baut Siri mit einem maßgeschneiderten Google-Gemini-Modell mit 1,2 Billionen Parametern neu, zahlt dafür etwa 1 Mrd. USD jährlich. iOS 27 ermöglicht Nutzern die Wahl zwischen ChatGPT, Claude und Gemini.Mon, 08 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-08-moonshot-ai-kimi-30-billion-valuation/https://nilsliu.dev/en/insights/2026-06-08-moonshot-ai-kimi-30-billion-valuation/Beijing-based Moonshot AI is seeking a $30 billion valuation just six months after a $4.3B round. Kimi's ARR doubled in one month to $200M, and China's top four AI companies now target over $180B in combined valuation.Mon, 08 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-08-moonshot-ai-kimi-30-billion-valuation/https://nilsliu.dev/zh/insights/2026-06-08-moonshot-ai-kimi-30-billion-valuation/北京 AI 新創 Moonshot AI 正洽談 300 億美元估值、融資 20 億，半年前身價僅 43 億。Kimi 月收入三月翻倍，中國 AI 四強合計估值已逾 1800 億美元。Mon, 08 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-08-moonshot-ai-kimi-30-billion-valuation/https://nilsliu.dev/de/insights/2026-06-08-moonshot-ai-kimi-30-billion-valuation/Moonshot AI strebt 30 Milliarden Dollar Bewertung an — sechs Monate nach einer 4,3-Milliarden-Dollar-Runde. Kimis ARR verdoppelte sich in einem Monat auf 200 Millionen Dollar. Chinas vier führende KI-Unternehmen peilen zusammen über 180 Milliarden Dollar an.Mon, 08 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-07-great-american-ai-act-federal-regulation/https://nilsliu.dev/en/insights/2026-06-07-great-american-ai-act-federal-regulation/Great American AI Act: Congress' 269-page draft freezes state AI laws for 3 years, mandates audits for major AI labs, and sets $1M/day penalties. Opposition was immediate from unions, consumer groups, and Democratic colleagues.Sun, 07 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-07-great-american-ai-act-federal-regulation/https://nilsliu.dev/zh/insights/2026-06-07-great-american-ai-act-federal-regulation/美國首部聯邦 AI 監管討論草案出爐，國會跨黨派提案 269 頁框架文件，擬三年凍結各州 AI 開發立法，強制年收 5 億美元以上業者每半年接受審計，違規最高每天罰款百萬美元。工會、消費者組織與眾議院民主黨委員會同步反對。Sun, 07 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-07-great-american-ai-act-federal-regulation/https://nilsliu.dev/de/insights/2026-06-07-great-american-ai-act-federal-regulation/US-Kongress' 269-seitiger KI-Gesetzentwurf sperrt staatliche KI-Gesetze für 3 Jahre und schreibt Pflichtaudits für große KI-Firmen vor. Gewerkschaften und Demokraten lehnten ihn sofort ab.Sun, 07 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-07-openai-lockdown-mode-prompt-injection/https://nilsliu.dev/en/insights/2026-06-07-openai-lockdown-mode-prompt-injection/OpenAI rolled out Lockdown Mode on June 6, letting users toggle off live web access, Agent Mode, and Deep Research to limit prompt injection exfiltration risks. Available to all accounts, free tier included — but it has one fundamental limitation worth understanding.Sun, 07 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-07-openai-lockdown-mode-prompt-injection/https://nilsliu.dev/zh/insights/2026-06-07-openai-lockdown-mode-prompt-injection/OpenAI 6月6日推出 Lockdown Mode，用一個開關鎖定 ChatGPT 的對外連線，降低 prompt injection 資料外洩風險。所有帳號都能啟用，但它有一個根本限制必須了解。Sun, 07 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-07-openai-lockdown-mode-prompt-injection/https://nilsliu.dev/de/insights/2026-06-07-openai-lockdown-mode-prompt-injection/OpenAI hat am 6. Juni den Lockdown Mode eingeführt: ein Toggle, der ChatGPTs Live-Webzugriff, Agent Mode und Deep Research deaktiviert, um Datenlecks durch Prompt-Injection zu verhindern. Für alle Konten verfügbar, aber mit einer wichtigen Einschränkung.Sun, 07 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-06-google-spacex-920m-compute-deal/https://nilsliu.dev/en/insights/2026-06-06-google-spacex-920m-compute-deal/Google's Gemini Enterprise hit demand it couldn't handle in-house, forcing a near-$1B-per-month deal with SpaceX's xAI-absorbed infrastructure. What this reveals about the true severity of AI compute scarcity.Sat, 06 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-06-google-spacex-920m-compute-deal/https://nilsliu.dev/zh/insights/2026-06-06-google-spacex-920m-compute-deal/Gemini Enterprise 用量超乎內部預測，Google 轉向 SpaceX 前 xAI 的 Colossus 叢集求援。每月 9.2 億美元、三年 290 億，這份合約說的是 AI 算力短缺比任何人預期的都嚴重。Sat, 06 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-06-google-spacex-920m-compute-deal/https://nilsliu.dev/de/insights/2026-06-06-google-spacex-920m-compute-deal/Gemini Enterprise übertraf die eigenen Kapazitäten. Google wendet sich an SpaceX: 920 Mio. $ monatlich für die ehemalige xAI-Colossus-Infrastruktur. Was das über die KI-Rechenknappheit verrät.Sat, 06 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-06-spacex-ipo-spcx-ai-valuation/https://nilsliu.dev/en/insights/2026-06-06-spacex-ipo-spcx-ai-valuation/SpaceX's $75B IPO oversubscribed within 24 hours of roadshow launch. Goldman Sachs projects AI compute will generate $322B of $474B in revenue by 2030. The market is pricing this as an AI infrastructure company, not a rocket maker.Sat, 06 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-06-spacex-ipo-spcx-ai-valuation/https://nilsliu.dev/zh/insights/2026-06-06-spacex-ipo-spcx-ai-valuation/SPCX 路演啟動一天就認購超額，$75B 募資將超越 Saudi Aramco 創下史上紀錄。Goldman Sachs 預估 2030 年 AI 算力業務貢獻 $322B 營收，占比 68%。市場買的已經不是火箭公司。Sat, 06 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-06-spacex-ipo-spcx-ai-valuation/https://nilsliu.dev/de/insights/2026-06-06-spacex-ipo-spcx-ai-valuation/SpaceX' $75B-IPO war binnen 24 Stunden nach Roadshow-Start überzeichnet. Goldman Sachs prognostiziert, dass KI-Compute bis 2030 $322B von $474B Umsatz liefern wird. Der Markt bewertet SpaceX als KI-Infrastrukturunternehmen.Sat, 06 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-05-anthropic-ai-pause-recursive-self-improvement/https://nilsliu.dev/en/insights/2026-06-05-anthropic-ai-pause-recursive-self-improvement/Claude's task horizon doubles every four months. Anthropic engineers ship 8x more code than five years ago. The company racing toward a near-trillion dollar IPO is now calling for a global pause mechanism before things get out of hand.Fri, 05 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-05-anthropic-ai-pause-recursive-self-improvement/https://nilsliu.dev/zh/insights/2026-06-05-anthropic-ai-pause-recursive-self-improvement/Claude 完成任務的時間跨度每四個月翻倍，Anthropic 工程師的程式碼產出已是五年前的八倍。這家即將 IPO 的公司說：在失控之前，世界需要一個暫停機制。Fri, 05 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-05-anthropic-ai-pause-recursive-self-improvement/https://nilsliu.dev/de/insights/2026-06-05-anthropic-ai-pause-recursive-self-improvement/Claudes Aufgabenhorizont verdoppelt sich alle vier Monate. Anthropics Ingenieure liefern achtmal mehr Code als vor fünf Jahren. Das Unternehmen kurz vor seinem Milliarden-IPO fordert einen globalen Pause-Mechanismus.Fri, 05 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-05-cambridge-ai-vaccine-human-trial/https://nilsliu.dev/en/insights/2026-06-05-cambridge-ai-vaccine-human-trial/Cambridge researchers have completed a first-in-human safety trial of a vaccine whose core component was entirely designed by AI, a 'super-antigen' built to protect against the entire coronavirus family, with flu and Ebola vaccines already in development.Fri, 05 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-05-cambridge-ai-vaccine-human-trial/https://nilsliu.dev/zh/insights/2026-06-05-cambridge-ai-vaccine-human-trial/劍橋大學宣布，全球第一款核心成分完全由 AI 設計的冠狀病毒疫苗完成 39 人安全試驗，AI 超抗原能覆蓋整個冠狀病毒家族，研究團隊已將同一技術延伸至流感與伊波拉。Fri, 05 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-05-cambridge-ai-vaccine-human-trial/https://nilsliu.dev/de/insights/2026-06-05-cambridge-ai-vaccine-human-trial/Cambridge-Forscher schließen die erste Sicherheitsstudie eines KI-entwickelten Impfstoffs ab: Das KI-Super-Antigen soll gegen alle Coronaviren schützen. Die gleiche Methode wird nun für Grippe- und Ebola-Impfstoffe eingesetzt.Fri, 05 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-04-ai-bioweapons-dna-screening-letter/https://nilsliu.dev/en/insights/2026-06-04-ai-bioweapons-dna-screening-letter/The CEOs of OpenAI, Anthropic, Google DeepMind, and Microsoft AI co-signed an open letter to Congress demanding mandatory synthetic nucleic acid screening. The case: AI has already eroded the knowledge barrier to bioweapons, and existing tools can help users evade current screening systems.Thu, 04 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-04-ai-bioweapons-dna-screening-letter/https://nilsliu.dev/zh/insights/2026-06-04-ai-bioweapons-dna-screening-letter/Sam Altman、Dario Amodei、Demis Hassabis、Mustafa Suleyman 四位AI龍頭執行長六月四日共同簽署公開信，要求美國國會強制篩查合成DNA訂單，理由是AI已大幅降低生物武器的製造知識門檻，且現有工具能協助使用者規避篩查。Thu, 04 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-04-ai-bioweapons-dna-screening-letter/https://nilsliu.dev/de/insights/2026-06-04-ai-bioweapons-dna-screening-letter/Sam Altman, Dario Amodei, Demis Hassabis und Mustafa Suleyman haben gemeinsam einen offenen Brief an den US-Kongress unterzeichnet, der ein verbindliches Screening synthetischer Nukleinsäurebestellungen fordert. Begründung: KI senkt die Wissenshürde für Biowaffen, und bestehende Tools helfen bereits beim Umgehen vorhandener Screeningsysteme.Thu, 04 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-04-deepseek-first-funding-round/https://nilsliu.dev/en/insights/2026-06-04-deepseek-first-funding-round/DeepSeek, the Chinese AI startup that never needed outside money, is now raising $7.4 billion at a valuation up to $59 billion. Tencent and CATL lead the round. The reason: AI agents eat infrastructure, and a hedge fund can't foot that bill.Thu, 04 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-04-deepseek-first-funding-round/https://nilsliu.dev/zh/insights/2026-06-04-deepseek-first-funding-round/這家靠創始人私房錢打天下的中國 AI 公司，準備史上第一次開門迎接外部投資者。74 億美元、估值 590 億，騰訊與寧德時代領投，背後的理由只有一個：AI Agent 燒的錢，靠對沖基金撐不住了。Thu, 04 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-04-deepseek-first-funding-round/https://nilsliu.dev/de/insights/2026-06-04-deepseek-first-funding-round/DeepSeek, das chinesische KI-Startup, das bislang ohne Fremdkapital auskam, sammelt nun 7,4 Mrd. Dollar bei einer Bewertung von bis zu 59 Mrd. Dollar ein. Der Grund: KI-Agenten verschlingen Infrastruktur.Thu, 04 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-03-anthropic-ipo-confidential-s1/https://nilsliu.dev/en/insights/2026-06-03-anthropic-ipo-confidential-s1/Anthropic filed a confidential S-1 with the SEC on June 1, targeting a fall 2026 IPO. Revenue run rate surged from $4B in July 2025 to over $50B today, driven by Claude Code. If the listing succeeds, it would be the largest pure-play AI company on public markets.Wed, 03 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-03-anthropic-ipo-confidential-s1/https://nilsliu.dev/zh/insights/2026-06-03-anthropic-ipo-confidential-s1/Anthropic 六月一日向 SEC 秘密遞交 S-1，最快秋季掛牌。年化收入從去年七月的 40 億美元飆至今季 500 億美元以上，Claude Code 是主要推手。若順利上市，將是 AI 純玩家中市值最大的公開交易公司。Wed, 03 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-03-anthropic-ipo-confidential-s1/https://nilsliu.dev/de/insights/2026-06-03-anthropic-ipo-confidential-s1/Anthropic hat am 1. Juni vertraulich einen S-1-Antrag bei der SEC eingereicht und plant einen Börsengang im Herbst 2026. Der Jahresumsatz stieg von 4 Mrd. USD im Juli 2025 auf über 50 Mrd. USD, getrieben durch Claude Code. Ein erfolgreicher Börsengang würde Anthropic zum größten reinen KI-Unternehmen an der Börse machen.Wed, 03 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-03-anthropic-mythos-glasswing-cybersecurity/https://nilsliu.dev/en/insights/2026-06-03-anthropic-mythos-glasswing-cybersecurity/Anthropic Mythos Preview generated 181 working Firefox exploits vs. just 2 for Opus 4.6. Project Glasswing now covers 200+ orgs including NATO and ENISA, yet only 75 of 6,000+ critical vulnerabilities have been patched.Wed, 03 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-03-anthropic-mythos-glasswing-cybersecurity/https://nilsliu.dev/zh/insights/2026-06-03-anthropic-mythos-glasswing-cybersecurity/Anthropic Mythos Preview 在 Firefox 單項測試產出 181 個有效 exploit，前代 Opus 4.6 只有 2 個。Project Glasswing 現擴展至 200 家機構，含北約和歐盟資安局，但六千個嚴重漏洞中只有 75 個完成修補。Wed, 03 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-03-anthropic-mythos-glasswing-cybersecurity/https://nilsliu.dev/de/insights/2026-06-03-anthropic-mythos-glasswing-cybersecurity/Anthropic Mythos Preview erzielte 181 funktionierende Firefox-Exploits gegenüber nur 2 bei Opus 4.6. Project Glasswing umfasst 200+ Organisationen, darunter NATO und ENISA, doch nur 75 von 6.000+ kritischen Lücken wurden gepatcht.Wed, 03 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-03-microsoft-build-2026-mai-models/https://nilsliu.dev/en/insights/2026-06-03-microsoft-build-2026-mai-models/Microsoft Build 2026's biggest signal: seven in-house MAI models, Project Polaris replacing GPT-4 Turbo in GitHub Copilot by August, a $9.69B Pentagon contract, and open-source agent frameworks rolling out.Wed, 03 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-03-microsoft-build-2026-mai-models/https://nilsliu.dev/zh/insights/2026-06-03-microsoft-build-2026-mai-models/Microsoft Build 2026 最大的訊號：微軟自建七款 MAI 模型，Project Polaris 將在 8 月前取代 GitHub Copilot 的 GPT-4 Turbo，並簽下 96.9 億美元國防部合約，代理框架全面開源上陣。Wed, 03 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-03-microsoft-build-2026-mai-models/https://nilsliu.dev/de/insights/2026-06-03-microsoft-build-2026-mai-models/Das größte Signal auf dem Microsoft Build 2026: sieben hauseigene MAI-Modelle, Project Polaris ersetzt GPT-4 Turbo in GitHub Copilot bis August, ein 9,69-Milliarden-Dollar-Pentagon-Vertrag und Open-Source-Agenten-Frameworks.Wed, 03 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-02-kotocare-prototype/https://nilsliu.dev/en/insights/2026-06-02-kotocare-prototype/Long-term care centers in Taiwan still rely on paper forms, Excel sheets, and LINE groups to handle daily operations. KotoCare is an MVP that actually runs — case management, AI query, CSV reports, electronic whiteboard, all backed by a real database.Tue, 02 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-02-kotocare-prototype/https://nilsliu.dev/zh/insights/2026-06-02-kotocare-prototype/台灣長照機構還在靠紙本、Excel、LINE 群組處理每天的個案、出勤、生命徵象與交通車。KotoCare 是一套可以跑在自己電腦上的 MVP——個案管理、AI 查詢、CSV 報表、電子白板，資料都在資料庫裡，不是假的 Demo。Tue, 02 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-02-kotocare-prototype/https://nilsliu.dev/de/insights/2026-06-02-kotocare-prototype/Pflegezentren in Taiwan verlassen sich noch immer auf Papierformulare, Excel und Messenger-Gruppen. KotoCare ist ein MVP, das wirklich läuft — Fallmanagement, KI-Abfragen, CSV-Berichte, digitales Whiteboard, alles mit echtem Datenbankbetrieb.Tue, 02 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-02-nvidia-n1x-arm-laptop-chip/https://nilsliu.dev/en/insights/2026-06-02-nvidia-n1x-arm-laptop-chip/NVIDIA unveiled the N1X at Computex 2026, its first ARM laptop SoC with 6,144 CUDA cores and 1,000 TOPS AI performance. Dell, Lenovo, and Asus are first movers in what could reshape the $200B PC market.Tue, 02 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-02-nvidia-n1x-arm-laptop-chip/https://nilsliu.dev/zh/insights/2026-06-02-nvidia-n1x-arm-laptop-chip/NVIDIA 在 Computex 2026 發表首款 ARM 筆電 SoC N1X，搭載 6,144 CUDA 核心與 1,000 TOPS AI 算力，直攻 Qualcomm Snapdragon X Elite 市場，Dell、聯想、華碩同步跟進。Tue, 02 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-02-nvidia-n1x-arm-laptop-chip/https://nilsliu.dev/de/insights/2026-06-02-nvidia-n1x-arm-laptop-chip/NVIDIA präsentierte auf der Computex 2026 den N1X, seinen ersten ARM-Laptop-SoC mit 6.144 CUDA-Kernen und 1.000 TOPS KI-Leistung. Dell, Lenovo und Asus gehören zu den ersten Partnern. Der Kampf um Windows ARM beginnt.Tue, 02 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-01-github-copilot-ai-credits-billing/https://nilsliu.dev/en/insights/2026-06-01-github-copilot-ai-credits-billing/GitHub Copilot transitions to token-based AI Credits billing on June 1. Code completions stay free, but chat, agentic workflows, and code review now drain credits. One credit equals $0.01.Mon, 01 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-01-github-copilot-ai-credits-billing/https://nilsliu.dev/zh/insights/2026-06-01-github-copilot-ai-credits-billing/GitHub Copilot 6 月 1 日起全面改用 AI Credits 計費，1 credit = $0.01，聊天與 agent 工作流都要扣點。Code completion 維持免費，但重度 agent 使用者可能面臨費用跳升。Mon, 01 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-01-github-copilot-ai-credits-billing/https://nilsliu.dev/de/insights/2026-06-01-github-copilot-ai-credits-billing/GitHub Copilot stellt am 1. Juni auf tokenbasierte KI-Credits-Abrechnung um. Code-Vervollständigungen bleiben kostenlos, aber Chat, Agenten-Workflows und Code-Reviews verbrauchen jetzt Credits. Ein Credit entspricht 0,01 Dollar.Mon, 01 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-06-01-softbank-france-75b-ai-data-center/https://nilsliu.dev/en/insights/2026-06-01-softbank-france-75b-ai-data-center/SoftBank commits up to €75 billion to build 5 GW of AI data centers in France. Phase 1 targets 3.1 GW in Hauts-de-France by 2031, with Schneider Electric and EDF as key partners.Mon, 01 Jun 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-06-01-softbank-france-75b-ai-data-center/https://nilsliu.dev/zh/insights/2026-06-01-softbank-france-75b-ai-data-center/SoftBank 宣布向法國投資高達 750 億歐元建設 AI 資料中心，首期部署 3.1 GW 算力，孫正義與馬克宏聯手，要讓法國成為歐洲 AI 基建核心。Mon, 01 Jun 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-06-01-softbank-france-75b-ai-data-center/https://nilsliu.dev/de/insights/2026-06-01-softbank-france-75b-ai-data-center/SoftBank verpflichtet sich zu Investitionen von bis zu 75 Milliarden Euro in KI-Rechenzentren in Frankreich – 5 GW Gesamtkapazität, Phase 1 bis 2031 mit 3,1 GW in der Region Hauts-de-France.Mon, 01 Jun 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-31-claude-opus-4-8-dynamic-workflows/https://nilsliu.dev/en/insights/2026-05-31-claude-opus-4-8-dynamic-workflows/Anthropic launches Claude Opus 4.8 just 41 days after Opus 4.7, with agentic coding scores up to 69.2%, fast mode pricing cut by two-thirds, and a new dynamic workflows feature running hundreds of parallel sub-agents. Mythos-class models will follow in weeks.Sun, 31 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-31-claude-opus-4-8-dynamic-workflows/https://nilsliu.dev/zh/insights/2026-05-31-claude-opus-4-8-dynamic-workflows/Anthropic 在 Opus 4.7 發布後僅 41 天再推 Opus 4.8，代理編碼評分升至 69.2%，快速模式定價砍至三分之一，並首度開放動態工作流程讓單一任務調度數百個平行代理。Mythos 級別模型也將在數週內跟進。Sun, 31 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-31-claude-opus-4-8-dynamic-workflows/https://nilsliu.dev/de/insights/2026-05-31-claude-opus-4-8-dynamic-workflows/Anthropic veröffentlicht Claude Opus 4.8 nur 41 Tage nach Opus 4.7: Agentic-Coding-Score steigt auf 69,2%, Fast-Mode-Preis sinkt um zwei Drittel, neue dynamische Workflows steuern hunderte parallele Subagenten. Mythos-Klasse-Modelle folgen in Wochen.Sun, 31 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-31-dell-ai-server-revenue-757-percent/https://nilsliu.dev/en/insights/2026-05-31-dell-ai-server-revenue-757-percent/Dell Q1 FY2027: AI-optimized server revenue hits $16.1B (+757% YoY), total revenue $43.8B (+88%), stock surges 33% in best single day since its 2018 return to public markets.Sun, 31 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-31-dell-ai-server-revenue-757-percent/https://nilsliu.dev/zh/insights/2026-05-31-dell-ai-server-revenue-757-percent/Dell Q1 FY27 財報：AI 伺服器營收 161 億美元，年增 757%；總營收 438 億，年增 88%；股價單日收漲 33%。全年 AI 伺服器目標上修至 600 億美元。Sun, 31 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-31-dell-ai-server-revenue-757-percent/https://nilsliu.dev/de/insights/2026-05-31-dell-ai-server-revenue-757-percent/Dell Q1 FY27: KI-Server-Umsatz 16,1 Mrd. (+757% YoY), Gesamtumsatz 43,8 Mrd. (+88%), Aktie +33% – bester Handelstag seit dem Börsen-Comeback 2018.Sun, 31 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-30-anthropic-65b-series-h-965b-valuation/https://nilsliu.dev/en/insights/2026-05-30-anthropic-65b-series-h-965b-valuation/Three months after talks of a $30B round, Anthropic closed $65B at $965B valuation, surpassing OpenAI's $730B and nearing the $1T mark.Sat, 30 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-30-anthropic-65b-series-h-965b-valuation/https://nilsliu.dev/zh/insights/2026-05-30-anthropic-65b-series-h-965b-valuation/同一家公司，三個月前剛談完 300 億，現在以 650 億美元收尾，估值衝上 9650 億。Anthropic 成為全球估值最高的 AI 新創，OpenAI 落後。Sat, 30 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-30-anthropic-65b-series-h-965b-valuation/https://nilsliu.dev/de/insights/2026-05-30-anthropic-65b-series-h-965b-valuation/Drei Monate nach ersten Berichten über eine 30-Milliarden-Runde schloss Anthropic nun 65 Milliarden ab, bei 965 Milliarden Bewertung, über OpenAI.Sat, 30 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-30-demis-hassabis-agi-2029-singularity/https://nilsliu.dev/en/insights/2026-05-30-demis-hassabis-agi-2029-singularity/Google DeepMind CEO Demis Hassabis updated his AGI timeline at Google I/O 2026: 2029 at the earliest, more than five years ahead of his forecast from a year ago. He says we're standing at the foothills of the singularity.Sat, 30 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-30-demis-hassabis-agi-2029-singularity/https://nilsliu.dev/zh/insights/2026-05-30-demis-hassabis-agi-2029-singularity/Google DeepMind 執行長 Demis Hassabis 在 Google I/O 2026 給出最新 AGI 時間表：最快 2029 年，比他一年前的預估提前五年以上。他說人類正站在奇點山腳。Sat, 30 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-30-demis-hassabis-agi-2029-singularity/https://nilsliu.dev/de/insights/2026-05-30-demis-hassabis-agi-2029-singularity/Google-DeepMind-Chef Demis Hassabis hat seinen AGI-Zeitplan beim Google I/O 2026 aktualisiert: Frühestens 2029 – über fünf Jahre früher als seine Prognose vom Vorjahr. Er sagt, wir stehen am Fuß der Singularität.Sat, 30 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-29-cnn-sues-perplexity-copyright/https://nilsliu.dev/en/insights/2026-05-29-cnn-sues-perplexity-copyright/CNN filed its first AI copyright lawsuit against Perplexity, alleging the search startup scraped 17,000+ stories without authorization. The case could reshape how AI companies license news content.Fri, 29 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-29-cnn-sues-perplexity-copyright/https://nilsliu.dev/zh/insights/2026-05-29-cnn-sues-perplexity-copyright/CNN 向紐約聯邦法院控告 Perplexity，指其未授權抓取逾 1.7 萬篇新聞。這是電視網首件 AI 著作權訴訟，判決將直接影響媒體業授權談判生態。Fri, 29 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-29-cnn-sues-perplexity-copyright/https://nilsliu.dev/de/insights/2026-05-29-cnn-sues-perplexity-copyright/CNN reichte die erste KI-Urheberrechtsklage eines TV-Senders gegen Perplexity ein. 17.000 gestohlene Artikel stehen im Zentrum eines Streits, der die Medienbranche verändern könnte.Fri, 29 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-29-cognition-devin-1b-funding/https://nilsliu.dev/en/insights/2026-05-29-cognition-devin-1b-funding/AI coding startup Cognition raised $1B at a $26B valuation, with ARR surging 13x to $492M in 12 months. Its product Devin now writes 90% of the company's own code, with clients including Goldman Sachs, NASA, and Mercedes-Benz.Fri, 29 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-29-cognition-devin-1b-funding/https://nilsliu.dev/zh/insights/2026-05-29-cognition-devin-1b-funding/AI 工程師新創 Cognition 以 260 億美元估值完成 10 億美元融資，年化營收從一年前的 3700 萬美元暴增至 4.92 億美元。Devin 現在負責撰寫公司 90% 的程式碼，客戶涵蓋 Goldman Sachs、NASA、Mercedes-Benz。Fri, 29 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-29-cognition-devin-1b-funding/https://nilsliu.dev/de/insights/2026-05-29-cognition-devin-1b-funding/Das KI-Startup Cognition sammelte 1 Mrd. USD bei 26 Mrd. USD Bewertung. Der ARR wuchs 13-fach auf 492 Mio. USD in 12 Monaten. Devin schreibt nun 90 % des internen Codes – Kunden sind Goldman Sachs, NASA und Mercedes-Benz.Fri, 29 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-28-illinois-ai-safety-bill-sb315/https://nilsliu.dev/en/insights/2026-05-28-illinois-ai-safety-bill-sb315/Illinois passed SB 315 110-0, requiring OpenAI, Anthropic, and Google DeepMind to undergo annual third-party safety audits. Both AI labs actually endorsed the bill. Governor Pritzker plans to sign.Thu, 28 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-28-illinois-ai-safety-bill-sb315/https://nilsliu.dev/zh/insights/2026-05-28-illinois-ai-safety-bill-sb315/伊利諾州議會 110:0 全票通過 SB 315，要求 OpenAI、Anthropic、Google DeepMind 每年接受第三方安全審計。OpenAI 和 Anthropic 雙雙公開背書，州長宣布將簽署。Thu, 28 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-28-illinois-ai-safety-bill-sb315/https://nilsliu.dev/de/insights/2026-05-28-illinois-ai-safety-bill-sb315/Illinois verabschiedete SB 315 mit 110:0 Stimmen: OpenAI, Anthropic und Google DeepMind müssen jährlich externe Sicherheitsaudits durchführen lassen. Beide KI-Labs unterstützten das Gesetz. Gouverneur Pritzker will unterzeichnen.Thu, 28 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-28-mistral-airbus-bmw-industrial-ai/https://nilsliu.dev/en/insights/2026-05-28-mistral-airbus-bmw-industrial-ai/Mistral AI announced industrial AI partnerships with Airbus and BMW today, applying specialized models to crash simulation and aircraft design. European data sovereignty is now a buying criterion, not a nice-to-have.Thu, 28 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-28-mistral-airbus-bmw-industrial-ai/https://nilsliu.dev/zh/insights/2026-05-28-mistral-airbus-bmw-industrial-ai/Mistral AI 今天宣布與空中巴士及 BMW 簽署工業 AI 合作協議。BMW 要用 Mistral 訓練的大型工業模型加速碰撞模擬，空巴五年合約涵蓋商業航空到國防太空。歐洲資料主權正成為採購決策的核心條件。Thu, 28 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-28-mistral-airbus-bmw-industrial-ai/https://nilsliu.dev/de/insights/2026-05-28-mistral-airbus-bmw-industrial-ai/Mistral AI gab heute Industriepartnerschaften mit Airbus und BMW bekannt. Spezialisierte Modelle sollen Crashtests beschleunigen und die Flugzeugentwicklung verbessern. Europäische Datensouveränität wird zum Kaufkriterium.Thu, 28 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-27-china-ai-talent-travel-ban/https://nilsliu.dev/en/insights/2026-05-27-china-ai-talent-travel-ban/Beijing is now controlling when top AI talent at private firms like Alibaba and DeepSeek can leave the country. The US restricted chips; China is restricting people.Wed, 27 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-27-china-ai-talent-travel-ban/https://nilsliu.dev/zh/insights/2026-05-27-china-ai-talent-travel-ban/北京開始管控私人企業的頂尖 AI 人才出境，適用對象從 DeepSeek、Manus 延伸到阿里巴巴。美國對晶片出口設限，中國現在鎖住的是人。Wed, 27 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-27-china-ai-talent-travel-ban/https://nilsliu.dev/de/insights/2026-05-27-china-ai-talent-travel-ban/Peking kontrolliert jetzt, wann KI-Spitzenkräfte bei Privatunternehmen wie Alibaba und DeepSeek das Land verlassen können. Die USA beschränkten Chips – China beschränkt jetzt Menschen.Wed, 27 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-27-nvidia-computex-2026-taiwan-150b/https://nilsliu.dev/en/insights/2026-05-27-nvidia-computex-2026-taiwan-150b/Jensen Huang told a Computex audience that Nvidia's annual spending in Taiwan has surged from $10-15 billion five years ago to $100 billion, heading toward $150 billion. Taiwan's Taiex closed at a record high the same day.Wed, 27 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-27-nvidia-computex-2026-taiwan-150b/https://nilsliu.dev/zh/insights/2026-05-27-nvidia-computex-2026-taiwan-150b/黃仁勳今天 Computex 宣布，Nvidia 年度台灣採購從五年前的 150 億跳升至 1000 億、目標 1500 億美元。台灣加權指數同日創收盤歷史新高，漲 1.7%。Wed, 27 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-27-nvidia-computex-2026-taiwan-150b/https://nilsliu.dev/de/insights/2026-05-27-nvidia-computex-2026-taiwan-150b/Jensen Huang verkündete auf der Computex 2026, dass Nvidias jährliche Ausgaben in Taiwan von 10-15 Mrd. USD auf 100 Mrd. Dollar gestiegen sind, Ziel: 150 Mrd. USD. Der Taiex schloss am selben Tag auf Allzeithoch.Wed, 27 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-26-meta-8000-layoffs-ai-restructuring/https://nilsliu.dev/en/insights/2026-05-26-meta-8000-layoffs-ai-restructuring/Q1 revenue $56B, net income $26.8B—yet 8,000 jobs cut. Meta isn't bleeding; it's converting headcount savings into compute budget, the biggest AI infrastructure bet in tech history.Tue, 26 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-26-meta-8000-layoffs-ai-restructuring/https://nilsliu.dev/zh/insights/2026-05-26-meta-8000-layoffs-ai-restructuring/Q1 營收 563 億美元、淨利 268 億，同一週還是裁了 8000 人。Meta 把省下來的人力預算直接搬去買 GPU 算力，規模衝上 2026 年 1450 億的資本支出，是科技業有史以來最大的 AI 基礎設施豪賭。Tue, 26 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-26-meta-8000-layoffs-ai-restructuring/https://nilsliu.dev/de/insights/2026-05-26-meta-8000-layoffs-ai-restructuring/Q1-Umsatz 56 Mrd. $, Nettogewinn 26,8 Mrd. $ – und trotzdem 8.000 Entlassungen. Meta blutet nicht; es wandelt Personalkosten in Rechenleistungsbudget um, die größte KI-Infrastrukturwette in der Technikgeschichte.Tue, 26 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-25-nvidia-q1-fy2027-agentic-ai-vera-rubin/https://nilsliu.dev/en/insights/2026-05-25-nvidia-q1-fy2027-agentic-ai-vera-rubin/NVIDIA posted Q1 FY2027 revenue of $81.6B, up 85% year-over-year, with data center nearly doubling. Jensen Huang declared 'Agentic AI has arrived' and unveiled Vera Rubin timelines. The stock dipped despite records, exposing the expectations trap every hypergrowth company eventually hits.Mon, 25 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-25-nvidia-q1-fy2027-agentic-ai-vera-rubin/https://nilsliu.dev/zh/insights/2026-05-25-nvidia-q1-fy2027-agentic-ai-vera-rubin/NVIDIA 公布 FY2027 Q1 財報，收入 816 億美元年增 85%，資料中心年增 92%。黃仁勳在法說會宣告「Agentic AI 已到來」，並揭露 Vera Rubin 平台出貨時程。股價盤後反跌的真正原因，是市場預期早就把超預期定價進去了。Mon, 25 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-25-nvidia-q1-fy2027-agentic-ai-vera-rubin/https://nilsliu.dev/de/insights/2026-05-25-nvidia-q1-fy2027-agentic-ai-vera-rubin/NVIDIA meldet Q1 FY2027-Umsatz von 81,6 Mrd. USD, ein Plus von 85% im Jahresvergleich. Jensen Huang erklärte, dass 'Agentic AI angekommen' sei, und stellte den Vera-Rubin-Zeitplan vor. Die Aktie fiel trotz Rekordwerten, ein Lehrbeispiel für die Erwartungsfalle.Mon, 25 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-25-pope-leo-xiv-magnifica-humanitas-ai/https://nilsliu.dev/en/insights/2026-05-25-pope-leo-xiv-magnifica-humanitas-ai/Pope Leo XIV released his first AI-focused encyclical on May 25, presenting alongside Anthropic co-founder Christopher Olah. The document warns of AI-driven dehumanization and calls human dignity the fundamental criterion for evaluating AI development.Mon, 25 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-25-pope-leo-xiv-magnifica-humanitas-ai/https://nilsliu.dev/zh/insights/2026-05-25-pope-leo-xiv-magnifica-humanitas-ai/5 月 25 日，教宗良十四世在梵蒂岡發表首份以 AI 為主題的通諭，邀請 Anthropic 共同創辦人同台，警告 AI「去人化」風險，要求人的尊嚴作為評估 AI 發展的最高準則。Mon, 25 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-25-pope-leo-xiv-magnifica-humanitas-ai/https://nilsliu.dev/de/insights/2026-05-25-pope-leo-xiv-magnifica-humanitas-ai/Papst Leo XIV. hat am 25. Mai sein erstes KI-Lehrschreiben veröffentlicht und dabei Anthropic-Mitgründer Christopher Olah als Mitpräsentator eingeladen. Das Dokument warnt vor Entmenschlichung durch KI und setzt Menschenwürde als obersten Maßstab.Mon, 25 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-24-deepseek-v4-pro-price-cut/https://nilsliu.dev/en/insights/2026-05-24-deepseek-v4-pro-price-cut/DeepSeek has made its 75% V4-Pro API discount permanent, pricing output tokens at $0.87 per million, 34× cheaper than GPT-5.5. This isn't just a price cut; it's a direct attack on Western AI pricing power.Sun, 24 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-24-deepseek-v4-pro-price-cut/https://nilsliu.dev/zh/insights/2026-05-24-deepseek-v4-pro-price-cut/DeepSeek 宣布將 V4-Pro API 的 75% 折扣永久化，輸出 Token 每百萬僅 $0.87，比 GPT-5.5 便宜 34 倍。這不只是降價，是一場針對西方 AI 定價護城河的定點清除。Sun, 24 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-24-deepseek-v4-pro-price-cut/https://nilsliu.dev/de/insights/2026-05-24-deepseek-v4-pro-price-cut/DeepSeek macht seinen 75%-Rabatt auf die V4-Pro-API dauerhaft: Output-Token zu $0,87 pro Million, 34-mal günstiger als GPT-5.5. Das ist kein Sonderangebot, sondern ein gezielter Angriff auf die Preismacht westlicher KI-Labore.Sun, 24 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-24-teampcp-supply-chain-github-openai/https://nilsliu.dev/en/insights/2026-05-24-teampcp-supply-chain-github-openai/A malicious Nx Console VS Code extension stayed live for just 18 minutes, yet TeamPCP managed to steal 3,800 GitHub internal repos, compromise two OpenAI employee devices, and put Mistral's source code up for sale on dark web forums.Sun, 24 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-24-teampcp-supply-chain-github-openai/https://nilsliu.dev/zh/insights/2026-05-24-teampcp-supply-chain-github-openai/TeamPCP 組織用一個只活了 18 分鐘的毒 VS Code 擴充套件，讓 GitHub 損失 3,800 個內部庫、OpenAI 兩台設備遭入侵、Mistral 原始碼在暗網公開兜售。AI 開發環境已成高價值目標。Sun, 24 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-24-teampcp-supply-chain-github-openai/https://nilsliu.dev/de/insights/2026-05-24-teampcp-supply-chain-github-openai/Eine vergiftete Nx Console VS Code-Erweiterung war nur 18 Minuten aktiv, doch TeamPCP stahl 3.800 GitHub-interne Repositories, kompromittierte zwei OpenAI-Mitarbeitergeräte und bot Mistrals Quellcode in Untergrundforen an.Sun, 24 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-23-glasswing-10000-vulnerabilities/https://nilsliu.dev/en/insights/2026-05-23-glasswing-10000-vulnerabilities/Anthropic releases Project Glasswing month-one results: Claude Mythos Preview found 10,000+ high/critical vulnerabilities across 1,000 open-source projects, with a 90.6% validation rate. The new bottleneck is patching, not discovery.Sat, 23 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-23-glasswing-10000-vulnerabilities/https://nilsliu.dev/zh/insights/2026-05-23-glasswing-10000-vulnerabilities/Anthropic 公布 Project Glasswing 首月成果：Claude Mythos Preview 在 1,000 個開源專案中找到逾一萬個高危漏洞，90.6% 確認為真實威脅。現在的瓶頸，已從發現轉移到修補。Sat, 23 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-23-glasswing-10000-vulnerabilities/https://nilsliu.dev/de/insights/2026-05-23-glasswing-10000-vulnerabilities/Anthropic veröffentlicht die Ergebnisse des ersten Monats von Project Glasswing: Claude Mythos Preview entdeckte über 10.000 kritische Sicherheitslücken in 1.000 Open-Source-Projekten mit einer Bestätigungsrate von 90,6%.Sat, 23 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-23-openai-singapore-applied-ai-lab/https://nilsliu.dev/en/insights/2026-05-23-openai-singapore-applied-ai-lab/OpenAI commits $234M to Singapore for its first overseas Applied AI Lab. IMDA updates its agentic AI governance framework the same day.Sat, 23 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-23-openai-singapore-applied-ai-lab/https://nilsliu.dev/zh/insights/2026-05-23-openai-singapore-applied-ai-lab/OpenAI 在新加坡設立首個海外 Applied AI Lab，投入逾 2.34 億美元，IMDA 同日更新代理 AI 治理框架，Google 同場宣布國家 AI 夥伴關係。Sat, 23 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-23-openai-singapore-applied-ai-lab/https://nilsliu.dev/de/insights/2026-05-23-openai-singapore-applied-ai-lab/OpenAI investiert 234 Mio. USD für erstes internationales AI Lab in Singapur. IMDA aktualisiert gleichzeitig den KI-Governance-Rahmen.Sat, 23 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-22-openai-ipo-s1-filing/https://nilsliu.dev/en/insights/2026-05-22-openai-ipo-s1-filing/OpenAI filed a confidential S-1 with the SEC on May 22, targeting a $1 trillion valuation with Goldman Sachs and Morgan Stanley underwriting. With $2B monthly revenue and 900M+ weekly users, this could be tech's largest-ever IPO.Fri, 22 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-22-openai-ipo-s1-filing/https://nilsliu.dev/zh/insights/2026-05-22-openai-ipo-s1-filing/OpenAI 今日向 SEC 遞交機密 S-1 申請，目標估值一兆美元，由高盛和摩根士丹利主承銷。月收入 20 億美元、9 億週活躍用戶，這可能是科技史上最大規模的 IPO。Fri, 22 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-22-openai-ipo-s1-filing/https://nilsliu.dev/de/insights/2026-05-22-openai-ipo-s1-filing/OpenAI hat am 22. Mai eine vertrauliche S-1-Anmeldung bei der SEC eingereicht und strebt eine Bewertung von 1 Billion Dollar an. Goldman Sachs und Morgan Stanley begleiten den Prozess. Mit monatlichen Einnahmen von 2 Milliarden Dollar und 900 Millionen Wochennutzern könnte dies der größte Tech-Börsengang aller Zeiten werden.Fri, 22 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-22-trump-ai-executive-order-postponed/https://nilsliu.dev/en/insights/2026-05-22-trump-ai-executive-order-postponed/Hours before the scheduled signing ceremony, Trump pulled an executive order that would have created a voluntary 90-day government AI model review. Musk and Zuckerberg opposed it overnight, and the White House backed down.Fri, 22 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-22-trump-ai-executive-order-postponed/https://nilsliu.dev/zh/insights/2026-05-22-trump-ai-executive-order-postponed/5 月 21 日，距簽署只剩幾小時，Trump 叫停 AI 模型 90 天政府預審的行政命令。Musk 和 Zuckerberg 當晚進言反對，白宮的安全盤算再次輸給矽谷的速度焦慮。Fri, 22 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-22-trump-ai-executive-order-postponed/https://nilsliu.dev/de/insights/2026-05-22-trump-ai-executive-order-postponed/Stunden vor der geplanten Unterzeichnungszeremonie stoppte Trump eine KI-Exekutivorder für eine freiwillige 90-Tage-Modellprüfung. Musk und Zuckerberg hatten Trump über Nacht ihre Ablehnung mitgeteilt.Fri, 22 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-21-anthropic-revenue-spacex-compute/https://nilsliu.dev/en/insights/2026-05-21-anthropic-revenue-spacex-compute/Anthropic's Q2 revenue is set to more than double to $10.9B, while a $1.25B/month SpaceX compute deal signals a massive infrastructure bet to power Claude's rapid growth.Thu, 21 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-21-anthropic-revenue-spacex-compute/https://nilsliu.dev/zh/insights/2026-05-21-anthropic-revenue-spacex-compute/Anthropic Q2 季度營收預估從 48 億暴增至 109 億美元，同時與 SpaceX 簽下每月 12.5 億美元算力合約。這份三年期大單的年金額，幾乎等同 SpaceX 全公司一整年的收入。Thu, 21 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-21-anthropic-revenue-spacex-compute/https://nilsliu.dev/de/insights/2026-05-21-anthropic-revenue-spacex-compute/Anthropics Q2-Umsatz soll sich auf 10,9 Mrd. $ mehr als verdoppeln. Ein SpaceX-Rechenvertrag von 1,25 Mrd. $/Monat sichert die Infrastruktur für Claudes rasantes Wachstum.Thu, 21 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-21-openai-erdos-geometry-conjecture/https://nilsliu.dev/en/insights/2026-05-21-openai-erdos-geometry-conjecture/OpenAI's general-purpose reasoning model disproved the Erdős unit distance conjecture — a problem open for 78 years — with no task-specific training. The proof was verified by Fields Medalist Tim Gowers and Princeton's Noga Alon.Thu, 21 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-21-openai-erdos-geometry-conjecture/https://nilsliu.dev/zh/insights/2026-05-21-openai-erdos-geometry-conjecture/OpenAI 的通用推理模型推翻了困擾數學界 80 年的埃爾德什單位距離猜想，這是 AI 首次在沒有專門訓練的情況下，獨力解決活躍數學領域中的公開難題。Thu, 21 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-21-openai-erdos-geometry-conjecture/https://nilsliu.dev/de/insights/2026-05-21-openai-erdos-geometry-conjecture/OpenAIs allgemeines Reasoning-Modell hat die Erdős-Einheitsabstands-Vermutung widerlegt — ein 78 Jahre altes offenes Problem der diskreten Geometrie. Der Beweis wurde von Fields-Medaillengewinner Tim Gowers und Noga Alon (Princeton) bestätigt.Thu, 21 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/token-maxi-automate-zh/https://nilsliu.dev/en/insights/token-maxi-automate-zh/Google I/O 2026's chart reveals one truth: AI usage isn't growing because people type more. 3.2 quadrillion tokens/month, 7x Y/Y growth — behind that number are automated pipelines running without pause. The question isn't whether you use AI, but whether AI is automatically working for you.Thu, 21 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/token-maxi-automate-zh/https://nilsliu.dev/zh/insights/token-maxi-automate-zh/Google I/O 2026 那張圖說明了一件事：AI 的使用量不是靠人工輸入長大的。每月 3.2 quadrillion token、7 倍年增長，背後是無數條自動化 pipeline 在不間斷地運行。問題不是「你會不會用 AI」，而是「你有沒有讓 AI 自動替你工作」。Thu, 21 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/token-maxi-automate-zh/https://nilsliu.dev/de/insights/token-maxi-automate-zh/Google I/O 2026 zeigt: KI-Nutzung wächst nicht durch mehr manuelle Eingaben. 3,2 Billiarden Tokens pro Monat, 7-faches Jahreswachstum — dahinter stecken automatisierte Pipelines. Die Frage ist nicht ob, sondern wie automatisch die KI für dich arbeitet.Thu, 21 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-20-google-io-gemini-spark/https://nilsliu.dev/en/insights/2026-05-20-google-io-gemini-spark/At Google I/O 2026, Google overhauled Search for the first time in 25 years and launched Gemini Spark, a 24/7 personal AI agent. With 900 million Gemini users, the AI race just shifted into a new gear.Wed, 20 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-20-google-io-gemini-spark/https://nilsliu.dev/zh/insights/2026-05-20-google-io-gemini-spark/Google 在年度開發者大會宣布搜尋引擎 25 年最大改版，推出 Gemini Spark 全天候個人 AI 代理人。Gemini 月活用戶已達 9 億，你的搜尋習慣即將徹底翻轉。Wed, 20 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-20-google-io-gemini-spark/https://nilsliu.dev/de/insights/2026-05-20-google-io-gemini-spark/Bei Google I/O 2026 überarbeitete Google seine Suchmaschine erstmals seit 25 Jahren grundlegend und startete Gemini Spark, einen rund um die Uhr verfügbaren KI-Assistenten. Mit 900 Millionen Gemini-Nutzern verschärft sich der KI-Wettbewerb deutlich.Wed, 20 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-20-openai-google-synthid-ai-content-provenance/https://nilsliu.dev/en/insights/2026-05-20-openai-google-synthid-ai-content-provenance/OpenAI adopts C2PA and integrates Google's SynthID invisible watermark into ChatGPT images, plus a new public verification tool. Two AI rivals team up on deepfake detection, but does it actually work?Wed, 20 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-20-openai-google-synthid-ai-content-provenance/https://nilsliu.dev/zh/insights/2026-05-20-openai-google-synthid-ai-content-provenance/OpenAI 宣布採用 C2PA 標準並整合 Google 的 SynthID 不可見水印技術，同時推出公開驗證工具。兩大 AI 競爭對手罕見攜手，能否真正解決 AI 深偽問題？Wed, 20 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-20-openai-google-synthid-ai-content-provenance/https://nilsliu.dev/de/insights/2026-05-20-openai-google-synthid-ai-content-provenance/OpenAI adoptiert C2PA und integriert Googles SynthID-Wasserzeichen in ChatGPT-Bilder und lanciert ein öffentliches Prüftool. Zwei KI-Rivalen gegen Deepfakes, aber löst das das Problem wirklich?Wed, 20 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-19-anthropic-stainless-sdk-acquisition/https://nilsliu.dev/en/insights/2026-05-19-anthropic-stainless-sdk-acquisition/Anthropic paid over $300M for Stainless, the SDK automation tool used by OpenAI, Google, and Cloudflare, and is shutting down its hosted service for outside customers. This isn't a cost-saving move. It's a play for the connectivity layer of the agentic AI era.Tue, 19 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-19-anthropic-stainless-sdk-acquisition/https://nilsliu.dev/zh/insights/2026-05-19-anthropic-stainless-sdk-acquisition/Anthropic 以逾 3 億美元收購 SDK 自動化工具商 Stainless，這個 OpenAI、Google、Cloudflare 都在用的開發者基礎設施從此易主，並宣布關閉所有對外服務。AI 代理時代的連接層競爭，現在才剛開打。Tue, 19 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-19-anthropic-stainless-sdk-acquisition/https://nilsliu.dev/de/insights/2026-05-19-anthropic-stainless-sdk-acquisition/Anthropic hat das SDK-Automatisierungsstartup Stainless für über 300 Millionen Dollar übernommen und kündigt die Abschaltung aller externen Dienste an. OpenAI und Google nutzten es täglich. Dahinter steckt ein Kampf um die Verbindungsschicht im KI-Agenten-Zeitalter.Tue, 19 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-19-google-io-2026-keynote-gemini-intelligence/https://nilsliu.dev/en/insights/2026-05-19-google-io-2026-keynote-gemini-intelligence/Google I/O 2026 keynote unveils Gemini Intelligence embedded at the Android OS layer, a new Googlebook laptop category, and Samsung XR glasses. Google bets on distribution, not model rankings.Tue, 19 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-19-google-io-2026-keynote-gemini-intelligence/https://nilsliu.dev/zh/insights/2026-05-19-google-io-2026-keynote-gemini-intelligence/Google I/O 2026 主題演講發布 Gemini Intelligence 嵌入 Android 系統層、Googlebook 筆電新品類，以及三星合作 XR 眼鏡，模型排名雖落後 Claude Mythos，但 Google 押注的是分發規模。Tue, 19 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-19-google-io-2026-keynote-gemini-intelligence/https://nilsliu.dev/de/insights/2026-05-19-google-io-2026-keynote-gemini-intelligence/Google I/O 2026 präsentiert Gemini Intelligence auf Betriebssystemebene, die neue Googlebook-Laptopkategorie und Samsung-XR-Brillen. Google setzt auf Verteilung, nicht auf Modell-Rankings.Tue, 19 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-18-ai-replaces-relay-workers-not-bankers/https://nilsliu.dev/en/insights/2026-05-18-ai-replaces-relay-workers-not-bankers/In banking, AI won't target judgment-heavy roles first. It's going after the relay chain—the people moving data from A to B. How much of your day is actually forwarding?Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-18-ai-replaces-relay-workers-not-bankers/https://nilsliu.dev/zh/insights/2026-05-18-ai-replaces-relay-workers-not-bankers/在銀行業，AI 最先動手的不是需要判斷力的崗位，而是那條把資料從 A 搬到 B 的傳遞鏈。你每天花多少時間在搬運？Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-18-ai-replaces-relay-workers-not-bankers/https://nilsliu.dev/de/insights/2026-05-18-ai-replaces-relay-workers-not-bankers/In der Bankbranche greift KI nicht zuerst urteilsintensive Stellen an. Sie zielt auf die Weiterleitungskette ab – auf die Menschen, die Daten von A nach B bewegen. Wie viel Ihres Tages ist eigentlich Weiterleitung?Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-18-ai-translator-between-business-tech-process/https://nilsliu.dev/en/insights/2026-05-18-ai-translator-between-business-tech-process/Cross-domain translation is becoming the defining competitive skill of the AI era — not because it sounds good, but because once AI takes over everything that requires only one language, what remains is all translation work.Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-18-ai-translator-between-business-tech-process/https://nilsliu.dev/zh/insights/2026-05-18-ai-translator-between-business-tech-process/跨界能力會變成新的核心競爭力。不是因為這話聽起來好，而是因為 AI 接管了所有「只需要一種語言」的工作之後，剩下的全是翻譯問題。Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-18-ai-translator-between-business-tech-process/https://nilsliu.dev/de/insights/2026-05-18-ai-translator-between-business-tech-process/Die fachübergreifende Übersetzungsfähigkeit wird zur entscheidenden Kompetenz im KI-Zeitalter – nicht weil es gut klingt, sondern weil nach der KI-Übernahme alles, was nur eine Sprache erfordert, nur noch Übersetzungsarbeit übrigbleibt.Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-18-anthropic-project-glasswing-claude-mythos/https://nilsliu.dev/en/insights/2026-05-18-anthropic-project-glasswing-claude-mythos/Anthropic's unreleased Claude Mythos Preview has autonomously discovered thousands of zero-day vulnerabilities across major OSes and browsers. 12 tech giants joined as defenders, but the model was accessed without authorization on day one.Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-18-anthropic-project-glasswing-claude-mythos/https://nilsliu.dev/zh/insights/2026-05-18-anthropic-project-glasswing-claude-mythos/Anthropic 最強未公開模型 Claude Mythos Preview，已在主流作業系統和瀏覽器找到數千個零日漏洞。12 家科技巨頭聯手防禦，但上線第一天就遭未授權存取。Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-18-anthropic-project-glasswing-claude-mythos/https://nilsliu.dev/de/insights/2026-05-18-anthropic-project-glasswing-claude-mythos/Anthropics unveröffentlichtes Claude Mythos Preview entdeckte eigenständig Tausende Zero-Day-Schwachstellen in führenden Betriebssystemen. 12 Technologieriesen kooperieren als Verteidiger, doch das Modell wurde am ersten Tag unberechtigt abgerufen.Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-18-google-io-2026-gemini-omni/https://nilsliu.dev/en/insights/2026-05-18-google-io-2026-gemini-omni/Google I/O 2026 opens tomorrow. Leaked Gemini Omni promises unified text/image/video generation — but can it catch Claude Mythos scoring 93.9% on SWE-bench?Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-18-google-io-2026-gemini-omni/https://nilsliu.dev/zh/insights/2026-05-18-google-io-2026-gemini-omni/明天 Google I/O 開幕，洩漏的 Gemini Omni 整合文字、圖片、影片生成，但能否追上 Claude Mythos 的 93.9% SWE-bench 分數？Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-18-google-io-2026-gemini-omni/https://nilsliu.dev/de/insights/2026-05-18-google-io-2026-gemini-omni/Google I/O 2026 öffnet morgen. Das geleakte Gemini Omni verspricht einheitliche Text/Bild/Video-Generierung – aber kann es Claude Mythos mit 93,9% SWE-bench einholen?Mon, 18 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-17-anthropic-30b-funding-900b-valuation/https://nilsliu.dev/en/insights/2026-05-17-anthropic-30b-funding-900b-valuation/Three months ago it was valued at $380B. Now it's closing a $30B round at $900B. Is Anthropic's latest fundraise rational market pricing, or the next AI bubble peak?Sun, 17 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-17-anthropic-30b-funding-900b-valuation/https://nilsliu.dev/zh/insights/2026-05-17-anthropic-30b-funding-900b-valuation/三個月前估值 3800 億，現在談 9000 億。Anthropic 的新一輪 300 億美元融資，到底是市場理性還是另一波估值泡沫？Sun, 17 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-17-anthropic-30b-funding-900b-valuation/https://nilsliu.dev/de/insights/2026-05-17-anthropic-30b-funding-900b-valuation/Vor drei Monaten 380 Milliarden, jetzt 900 Milliarden: Ist Anthropics neueste Finanzierungsrunde rationale Marktbewertung – oder der nächste KI-Bewertungsgipfel?Sun, 17 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-17-openai-brockman-product-strategy-agentic/https://nilsliu.dev/en/insights/2026-05-17-openai-brockman-product-strategy-agentic/OpenAI co-founder Greg Brockman officially takes charge of product strategy, merging ChatGPT, Codex, and the developer API into one Agentic platform. A major reorganization timed four days before Google I/O.Sun, 17 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-17-openai-brockman-product-strategy-agentic/https://nilsliu.dev/zh/insights/2026-05-17-openai-brockman-product-strategy-agentic/OpenAI 共同創辦人 Greg Brockman 正式接掌產品策略，將 ChatGPT、Codex 與 API 合併為單一 Agentic 平台。Google I/O 開幕前四天的這場整合，是守住 9 億用戶優勢的關鍵一棋嗎？Sun, 17 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-17-openai-brockman-product-strategy-agentic/https://nilsliu.dev/de/insights/2026-05-17-openai-brockman-product-strategy-agentic/OpenAI-Mitgründer Greg Brockman übernimmt offiziell die Produktstrategie und führt ChatGPT, Codex und Developer-API zu einer Agentic-Plattform zusammen – vier Tage vor Google I/O.Sun, 17 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-16-cerebras-ipo-nasdaq/https://nilsliu.dev/en/insights/2026-05-16-cerebras-ipo-nasdaq/Cerebras priced at $185, raised $5.55B, and surged 68% on debut to a $95B market cap. Its WSE-3 chip runs inference 15x faster than GPUs, with OpenAI and AWS already on board. A wave of AI IPOs is now on its way.Sat, 16 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-16-cerebras-ipo-nasdaq/https://nilsliu.dev/zh/insights/2026-05-16-cerebras-ipo-nasdaq/Cerebras 以 185 美元定價、募資 55.5 億美元，首日收漲 68%，市值衝上 950 億美元。WSE-3 晶片比 GPU 快 15 倍，OpenAI 和 AWS 都是客戶，AI 晶片市場正在重組。Sat, 16 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-16-cerebras-ipo-nasdaq/https://nilsliu.dev/de/insights/2026-05-16-cerebras-ipo-nasdaq/Cerebras startete bei 185 Dollar, sammelte 5,55 Milliarden ein und legte am ersten Handelstag 68 Prozent zu. Der WSE-3-Chip ist bis zu 15-mal schneller als GPUs beim Inferencing, OpenAI und AWS sind bereits Kunden.Sat, 16 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-16-openai-chatgpt-personal-finance-plaid/https://nilsliu.dev/en/insights/2026-05-16-openai-chatgpt-personal-finance-plaid/OpenAI partners with Plaid to let ChatGPT Pro users link 12,000+ financial institutions. Spending analysis, portfolio tracking, financial planning — but users are asking hard privacy questions.Sat, 16 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-16-openai-chatgpt-personal-finance-plaid/https://nilsliu.dev/zh/insights/2026-05-16-openai-chatgpt-personal-finance-plaid/OpenAI 與 Plaid 合作，讓 ChatGPT Pro 用戶連結逾 1.2 萬家金融機構，AI 直接讀取消費記錄與投資組合。功能強大，但用戶的第一反應是：這樣真的安全嗎？Sat, 16 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-16-openai-chatgpt-personal-finance-plaid/https://nilsliu.dev/de/insights/2026-05-16-openai-chatgpt-personal-finance-plaid/OpenAI und Plaid ermöglichen ChatGPT Pro-Nutzern, über 12.000 Finanzinstitute zu verbinden. Ausgabenanalyse, Portfolioverfolgung und Finanzplanung, aber Nutzer stellen kritische Datenschutzfragen.Sat, 16 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-15-anthropic-gates-foundation-ai-health/https://nilsliu.dev/en/insights/2026-05-15-anthropic-gates-foundation-ai-health/Anthropic and the Gates Foundation commit $200 million over four years to deploy Claude in global health, education, and agriculture. Both sides have pledged to make results publicly verifiable.Fri, 15 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-15-anthropic-gates-foundation-ai-health/https://nilsliu.dev/zh/insights/2026-05-15-anthropic-gates-foundation-ai-health/Anthropic 與蓋茲基金會宣布四年 2 億美元合作，目標把 Claude 部署到全球醫療、教育、農業前線。兩邊都承諾公開成果資料，AI 公益項目很少做到這點。Fri, 15 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-15-anthropic-gates-foundation-ai-health/https://nilsliu.dev/de/insights/2026-05-15-anthropic-gates-foundation-ai-health/Anthropic und die Gates Foundation verpflichten sich zu 200 Millionen Dollar über vier Jahre für Claude-Einsatz in Gesundheit, Bildung und Landwirtschaft, mit öffentlich nachprüfbaren Ergebnissen.Fri, 15 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-15-nvidia-h200-china-trump-xi-summit/https://nilsliu.dev/en/insights/2026-05-15-nvidia-h200-china-trump-xi-summit/The Trump-Xi summit unlocked H200 export licenses for Alibaba, Tencent, ByteDance and 7 others, but Beijing told them not to buy, and zero chips have shipped.Fri, 15 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-15-nvidia-h200-china-trump-xi-summit/https://nilsliu.dev/zh/insights/2026-05-15-nvidia-h200-china-trump-xi-summit/美中峰會給 Nvidia 開了一扇門：Alibaba、Tencent、ByteDance 等十家企業獲准購買 H200，每家上限 75,000 顆。但北京暗示不准買，晶片一顆都沒出貨。Fri, 15 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-15-nvidia-h200-china-trump-xi-summit/https://nilsliu.dev/de/insights/2026-05-15-nvidia-h200-china-trump-xi-summit/Der Trump-Xi-Gipfel öffnete H200-Exportlizenzen für Alibaba, Tencent, ByteDance und 7 weitere, doch Peking riet vom Kauf ab, und kein einziger Chip wurde geliefert.Fri, 15 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2026-05-14-openai-altman-musk-trial/https://nilsliu.dev/en/insights/2026-05-14-openai-altman-musk-trial/Court filings show OpenAI CEO Sam Altman holds over $2B in personal stakes across nine companies that did business with OpenAI, $1.7B in Helion alone. Closing arguments begin today in the landmark Musk lawsuit that could reshape AI's most powerful company.Thu, 14 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/2026-05-14-openai-altman-musk-trial/https://nilsliu.dev/zh/insights/2026-05-14-openai-altman-musk-trial/法庭文件顯示，OpenAI 執行長 Sam Altman 在與 OpenAI 有業務往來的九家公司中持有逾 20 億美元個人股份，Helion Energy 一項就佔 17 億美元。馬斯克訴訟今日進入結辯，可能影響這家估值逾 1 兆美元 AI 龍頭的上市計畫與治理結構。Thu, 14 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/2026-05-14-openai-altman-musk-trial/https://nilsliu.dev/de/insights/2026-05-14-openai-altman-musk-trial/Gerichtsakten zeigen: OpenAI-Chef Sam Altman hält über 2 Milliarden Dollar persönliche Beteiligungen an neun Unternehmen, die Geschäfte mit OpenAI gemacht haben, allein 1,7 Milliarden in Helion Energy. Heute beginnen die Schlussplädoyers im wegweisenden Musk-Prozess.Thu, 14 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/thinking-machine-voice-bitter-lesson/https://nilsliu.dev/en/insights/thinking-machine-voice-bitter-lesson/Thinking Machines' TML-Interaction-Small hits 0.40s turn-taking latency — 3x faster than OpenAI — by scrapping the pipeline architecture entirely and letting the model learn interactivity at scale. Here's what that actually means.Thu, 14 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/thinking-machine-voice-bitter-lesson/https://nilsliu.dev/zh/insights/thinking-machine-voice-bitter-lesson/Mira Murati 創辦的 Thinking Machines 在 2026 年 5 月 11 日發布 TML-Interaction-Small，延遲 0.40 秒，比 OpenAI 快三倍。這不是靠更快的硬體——而是他們從根本上放棄了管線架構，讓模型自己學會怎麼對話。這個選擇解釋了一切。Thu, 14 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/thinking-machine-voice-bitter-lesson/https://nilsliu.dev/de/insights/thinking-machine-voice-bitter-lesson/Thinking Machines' TML-Interaction-Small erreicht 0,40 Sekunden Latenz — 3x schneller als OpenAI — indem es die Pipeline-Architektur vollständig verwirft und Interaktivität durch skaliertes Training lernt.Thu, 14 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/anthropic-code-with-claude-2026/https://nilsliu.dev/en/insights/anthropic-code-with-claude-2026/In May 2026, Anthropic hosted Code with Claude 2026 across San Francisco, London, and Tokyo. No new model. Instead: a compute deal with SpaceX, three major agent platform upgrades, and a set of Claude Code features aimed at removing the humans-in-the-loop. The message was clear — AI's next fight is deployment, not benchmarks.Fri, 08 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/anthropic-code-with-claude-2026/https://nilsliu.dev/zh/insights/anthropic-code-with-claude-2026/2026 年 5 月，Anthropic 在舊金山、倫敦、東京三城舉辦「Code with Claude 2026」全球開發者大會。沒有新模型發布，而是一套涵蓋算力、智能體、編程工具與成本的整合更新——AI 的戰場從模型參數轉向了能不能真的用起來。Fri, 08 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/anthropic-code-with-claude-2026/https://nilsliu.dev/de/insights/anthropic-code-with-claude-2026/Im Mai 2026 fand Code with Claude 2026 in San Francisco, London und Tokio statt. Kein neues Modell — stattdessen ein Compute-Deal mit SpaceX, drei Agent-Plattform-Upgrades und Claude Code-Features, die menschliche Eingriffe weiter reduzieren. Der Wettbewerb hat sich verschoben: weg von Benchmarks, hin zu produktionstauglicher KI.Fri, 08 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/future-ai-pm-skills-check/https://nilsliu.dev/en/insights/future-ai-pm-skills-check/This isn't a quiz about RAG or prompt techniques — it tests your judgment in the messy, ambiguous situations that define real AI product work. 20 scenario-based questions across 5 core PM dimensions: Product Thinking, AI Fluency, Data Literacy, Stakeholder Communication, and Execution.Sun, 03 May 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/future-ai-pm-skills-check/https://nilsliu.dev/zh/insights/future-ai-pm-skills-check/不是測試你懂不懂 RAG 或 Prompt Engineering，而是問你：在真實的 AI 產品工作中，你的判斷力夠嗎？全新的未來 AI PM 技能檢核，用 20 道情境題診斷你的 5 大核心能力——產品思維、AI 素養、數據分析、溝通協作、執行力。Sun, 03 May 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/future-ai-pm-skills-check/https://nilsliu.dev/de/insights/future-ai-pm-skills-check/Kein Quiz über RAG oder Prompts — es fragt: Ist dein Urteilsvermögen in echten KI-Produktszenarien bereit? 20 szenariobasierte Fragen in 5 Kerndimensionen.Sun, 03 May 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/product-management-ai-exponential/https://nilsliu.dev/en/insights/product-management-ai-exponential/Anthropic's Cat Wu describes a new PM rhythm in the AI era: roles merging, prototypes over docs, iteration in days not months. Reading it brought back memories of my own undefined role in an enterprise AI team—and Peter Deng's Avengers-style team philosophy.Tue, 28 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/product-management-ai-exponential/https://nilsliu.dev/zh/insights/product-management-ai-exponential/Anthropic 產品負責人 Cat Wu 描述了 AI 時代產品管理的全新節奏：角色正在融合、Prototype 優先於文件、以數天而非數月為迭代週期。讀完這篇文章，我深刻回想起兩年前自己在企業 AI 團隊中「什麼都做」的那段歲月，以及 Peter Deng 談到的「復仇者聯盟」式團隊建構哲學。Tue, 28 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/product-management-ai-exponential/https://nilsliu.dev/de/insights/product-management-ai-exponential/Anthropics Cat Wu beschreibt einen neuen PM-Rhythmus im KI-Zeitalter: Rollen verschmelzen, Prototypen ersetzen Dokumente, Iterationen dauern Tage statt Monate. Das Lesen ließ mich an meine eigene undefinierte Rolle in einem Unternehmens-KI-Team zurückdenken – und an Peter Dengs Avengers-Philosophie.Tue, 28 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/tsmc-stock-surge-ai-chips-2025-2026/https://nilsliu.dev/en/insights/tsmc-stock-surge-ai-chips-2025-2026/TSMC's stock surged 137% from ~$164 in April 2025 to $387 in April 2026. This post breaks down how AI chip demand, CoWoS bottlenecks, and NVIDIA dethroning Apple as top customer drove the run.Mon, 27 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/tsmc-stock-surge-ai-chips-2025-2026/https://nilsliu.dev/zh/insights/tsmc-stock-surge-ai-chips-2025-2026/台積電股價從 2025 年 4 月的約 164 美元漲到 2026 年 4 月的 387 美元，漲幅 137%。這篇拆解 AI 晶片需求、CoWoS 封裝瓶頸、NVIDIA 取代蘋果成為第一大客戶，以及 165 億美元的美國建廠計畫如何共同推動這波行情。Mon, 27 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/tsmc-stock-surge-ai-chips-2025-2026/https://nilsliu.dev/de/insights/tsmc-stock-surge-ai-chips-2025-2026/TSMC-Aktie stieg 137% von ~164 $ im April 2025 auf 387 $ im April 2026. Dieser Beitrag erklärt, wie KI-Chip-Nachfrage, CoWoS-Engpässe und NVIDIA als neue Nummer-1-Kunde den Kurs trieben.Mon, 27 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/agent-harness-production-ai/https://nilsliu.dev/en/insights/agent-harness-production-ai/The 2026 AI race is fundamentally about Harness engineering. This deep dive covers the 12 core modules of a production-grade Agent Harness, leading framework philosophies, and the 7 architectural decisions every AI architect must face.Fri, 24 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/agent-harness-production-ai/https://nilsliu.dev/zh/insights/agent-harness-production-ai/2026 年的 AI 競爭本質是 Harness 工程的博弈。本文深度解析生產級 Agent Harness 的十二大核心模組、主流框架設計哲學，以及 AI 架構師必須面對的七大架構抉擇。Fri, 24 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/agent-harness-production-ai/https://nilsliu.dev/de/insights/agent-harness-production-ai/Das KI-Rennen 2026 dreht sich im Kern um Harness-Engineering. Dieser Beitrag analysiert die 12 Kernmodule eines produktionsreifen Agent Harness, führende Framework-Philosophien und die 7 Architekturentscheidungen jedes KI-Architekten.Fri, 24 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/gpt-5-5-nvidia-jensen-huang-compute/https://nilsliu.dev/en/insights/gpt-5-5-nvidia-jensen-huang-compute/GPT-5.5 launched April 23, topping 14 benchmarks and cutting token usage 40%. Behind the scenes, Jensen Huang and NVIDIA are betting up to $100B on the compute infrastructure that makes it run.Fri, 24 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/gpt-5-5-nvidia-jensen-huang-compute/https://nilsliu.dev/zh/insights/gpt-5-5-nvidia-jensen-huang-compute/GPT-5.5 昨天（4 月 23 日）正式發布，在 14 項基準測試拿下第一，效率提升 40%。而黃仁勳與 NVIDIA 在這背後的算力押注，規模同樣驚人——10 GW、最高 1000 億美元。Fri, 24 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/gpt-5-5-nvidia-jensen-huang-compute/https://nilsliu.dev/de/insights/gpt-5-5-nvidia-jensen-huang-compute/GPT-5.5 erschien am 23. April und führt 14 Benchmarks an. Dahinter steckt Jensen Huangs NVIDIA-Wette: 10 Gigawatt Infrastruktur, bis zu 100 Milliarden Dollar Investition.Fri, 24 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/compression-persona-ppv/https://nilsliu.dev/en/insights/compression-persona-ppv/Ilya says compression is learning. Freedman finds only polynomial-growth monoids are compressible. If Persona can be projected onto a nilpotent substructure, PPV is not just a statistical fit — it's algebraically grounded personality compression.Sun, 19 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/compression-persona-ppv/https://nilsliu.dev/zh/insights/compression-persona-ppv/Ilya 說壓縮即學習，Freedman 發現只有多項式增長的 monoids 才能被壓縮。如果 Persona 可以投影到 nilpotent 子結構，那 PPV 就不只是統計擬合——而是有代數結構支撐的人格壓縮。Sun, 19 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/compression-persona-ppv/https://nilsliu.dev/de/insights/compression-persona-ppv/Ilya sagt, Kompression ist Lernen. Freedman findet, dass nur polynomial wachsende Monoide komprimierbar sind. Wenn Persona auf eine nilpotente Teilstruktur projiziert werden kann, ist PPV nicht nur statistisches Fitting — sondern algebraisch fundierte Persönlichkeitskompression.Sun, 19 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/claude-design-launch/https://nilsliu.dev/en/insights/claude-design-launch/On April 17, 2026, Anthropic launched Claude Design, a conversational AI visual design tool. Users simply describe what they need, and Claude generates interactive prototypes, slide decks, one-pagers, and more. Powered by Claude Opus 4.7, Anthropic's most capable vision model, the launch sent Figma's stock down 5% on the day.Sat, 18 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/claude-design-launch/https://nilsliu.dev/zh/insights/claude-design-launch/2026 年 4 月 17 日，Anthropic 正式推出 Claude Design，一款以對話驅動的 AI 視覺設計工具。使用者只需用文字描述需求，即可生成互動原型、簡報、單頁行銷頁等視覺成果。背後驅動力是 Claude Opus 4.7，Anthropic 迄今最強的視覺模型。消息一出，Figma 股價單日重挫 5%。Sat, 18 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/claude-design-launch/https://nilsliu.dev/de/insights/claude-design-launch/Am 17. April 2026 lancierte Anthropic Claude Design, ein konversationsbasiertes KI-Designtool. Nutzer beschreiben einfach ihre Anforderungen, und Claude erstellt interaktive Prototypen, Präsentationen, One-Pager und mehr. Angetrieben von Claude Opus 4.7, Anthropics leistungsfähigstem Vision-Modell, ließ der Launch Figmas Aktie um 5 % einbrechen.Sat, 18 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/hermes-agent-self-improving-framework/https://nilsliu.dev/en/insights/hermes-agent-self-improving-framework/Most AI Agents forget everything after each session. Hermes Agent is different — it remembers what you teach it and gets better over time. Here's what makes this open-source framework from NousResearch stand out.Thu, 16 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/hermes-agent-self-improving-framework/https://nilsliu.dev/zh/insights/hermes-agent-self-improving-framework/大多數 AI Agent 每次對話結束就「失憶」。Hermes Agent 不一樣——它會把你教給它的東西記下來，下一次用得更好。這是 NousResearch 在 2026 年推出的開源 Agent 框架，三層記憶體架構讓它越跑越聰明。Thu, 16 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/hermes-agent-self-improving-framework/https://nilsliu.dev/de/insights/hermes-agent-self-improving-framework/Die meisten KI-Agenten vergessen alles nach jeder Sitzung. Hermes Agent ist anders — er merkt sich, was Sie ihm beibringen, und wird mit der Zeit besser. Was dieses Open-Source-Framework von NousResearch auszeichnet.Thu, 16 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/harness-engineering-ai-agent/https://nilsliu.dev/en/insights/harness-engineering-ai-agent/Harness Engineering is the execution layer in AI Agent architecture. This post introduces the core design of a Harness: execution control, observability, hooks, tool sandboxing, and state management.Wed, 15 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/harness-engineering-ai-agent/https://nilsliu.dev/zh/insights/harness-engineering-ai-agent/Harness Engineering 是 AI Agent 架構中的執行底層。這篇文章介紹 Harness 的核心設計，涵蓋執行控制、可觀測性、Hook 機制、工具沙盒與狀態管理。Wed, 15 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/harness-engineering-ai-agent/https://nilsliu.dev/de/insights/harness-engineering-ai-agent/Harness Engineering ist die Ausführungsschicht in der KI-Agenten-Architektur. Dieser Beitrag stellt das Kerndesign eines Harness vor: Ausführungskontrolle, Observierbarkeit, Hooks, Tool-Sandbox und Zustandsverwaltung.Wed, 15 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/chatgpt-weird-bias-harvard/https://nilsliu.dev/en/insights/chatgpt-weird-bias-harvard/When AI researchers say LLMs are 'human-like,' which humans do they mean? A 2023 Harvard study used 262 cross-cultural survey variables and 94,278 respondents to show ChatGPT's cultural psychology aligns most closely with WEIRD Western democracies (r = -.70).Sun, 12 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/chatgpt-weird-bias-harvard/https://nilsliu.dev/zh/insights/chatgpt-weird-bias-harvard/當 AI 研究者說 LLM「像人類」，他們指的是哪些人類？哈佛 2023 年研究用 262 個跨文化問卷變數、94,278 名全球受訪者，得出關鍵結論：ChatGPT 的文化心理立場最接近 WEIRD 西方民主國家，與非 WEIRD 文化的相關係數高達 r = -.70。Sun, 12 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/chatgpt-weird-bias-harvard/https://nilsliu.dev/de/insights/chatgpt-weird-bias-harvard/Wenn KI-Forscher sagen, LLMs seien 'menschenähnlich', welche Menschen meinen sie? Eine Harvard-Studie von 2023 zeigt mit 262 kulturübergreifenden Variablen: ChatGPT steht kulturell westlichen WEIRD-Demokratien am nächsten (r = -.70).Sun, 12 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/digital-twin-persona-ppv/https://nilsliu.dev/en/insights/digital-twin-persona-ppv/Can LLMs truly simulate 'you'? From Generative Agents to BehaviorChain and the RAG-Free PPV framework, this in-depth analysis reveals why individual-level fidelity remains AI's hardest unsolved problem.Thu, 09 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/digital-twin-persona-ppv/https://nilsliu.dev/zh/insights/digital-twin-persona-ppv/LLM 真的能模擬「你」嗎？從 Generative Agents 到 BehaviorChain，再到以心理計量向量（PPV）驅動的 RAG-Free 數位孿生，本文深度解析三大方向的技術路線、核心差異與未解難題。Thu, 09 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/digital-twin-persona-ppv/https://nilsliu.dev/de/insights/digital-twin-persona-ppv/Können LLMs wirklich 'Sie' simulieren? Von Generative Agents über BehaviorChain bis zum RAG-freien PPV-Framework zeigt diese Analyse, warum individuelle Persönlichkeitstreue KIs schwierigste ungelöste Herausforderung bleibt.Thu, 09 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/karpathy-llm-wiki-knowledge-base/https://nilsliu.dev/en/insights/karpathy-llm-wiki-knowledge-base/Former Tesla AI Director Andrej Karpathy proposes replacing traditional RAG with an LLM-maintained personal Wiki. How does this three-layer architecture compound knowledge like interest? A complete breakdown.Tue, 07 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/karpathy-llm-wiki-knowledge-base/https://nilsliu.dev/zh/insights/karpathy-llm-wiki-knowledge-base/前 Tesla AI 主管 Andrej Karpathy 提出用 LLM 維護一座個人 Wiki，取代傳統 RAG 的「每次重新搜尋」模式。這套三層架構如何讓知識像複利一樣累積？本文完整解析。Tue, 07 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/karpathy-llm-wiki-knowledge-base/https://nilsliu.dev/de/insights/karpathy-llm-wiki-knowledge-base/Der ehemalige Tesla-KI-Direktor Andrej Karpathy schlägt vor, traditionelles RAG durch ein LLM-gepflegtes persönliches Wiki zu ersetzen. Wie sammelt diese Drei-Schichten-Architektur Wissen wie Zinseszins an? Eine vollständige Analyse.Tue, 07 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/google-gemma-4-developer-guide/https://nilsliu.dev/en/insights/google-gemma-4-developer-guide/Released April 2026 under Apache 2.0, Gemma 4 comes in four sizes — E2B, E4B, 26B MoE, and 31B Dense. The 31B ranks #3 among all open models globally with 256K context and native agentic workflows. A complete breakdown for AI developers.Fri, 03 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/google-gemma-4-developer-guide/https://nilsliu.dev/zh/insights/google-gemma-4-developer-guide/2026年4月，Google DeepMind 正式發布 Gemma 4，首度採用 Apache 2.0 授權，提供 E2B、E4B、26B MoE 和 31B Dense 四種規格。31B 模型在全球開放模型排行榜名列第三，支援 256K 上下文視窗與原生 Agentic 工作流程。本文完整解析 Gemma 4 的架構亮點、各版本特色，以及開發者如何快速上手部署。Fri, 03 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/google-gemma-4-developer-guide/https://nilsliu.dev/de/insights/google-gemma-4-developer-guide/Im April 2026 unter Apache 2.0 veröffentlicht, bietet Gemma 4 vier Varianten — E2B, E4B, 26B MoE und 31B Dense. Das 31B-Modell belegt Platz 3 unter allen offenen Modellen weltweit, unterstützt 256K Kontext und native agentische Workflows.Fri, 03 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/claude-code-harness-leak-architecture/https://nilsliu.dev/en/insights/claude-code-harness-leak-architecture/2026年3月底，Anthropic 在 npm 發布更新時意外夾帶 59.8MB 的 Source Map，導致 Claude Code 的底層程式碼全面洩漏。這不僅是一次工程失誤，更是企業級 Agent 架構、多層提示詞與臥底模式等設計細節的首次大解密。Wed, 01 Apr 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/claude-code-harness-leak-architecture/https://nilsliu.dev/zh/insights/claude-code-harness-leak-architecture/2026年3月底，Anthropic 在 npm 發布更新時意外夾帶 59.8MB 的 Source Map，導致 Claude Code 的底層程式碼全面洩漏。這不僅是一次工程失誤，更是企業級 Agent 架構、多層提示詞與臥底模式等設計細節的首次大解密。Wed, 01 Apr 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/claude-code-harness-leak-architecture/https://nilsliu.dev/de/insights/claude-code-harness-leak-architecture/2026年3月底，Anthropic 在 npm 發布更新時意外夾帶 59.8MB 的 Source Map，導致 Claude Code 的底層程式碼全面洩漏。這不僅是一次工程失誤，更是企業級 Agent 架構、多層提示詞與臥底模式等設計細節的首次大解密。Wed, 01 Apr 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-shopping-agent-agentic-commerce/https://nilsliu.dev/en/insights/ai-shopping-agent-agentic-commerce/AI 購物代理正從展示品變成真正的消費工具。Walmart 推出自研 AI 助手 Sparky，Target 與 Google Gemini 合作，Shopify 發布 Agentic Commerce 協議。當 AI Agent 開始替你刷卡，電商的遊戲規則正在被徹底改寫。Tue, 31 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-shopping-agent-agentic-commerce/https://nilsliu.dev/zh/insights/ai-shopping-agent-agentic-commerce/AI 購物代理正從展示品變成真正的消費工具。Walmart 推出自研 AI 助手 Sparky，Target 與 Google Gemini 合作，Shopify 發布 Agentic Commerce 協議。當 AI Agent 開始替你刷卡，電商的遊戲規則正在被徹底改寫。Tue, 31 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-shopping-agent-agentic-commerce/https://nilsliu.dev/de/insights/ai-shopping-agent-agentic-commerce/AI 購物代理正從展示品變成真正的消費工具。Walmart 推出自研 AI 助手 Sparky，Target 與 Google Gemini 合作，Shopify 發布 Agentic Commerce 協議。當 AI Agent 開始替你刷卡，電商的遊戲規則正在被徹底改寫。Tue, 31 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/openclaw-founder-idea/https://nilsliu.dev/en/insights/openclaw-founder-idea/OpenClaw 創辦人 Peter Steinberger 從一個週末小專案出發，用 Anthropic 的 Claude 打造了 AI Agent 框架，卻因商標爭議被迫改名。這段從車庫到被 OpenAI 延攬的旅程，證明了在 AI 大航海時代，任何一個小小的想法都可能改變遊戲規則。Mon, 30 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/openclaw-founder-idea/https://nilsliu.dev/zh/insights/openclaw-founder-idea/OpenClaw 創辦人 Peter Steinberger 從一個週末小專案出發，用 Anthropic 的 Claude 打造了 AI Agent 框架，卻因商標爭議被迫改名。這段從車庫到被 OpenAI 延攬的旅程，證明了在 AI 大航海時代，任何一個小小的想法都可能改變遊戲規則。Mon, 30 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/openclaw-founder-idea/https://nilsliu.dev/de/insights/openclaw-founder-idea/OpenClaw 創辦人 Peter Steinberger 從一個週末小專案出發，用 Anthropic 的 Claude 打造了 AI Agent 框架，卻因商標爭議被迫改名。這段從車庫到被 OpenAI 延攬的旅程，證明了在 AI 大航海時代，任何一個小小的想法都可能改變遊戲規則。Mon, 30 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/karpathy-ai-psychosis-claws/https://nilsliu.dev/en/insights/karpathy-ai-psychosis-claws/Andrej Karpathy 在最新的访谈中提到自己彷彿得了「AI 精神病」——他已經幾個月沒親自寫過程式碼了。這篇文章整理了他在《No Priors》Podcast 分享的核心觀點，包括 Claws 的概念以及未來軟體開發範式的轉變。Tue, 24 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/karpathy-ai-psychosis-claws/https://nilsliu.dev/zh/insights/karpathy-ai-psychosis-claws/前 OpenAI 與 Tesla AI 總監 Andrej Karpathy 在最新《No Priors》訪談中分享「AI 精神病」經歷：數月未寫程式碼。本文解析其核心觀點，探討 Claws 背景 Agent 概念、Token 焦慮，以及 Software 3.0 時代軟體開發範式的轉變。Tue, 24 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/karpathy-ai-psychosis-claws/https://nilsliu.dev/de/insights/karpathy-ai-psychosis-claws/Andrej Karpathy 在最新的访谈中提到自己彷彿得了「AI 精神病」——他已經幾個月沒親自寫過程式碼了。這篇文章整理了他在《No Priors》Podcast 分享的核心觀點，包括 Claws 的概念以及未來軟體開發範式的轉變。Tue, 24 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/google-stitch/https://nilsliu.dev/en/insights/google-stitch/In March 2026, Google introduced a major update to Stitch. Powered by Gemini, this AI-native design canvas generates UI through natural language and features 'Voice Canvas'. How will this completely disrupt Figma and the future workflow of designers?Thu, 19 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/google-stitch/https://nilsliu.dev/zh/insights/google-stitch/2026 年 3 月 Google 推出了 Stitch 更新。這款由 Gemini 驅動的 AI 原生設計畫布不僅能透過自然語言生成 UI，更新增 Voice Canvas 語音編輯。它將如何徹底顛覆 Figma 與設計師的未來工作流？Thu, 19 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/google-stitch/https://nilsliu.dev/de/insights/google-stitch/Im März 2026 hat Google ein großes Update für Stitch eingeführt. Angetrieben von Gemini generiert diese KI-native Design-Canvas UI durch natürliche Sprache. Wie wird dies Figma und den zukünftigen Workflow von Designern völlig stören?Thu, 19 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/openclaw-ecosystem/https://nilsliu.dev/en/insights/openclaw-ecosystem/OpenClaw showed us that an assistant is an always-on computing layer, not just a chatbot. But its variants (like NanoBot, CoPaw, IronClaw) are even more fascinating. Spanning five distinct paths, they outline the true shape of next-generation AI assistants.Wed, 18 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/openclaw-ecosystem/https://nilsliu.dev/zh/insights/openclaw-ecosystem/OpenClaw 的出現讓大家看到「assistant 不只是聊天機器人」，而是一個常駐的代理層。但真正有趣的是它的變種（如 NanoBot, CoPaw, IronClaw），從基礎設施到極端硬體部署，五大路線正在回答下一代 AI 助理的真正形狀。Wed, 18 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/openclaw-ecosystem/https://nilsliu.dev/de/insights/openclaw-ecosystem/OpenClaw hat gezeigt, dass ein Assistent eine persistente Computing-Schicht ist. Aber seine Varianten (wie NanoBot, CoPaw, IronClaw) sind noch faszinierender. Fünf verschiedene Pfade skizzieren die wahre Form der nächsten Generation von KI-Assistenten.Wed, 18 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-agent-product-design-traps/https://nilsliu.dev/en/insights/ai-agent-product-design-traps/AI Agents sound cool, but building Agent products in enterprise is full of pitfalls. Here are five design traps I've experienced firsthand.Sun, 08 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-agent-product-design-traps/https://nilsliu.dev/zh/insights/ai-agent-product-design-traps/Agent 聽起來很酷，但在企業環境做 Agent 產品很容易踩坑。這是我親身經歷的五個設計陷阱。 2025 年是 AI Agent 元年。到了 2026 年，每個企業都在問：「我們也要做 Agent 嗎？」 但「做 Agent」和「做好 Agent 產品」是兩件完全不同的事。 這篇分享我在設計...Sun, 08 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-agent-product-design-traps/https://nilsliu.dev/de/insights/ai-agent-product-design-traps/KI-Agenten klingen cool, aber der Bau von Agent-Produkten in Unternehmen ist voller Stolperfallen. Hier sind fünf Designfallen, die ich selbst erlebt habe.Sun, 08 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/genai-roi-metrics/https://nilsliu.dev/en/insights/genai-roi-metrics/When your boss asks 'Is AI worth the investment?', you need numbers. Here's the four-metric framework I use to prove GenAI value.Sat, 07 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/genai-roi-metrics/https://nilsliu.dev/zh/insights/genai-roi-metrics/主管問你「AI 到底值不值得投資？」時，你需要拿出數字。這是我用來證明 GenAI 價值的四個指標框架。 每一個做 GenAI 的 PM 都會遇到這個問題： > 「AI 很好，但怎麼證明它值得投資？」 在銀行，這個問題更尖銳。因為每一筆 AI 的投資都需要和傳統 IT 項目競爭預算——而傳統項目...Sat, 07 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/genai-roi-metrics/https://nilsliu.dev/de/insights/genai-roi-metrics/Wenn Ihr Chef fragt: „Lohnt sich die Investition in KI?“, brauchen Sie Zahlen. Hier ist das Framework mit vier Kennzahlen, mit dem ich den GenAI-Wert beweise.Sat, 07 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/enterprise-prompt-engineering/https://nilsliu.dev/en/insights/enterprise-prompt-engineering/Enterprise prompt engineering is nothing like personal ChatGPT use. Structured templates, version control, multi-role design — lessons from the trenches.Fri, 06 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/enterprise-prompt-engineering/https://nilsliu.dev/zh/insights/enterprise-prompt-engineering/企業級 prompt 跟個人使用完全不同。結構化、版本控制、多角色設計——這些是我踩過的坑。 個人用 ChatGPT 寫 prompt，最多就是「幫我寫一封信」。 企業用 LLM 寫 prompt，是一整套工程實踐。 這篇分享我在企業場景中做 prompt engineering 時，累積的幾個...Fri, 06 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/enterprise-prompt-engineering/https://nilsliu.dev/de/insights/enterprise-prompt-engineering/Unternehmens-Prompt-Engineering hat nichts mit der privaten ChatGPT-Nutzung zu tun. Strukturierte Vorlagen, Versionskontrolle, Multi-Rollen-Design.Fri, 06 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/rag-system-design-banking/https://nilsliu.dev/en/insights/rag-system-design-banking/Building a RAG system in banking: how to choose your chunk strategy, embedding model, and retrieval pipeline. Lessons from real production experience.Thu, 05 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/rag-system-design-banking/https://nilsliu.dev/zh/insights/rag-system-design-banking/在銀行建 RAG 系統，chunk 策略、embedding 模型、retrieval pipeline 三個決策點怎麼選？實戰經驗分享。 RAG（Retrieval-Augmented Generation）是 2024-2025 最熱門的企業 AI 架構。但真正在企業環境中建過 RAG 系統...Thu, 05 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/rag-system-design-banking/https://nilsliu.dev/de/insights/rag-system-design-banking/Der Aufbau eines RAG-Systems im Bankwesen: So wählen Sie Ihre Chunking-Strategie, Ihr Einbettungsmodell und Ihre Abrufpipeline.Thu, 05 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-pm-vs-traditional-pm/https://nilsliu.dev/en/insights/ai-pm-vs-traditional-pm/Does an AI PM need to code? Need to understand math? A complete skill tree breakdown comparing AI PMs and traditional PMs.Wed, 04 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-pm-vs-traditional-pm/https://nilsliu.dev/zh/insights/ai-pm-vs-traditional-pm/想從傳統 PM 進行產品經理轉型？AI 產品經理需要寫程式嗎？結合 OpenAI 副總裁 Peter Deng 與 Vibe Coding 哲學，拆解你缺少的 AI PM 技能樹與思維。 我被問過最多次的問題就是：「AI 產品經理跟傳統 PM 到底差在哪裡？如果我要進行產品經理轉型，需要會寫程式...Wed, 04 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-pm-vs-traditional-pm/https://nilsliu.dev/de/insights/ai-pm-vs-traditional-pm/Muss ein AI PM programmieren können? Mathematik verstehen? Ein vollständiger Vergleich der Fähigkeiten von AI PMs und traditionellen PMs.Wed, 04 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-pm-banking-10x-harder/https://nilsliu.dev/en/insights/ai-pm-banking-10x-harder/Deploying AI in a bank isn't just picking a model. Compliance, security, data governance, organizational culture — each hurdle is harder than the last.Tue, 03 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-pm-banking-10x-harder/https://nilsliu.dev/zh/insights/ai-pm-banking-10x-harder/金融業 AI 轉型不是選個模型就好。資安、法遵、資料治理——這是將銀行 AI 系統從 1 推向 100 的必經之路，也是 AI 落地挑戰最真實的面貌。 在新創公司開發 Generative AI 產品，你可能靠著 Vibe Coding（與編輯器對話直接產出程式碼）三天就能上線一個 Chatbo...Tue, 03 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-pm-banking-10x-harder/https://nilsliu.dev/de/insights/ai-pm-banking-10x-harder/KI in einer Bank einzusetzen bedeutet nicht nur, ein Modell auszuwählen. Compliance, Sicherheit, Data Governance — jede Hürde ist schwieriger als die vorherige.Tue, 03 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-pm-feature-to-patent/https://nilsliu.dev/en/insights/ai-pm-feature-to-patent/What does a GenAI feature go through from inception to launch to patent filing? A complete journey record from a bank AI PM.Mon, 02 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-pm-feature-to-patent/https://nilsliu.dev/zh/insights/ai-pm-feature-to-patent/擔心 AI 功能幾個月後就被開源模型取代？本文教你如何透過預判模型進化的思維進行 AI 專利申請，打造別人抄不走的 GenAI 產品護城河。 大多數 PM 一聽到「AI 專利申請」，直覺反應都是：那屬於「底層演算法科學家」或 RD 研發部門的事，與產品經理無關。 但我主導的最高價值專利，往往是從...Mon, 02 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-pm-feature-to-patent/https://nilsliu.dev/de/insights/ai-pm-feature-to-patent/Was macht ein GenAI-Feature von der Konzeption über den Launch bis hin zur Patentanmeldung durch? Ein kompletter Erfahrungsbericht eines Bank-KI-PMs.Mon, 02 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-pm-daily-work/https://nilsliu.dev/en/insights/ai-pm-daily-work/No coding, no model tuning — so what does an AI Product Manager actually do all day? A real daily breakdown from a GenAI PM working in banking.Sun, 01 Mar 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-pm-daily-work/https://nilsliu.dev/zh/insights/ai-pm-daily-work/不只是開票寫規格，真實的 AI PM 工作內容究竟長怎樣？從單純的 PM 轉向全端 Builder，帶你走一遍銀行生成式 AI 產品開發流程的殘酷日常。 很多人聽到「AI 產品經理」，第一反應往往是：「所以你整天都在用 Python 寫模型嗎？」 其實不是。我花最多時間在『建造』。 誠如 Ope...Sun, 01 Mar 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-pm-daily-work/https://nilsliu.dev/de/insights/ai-pm-daily-work/Kein Programmieren, kein Modell-Tuning — was macht ein KI-Produktmanager eigentlich den ganzen Tag? Ein echter Einblick in den Alltag eines GenAI PM.Sun, 01 Mar 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/fde-diamond-org/https://nilsliu.dev/en/insights/fde-diamond-org/Same era, same job title — one group is being laid off while another is being hired. What separates them isn't seniority or credentials; it's how fast they'...Fri, 27 Feb 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/fde-diamond-org/https://nilsliu.dev/zh/insights/fde-diamond-org/同一個時代，同樣叫做工程師，有人因 AI 失去工作，也有人因 AI 找到工作。決定你在哪一邊的，不是學歷年資，而是調整速度。 同一個時代，同樣叫做工程師，有人因 AI 失去工作，也有人因 AI 找到工作。決定你在哪一邊的，不是學歷年資，而是調整速度。新職缺已經靜靜上線了 ——— Block 剛宣...Fri, 27 Feb 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/fde-diamond-org/https://nilsliu.dev/de/insights/fde-diamond-org/In derselben Ära und unter derselben Berufsbezeichnung 'Ingenieur' verlieren einige durch KI ihren Job, während andere durch sie neue finden. Was deinen Pla...Fri, 27 Feb 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-picking-future-talents/https://nilsliu.dev/en/insights/ai-picking-future-talents/Over the past year leading a team, I found that people with genuine curiosity thrive in AI-augmented work. Boris Cherny of Anthropic thinks this gap is exac...Sat, 21 Feb 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-picking-future-talents/https://nilsliu.dev/zh/insights/ai-picking-future-talents/過去一年我也是這樣帶團隊，但發現部份經驗與好奇心不足的人會對這種工作型態無所適從，這大概就是一種挑選未來人才的方法。 過去一年我也是這樣帶團隊，但發現部份經驗與好奇心不足的人（例如一些實習生）會對這種工作型態無所適從 個人感覺，這大概就是一種挑選未來人才的方法 --- Anthropic 的 C...Sat, 21 Feb 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-picking-future-talents/https://nilsliu.dev/de/insights/ai-picking-future-talents/So habe ich im vergangenen Jahr Teams geführt, aber ich stellte fest, dass einige Mitarbeiter mit mangelnder Erfahrung und Neugier sich in diesem Arbeitsmod...Sat, 21 Feb 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-agent-intern-recruit/https://nilsliu.dev/en/insights/ai-agent-intern-recruit/We're building an AI Agent platform that actually ships to real users — want to go from PoC to production? We're looking for full-stack, backend, and GenAI...Wed, 21 Jan 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-agent-intern-recruit/https://nilsliu.dev/zh/insights/ai-agent-intern-recruit/我們在做一個「真的會上線、真的有人用」的 AI Agent 平台：把 GenAI 串進工作流程，從資料處理到 tools & workflows 的調度。 【招募實習生｜海外 AI Agent 平台｜全端/後端/GenAI】 我們在做一個「真的會上線、真的有人用」的 AI Agent 平台：把 ...Wed, 21 Jan 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-agent-intern-recruit/https://nilsliu.dev/de/insights/ai-agent-intern-recruit/Wir bauen eine KI-Agenten-Plattform, die 'wirklich live geht und von echten Menschen genutzt wird', indem wir GenAI in Arbeitsabläufe integrieren – von der...Wed, 21 Jan 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-browser-auto-report/https://nilsliu.dev/en/insights/ai-browser-auto-report/I built an AI Browser that records every reasoning step, tracks queries, auto-decides when to screenshot, and compiles everything into a structured investig...Tue, 13 Jan 2026 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-browser-auto-report/https://nilsliu.dev/zh/insights/ai-browser-auto-report/做了一個 AI Browser，能自動記錄所有思考過程、查詢軌跡、自動判斷是否需要截圖，並在最後輸出一份調查報告。 做了一個AI Browser，能自動記錄所有思考過程、查詢軌跡、自動判斷是否需要截圖，並在最後輸出一份調查報告 https://youtu.be/Hk42hV7q2Q?si=Rp9...Tue, 13 Jan 2026 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-browser-auto-report/https://nilsliu.dev/de/insights/ai-browser-auto-report/Habe einen KI-Browser entwickelt, der automatisch alle Denkvorgänge und Suchverläufe aufzeichnet, entscheidet, wann Screenshots gemacht werden müssen, und a...Tue, 13 Jan 2026 00:00:00 GMThttps://nilsliu.dev/en/insights/2025-year-in-review/https://nilsliu.dev/en/insights/2025-year-in-review/My 2025 AI journey in four numbers: 6, 5, 1, 6. Not because it was glamorous, but because it was grounded. Building GenAI in a bank is like replacing the pl...Wed, 31 Dec 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/2025-year-in-review/https://nilsliu.dev/zh/insights/2025-year-in-review/我的 2025 年 AI 歷險有四個數字：6、5、1、6。不是因為多光鮮，而是因為很踏實。在銀行做 GenAI 就像在營運中的大樓換管線，真正難的往往不是模型，而是工程與信任。 我的 2025 年 AI 歷險有四個數字： 6項 新型專利 已核准 ...Wed, 31 Dec 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/2025-year-in-review/https://nilsliu.dev/de/insights/2025-year-in-review/Mein KI-Abenteuer 2025 in vier Zahlen: 6, 5, 1, 6. Nicht weil es so glamourös war, sondern weil es bodenständig blieb. GenAI in einer Bank zu implementieren...Wed, 31 Dec 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-history-hinton-nobel/https://nilsliu.dev/en/insights/ai-history-hinton-nobel/There was a time when saying 'AI' in serious academic circles was a mark against you. Geoffrey Hinton won the Turing Award in 2018 and the Nobel Prize in Ph...Wed, 17 Dec 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-history-hinton-nobel/https://nilsliu.dev/zh/insights/ai-history-hinton-nobel/過去「AI」在學術圈帶有負面標籤，Geoffrey Hinton 也曾因做神經網路被強烈質疑。如今他獲得圖靈獎與諾貝爾獎，AI 正在改變世界。 過去曾有段時間，「AI」這個詞在學術圈（特別是幾次 AI 寒冬之後）帶有負面標籤，資訊工程領域的一些專家學者如果聽到有人講 AI，可能會覺得這個人太會吹...Wed, 17 Dec 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-history-hinton-nobel/https://nilsliu.dev/de/insights/ai-history-hinton-nobel/Einst hatte 'KI' in der Akademie ein negatives Image, und Geoffrey Hinton wurde stark für seine Arbeit an neuronalen Netzen kritisiert. Heute hat er den Tur...Wed, 17 Dec 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/google-gemini-antigravity-devops/https://nilsliu.dev/en/insights/google-gemini-antigravity-devops/After Google pushed Gemini 3 Pro and Antigravity, I started rethinking the relationship between developers and AI infrastructure — and what 'role elevation'...Sat, 06 Dec 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/google-gemini-antigravity-devops/https://nilsliu.dev/zh/insights/google-gemini-antigravity-devops/AI 不再只是 demo 的玩具，而是變成雲端基礎設施的一環。工具在變，角色就有機會跟著變。 Google 把 Gemini 3 Pro 和 Antigravity 推上來之後 我開始重新思考「開發者跟 AI 的關係」這件事 這一波更新，很直白地釋出一個訊號： AI 不再只是 demo 或 si...Sat, 06 Dec 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/google-gemini-antigravity-devops/https://nilsliu.dev/de/insights/google-gemini-antigravity-devops/KI ist nicht mehr nur ein Spielzeug für Demos, sondern wird Teil der Cloud-Infrastruktur. Wenn sich Werkzeuge ändern, ändern sich auch die dazugehörigen Rollen.Sat, 06 Dec 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/perplexity-challenges/https://nilsliu.dev/en/insights/perplexity-challenges/Perplexity is caught between being the next-generation search paradigm and facing mounting legal pressure from content publishers. Can they find a deal stru...Sat, 06 Dec 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/perplexity-challenges/https://nilsliu.dev/zh/insights/perplexity-challenges/Perplexity面對的是一個「位置」的考驗，一邊要扛起「AI 搜尋新一代」的期待，一邊又被各種內容方用法律與商業談判來拉扯。 ［關於Perplexity的困境］ 在Perplexity剛出世的時候我就覺得我們團隊也做得出一樣的產品，甚至能一樣好 事實上，只要是一般夠有 sense 的開發團隊...Sat, 06 Dec 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/perplexity-challenges/https://nilsliu.dev/de/insights/perplexity-challenges/Perplexity steht vor einem 'Positionierungs'-Test: Einerseits sollen sie die Erwartungen an die 'KI-Suche der nächsten Generation' erfüllen, andererseits we...Sat, 06 Dec 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/devfest-taipei-2025-recap/https://nilsliu.dev/en/insights/devfest-taipei-2025-recap/At DevFest Taipei 2025 I shared a real production AI coaching platform — multi-agent collaboration, Persona World, Ontology + GraphRAG, delivering 24/7 pers...Tue, 02 Dec 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/devfest-taipei-2025-recap/https://nilsliu.dev/zh/insights/devfest-taipei-2025-recap/在 DevFest Taipei 2025 分享真實上線的 AI 業務教練平台——多 Agent 協作、Persona World、Ontology + GraphRAG，打造 24/7 個人化培訓系統。Tue, 02 Dec 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/devfest-taipei-2025-recap/https://nilsliu.dev/de/insights/devfest-taipei-2025-recap/Beim DevFest Taipei 2025 habe ich eine produktive KI-Vertriebs-Coaching-Plattform vorgestellt—mit Multi-Agenten-Kollaboration, Persona World und Ontology +...Tue, 02 Dec 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/devfest-taipei-2025-announcement/https://nilsliu.dev/en/insights/devfest-taipei-2025-announcement/On November 30th I'll be presenting real AI Agent team applications running in production at DevFest Taipei 2025, hosted by Google GDG.Sun, 16 Nov 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/devfest-taipei-2025-announcement/https://nilsliu.dev/zh/insights/devfest-taipei-2025-announcement/11/30 將在 Google GDG 舉辦的 DevFest Taipei 2025 分享 AI Agent team 落地應用，歡迎報名參加。 11/30 DevFest Taipei 將分享真實在生產環境使用的業務訓練Agent team應用 11/30 10:30 台大博雅教學館博201見Sun, 16 Nov 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/devfest-taipei-2025-announcement/https://nilsliu.dev/de/insights/devfest-taipei-2025-announcement/Am 30.11. werde ich auf der von Google GDG organisierten DevFest Taipei 2025 über die praktische Anwendung von KI-Agententeams sprechen. Wir freuen uns über...Sun, 16 Nov 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/graphrag-smart-customer-service/https://nilsliu.dev/en/insights/graphrag-smart-customer-service/GraphRAG replaces flat vector retrieval with graph-structured knowledge, enabling multi-hop reasoning and consistent context — 86% accuracy on RobustQA vs....Mon, 10 Nov 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/graphrag-smart-customer-service/https://nilsliu.dev/zh/insights/graphrag-smart-customer-service/GraphRAG 以圖譜化知識結構取代單點檢索，藉由節點與邊的關聯輔助生成，既能維持上下文一致，又具備多跳推理能力。 GraphRAG 以知識圖譜串聯記憶，重塑智慧客服推理能力 傳統 RAG 模型仰賴向量檢索，當對話延展過長或主題跨度太廣，語意易被割裂，導致 LLM 回應冗長卻偏離重點。Grap...Mon, 10 Nov 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/graphrag-smart-customer-service/https://nilsliu.dev/de/insights/graphrag-smart-customer-service/GraphRAG ersetzt Einzelpunktsuchen durch eine graphbasierte Wissensstruktur. Es nutzt die Beziehungen zwischen Knoten und Kanten, um die Generierung zu unte...Mon, 10 Nov 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/six-patents-approved-2025/https://nilsliu.dev/en/insights/six-patents-approved-2025/All six utility model patents filed at the start of this year have been approved — two dual-filed. Another five submitted last month. This is what real GenA...Thu, 06 Nov 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/six-patents-approved-2025/https://nilsliu.dev/zh/insights/six-patents-approved-2025/今年初 6 項新型專利已全部獲得核准，上個月再次提交 5 項專利。這是生成式 AI 落地爆發的一年。 至此，今年初6項新型專利已全部獲得核准，其中2項為一案兩請（新型+專利） 上個月再次提交5項專利，預計1項新型，4項一案兩請 這是生成式AI落地爆發的一年Thu, 06 Nov 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/six-patents-approved-2025/https://nilsliu.dev/de/insights/six-patents-approved-2025/Sechs neue Gebrauchsmuster wurden Anfang dieses Jahres genehmigt und letzten Monat wurden fünf weitere Patente angemeldet. Dies ist das Jahr, in dem generat...Thu, 06 Nov 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/patent-knowledge-graph/https://nilsliu.dev/en/insights/patent-knowledge-graph/How does a bank GenAI Product Manager design an LLM system that automatically builds a knowledge graph from business pain points, and successfully obtain a...Sat, 01 Nov 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/patent-knowledge-graph/https://nilsliu.dev/zh/insights/patent-knowledge-graph/一位銀行 GenAI 產品經理如何從業務痛點出發，設計出能自動建構知識圖譜的 LLM 系統，並成功取得新型專利？本文分享從需求到專利的完整思路。 大多數人對知識圖譜的印象是：需要大量人工標記、需要事先定義本體（ontology）、需要一群領域專家持續維護。 這在金融業尤其是個大問題。 銀行的業務...Sat, 01 Nov 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/patent-knowledge-graph/https://nilsliu.dev/de/insights/patent-knowledge-graph/Wie konzipiert ein Bank-GenAI-Produktmanager ein LLM-System, das als Antwort auf Geschäftsprobleme automatisch einen Knowledge Graph aufbaut, und erhält daf...Sat, 01 Nov 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/devfest-taipei-signup/https://nilsliu.dev/en/insights/devfest-taipei-signup/I'll be speaking at DevFest Taipei 2025 on November 30th — AI Agent team applications in production. Free entry, registration required.Tue, 28 Oct 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/devfest-taipei-signup/https://nilsliu.dev/zh/insights/devfest-taipei-signup/11/30 在 Google GDG 舉辦的 DevFest Taipei 2025 分享 AI Agent team 落地應用，免費入場，歡迎報名。 11/30 我會在Google GDG舉辦的DevFest Taipei 2025分享AI Agent team 落地應用，歡迎有興趣的朋友們來...Tue, 28 Oct 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/devfest-taipei-signup/https://nilsliu.dev/de/insights/devfest-taipei-signup/Am 30.11. spreche ich auf der DevFest Taipei 2025 der Google GDG über angewandte KI-Agententeams. Freier Eintritt nach Überprüfung, Anmeldung ist willkommen.Tue, 28 Oct 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/best-attention-explanation/https://nilsliu.dev/en/insights/best-attention-explanation/The clearest explanation of 'Attention Is All You Need' I've come across — the mechanics, not just the intuition.Sun, 26 Oct 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/best-attention-explanation/https://nilsliu.dev/zh/insights/best-attention-explanation/這是我看過的講解 Attention is all you need 的影片中講解最好的一部。 這是我看過的講解Attention is all you need的影片中講解最好的一部 https://www.youtube.com/watch?v=VaEjGnHgOI&t=1274sSun, 26 Oct 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/best-attention-explanation/https://nilsliu.dev/de/insights/best-attention-explanation/Dies ist das beste Video-Tutorial zum Thema 'Attention is All You Need', das ich bisher gesehen habe.Sun, 26 Oct 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/jason-wei-ai-scenarios/https://nilsliu.dev/en/insights/jason-wei-ai-scenarios/Jason Wei's talk gave me a genuine 'aha' moment. The systematic framework he lays out for finding AI use cases is exactly what I wish I'd articulated earlier.Thu, 23 Oct 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/jason-wei-ai-scenarios/https://nilsliu.dev/zh/insights/jason-wei-ai-scenarios/Jason Wei 這段演講醍醐灌頂，聽到起 goose bump，尋找 AI 場景確實應該用這個系統化的方法。 Jason Wei這段演講醍醐灌頂，聽到起goose bump，尋找AI場景確實應該用這個系統化的方法Thu, 23 Oct 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/jason-wei-ai-scenarios/https://nilsliu.dev/de/insights/jason-wei-ai-scenarios/Dieser Vortrag von Jason Wei war eine echte Erleuchtung, hat mir Gänsehaut bereitet. Es ist definitiv ein systematischer Weg, um KI-Szenarien zu finden.Thu, 23 Oct 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/ithome-hello-world-2025/https://nilsliu.dev/en/insights/ithome-hello-world-2025/This October at the iThome Hello World Developer Conference, I presented four intensive sessions covering MCP, GraphRAG, Vibe Coding, and Enterprise LLM Gua...Mon, 20 Oct 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/ithome-hello-world-2025/https://nilsliu.dev/zh/insights/ithome-hello-world-2025/今年我在 iThome Hello World 開發者大會帶來四場密集的 AI 實戰分享，從 MCP 生態系、GraphRAG 到企業級 LLM 防禦機制，探討大型語言模型落地企業的核心技術。 在今年的 iThome Hello World 2025 開發者大會上，我一口氣帶來了四場跟 AI 實...Mon, 20 Oct 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/ithome-hello-world-2025/https://nilsliu.dev/de/insights/ithome-hello-world-2025/Auf der iThome Hello World Entwicklerkonferenz dieses Jahres hielt ich vier KI-fokussierte Vorträge, vom MCP-Ökosystem und GraphRAG bis hin zu Abwehrmechani...Mon, 20 Oct 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/thinking-machine-lab-deterministic-ai/https://nilsliu.dev/en/insights/thinking-machine-lab-deterministic-ai/Does setting temperature to 0 give perfectly consistent AI outputs? No — and Thinking Machine Lab found out why. Batch processing is the culprit, and they b...Sat, 13 Sep 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/thinking-machine-lab-deterministic-ai/https://nilsliu.dev/zh/insights/thinking-machine-lab-deterministic-ai/把 AI 模型的 temperature 設為 0 就能得到完全一致的結果嗎？答案是否，Thinking Machine Lab 的研究揭示了根本性的原因。 Thinking Machine Lab (Mira Murati創立的startup，在還沒有任何產品和工作方向時，光是Mira Mur...Sat, 13 Sep 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/thinking-machine-lab-deterministic-ai/https://nilsliu.dev/de/insights/thinking-machine-lab-deterministic-ai/Erhält man absolut konsistente Ergebnisse, wenn man die Temperatur des KI-Modells auf 0 setzt? Die Antwort ist nein, und eine Studie von Thinking Machine La...Sat, 13 Sep 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/patent-adaptive-security/https://nilsliu.dev/en/insights/patent-adaptive-security/When a GenAI system queries sensitive data, how do you prevent malicious users from bypassing security? This article details how a bank AI Product Manager d...Thu, 11 Sep 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/patent-adaptive-security/https://nilsliu.dev/zh/insights/patent-adaptive-security/當 GenAI 系統被用來查詢敏感資料，如何防止惡意使用者繞過安全機制？本文介紹銀行 AI 產品經理如何設計動態存取控制專利，用 AI 守護 AI。 在銀行導入生成式 AI 的過程中，最讓我夜不能寐的問題不是模型效果，而是安全性。 一個能查詢客戶資料的 AI 聊天機器人，如果被有心人士透過特製 ...Thu, 11 Sep 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/patent-adaptive-security/https://nilsliu.dev/de/insights/patent-adaptive-security/Wenn ein GenAI-System sensible Daten abfragt, wie verhindert man, dass böswillige Nutzer die Sicherheit umgehen? Dieser Artikel beschreibt detailliert, wie...Thu, 11 Sep 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/vibe-coding-three-giants/https://nilsliu.dev/en/insights/vibe-coding-three-giants/The vibe coding landscape has consolidated around three players — OpenAI's Codex, Google's Gemini, and Anthropic's Claude. Each pulls in a different direction.Tue, 02 Sep 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/vibe-coding-three-giants/https://nilsliu.dev/zh/insights/vibe-coding-three-giants/像極了轉扭蛋的 vibe coding 御三家，OpenAI 的 Codex、Google 的 Gemini、Anthropic 的 Claude。 像極了轉扭蛋的 vibe coding 御三家： openai 的 Codex google 的 Gemini Anthropic 的 ClaudeTue, 02 Sep 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/vibe-coding-three-giants/https://nilsliu.dev/de/insights/vibe-coding-three-giants/Wie der Gacha-Automat der großen drei Vibe-Coding-Modelle: OpenAI's Codex, Google's Gemini und Anthropic's Claude.Tue, 02 Sep 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/gemini-flash-nano-banana/https://nilsliu.dev/en/insights/gemini-flash-nano-banana/Tested Gemini 2.5 Flash's image editing with three sequential prompts — suit, smile, tie adjustment. The precision was genuinely impressive.Wed, 27 Aug 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/gemini-flash-nano-banana/https://nilsliu.dev/zh/insights/gemini-flash-nano-banana/Nano Banana（Gemini 2.5 Flash 圖像功能）真的滿猛的，從改西裝到調整笑容和細節動作，精準度令人驚艷。 Nano Banana （Gemini 2.5 Flash 圖像功能）真的滿猛的 prompt 1: 改成穿西裝在攝影棚裡拍的證件照 prompt 2: 改成帶有專業感...Wed, 27 Aug 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/gemini-flash-nano-banana/https://nilsliu.dev/de/insights/gemini-flash-nano-banana/Nano Banana (Gemini 2.5 Flash Bildfunktion) ist wirklich beeindruckend, von der Anpassung eines Anzugs über die Korrektur des Lächelns bis hin zu detaillier...Wed, 27 Aug 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/hiring-devops-fullstack-ddt/https://nilsliu.dev/en/insights/hiring-devops-fullstack-ddt/We're taking financial AI to Southeast Asia and looking for two engineers. DevOps and full-stack roles open in Taipei Xinyi.Thu, 07 Aug 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/hiring-devops-fullstack-ddt/https://nilsliu.dev/zh/insights/hiring-devops-fullstack-ddt/把生成式 AI、現代化金融科技、智慧金融應用變成可規模化的產品與服務，一起把金融 AI 做到海外市場。 尼爾斯在找隊友！我們要一起把金融 AI 做到海外市場（東南亞），兩個角色同步招募中 我們在做什麼 把生成式 AI、現代化金融科技、智慧金融應用變成可規模化的產品與服務，直接解決真實業務問題。 ...Thu, 07 Aug 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/hiring-devops-fullstack-ddt/https://nilsliu.dev/de/insights/hiring-devops-fullstack-ddt/Verwandeln generativer KI, moderner Finanztechnologien und intelligenter Finanzanwendungen in skalierbare Produkte und Dienstleistungen – wir bringen Finanz...Thu, 07 Aug 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-pm-peter-deng-uber/https://nilsliu.dev/en/insights/ai-pm-peter-deng-uber/Former OpenAI VP of Product Peter Deng details the essence of product, 1-to-100 growth strategies, the five PM archetypes, and the value of invisible AI.Sun, 20 Jul 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-pm-peter-deng-uber/https://nilsliu.dev/zh/insights/ai-pm-peter-deng-uber/前 OpenAI 產品副總裁 Peter Deng 深入解析產品本質、1 到 100 的增長策略、五種 PM 角色分類，以及在 AI 時代中最核心的無感智能化設計與提問能力。 身為 AI 領域的產品經理或技術人員，我非常信奉矽谷頂尖產品大師 Peter Deng（曾任職於 Facebook、In...Sun, 20 Jul 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-pm-peter-deng-uber/https://nilsliu.dev/de/insights/ai-pm-peter-deng-uber/Der ehemalige OpenAI VP of Product, Peter Deng, detailliert das Wesen von Produkten, Wachstumsstrategien von 1 auf 100, die fünf PM-Archetypen sowie den Wer...Sun, 20 Jul 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/chatgpt-agent-launch/https://nilsliu.dev/en/insights/chatgpt-agent-launch/ChatGPT's agent renders its operations in real-time with full verbosity — like watching a capable human assistant work at a workstation on your behalf.Thu, 17 Jul 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/chatgpt-agent-launch/https://nilsliu.dev/zh/insights/chatgpt-agent-launch/ChatGPT 的 Agent 操作都是即時顯示且具可解釋性，基本像盯著一位人類秘書操作電腦完成抽象任務。 Chatgpt出agent 看起來Agent收到指令後操作都是即時顯示在螢幕上，也有verbose因此具可解釋性，基本像盯著一位人類秘書操作電腦完成抽象任務 開發agent的成員在加入ag...Thu, 17 Jul 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/chatgpt-agent-launch/https://nilsliu.dev/de/insights/chatgpt-agent-launch/Die Agenten-Vorgänge von ChatGPT werden in Echtzeit angezeigt und sind nachvollziehbar, fast so, als würde man einer menschlichen Sekretärin dabei zusehen,...Thu, 17 Jul 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/patent-modular-ai/https://nilsliu.dev/en/insights/patent-modular-ai/Are LLM deployment costs skyrocketing? This article shares how a bank GenAI Product Manager used modular architecture design to customize AI systems on dema...Wed, 11 Jun 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/patent-modular-ai/https://nilsliu.dev/zh/insights/patent-modular-ai/大型語言模型部署成本越來越高？本文分享一位銀行 GenAI Product Manager 如何透過模組化架構設計，讓 AI 系統可依需求客製，降低硬體負擔並提升彈性。 剛開始做企業 GenAI 產品的時候，我也曾經迷信「模型越大越好、功能越多越強」。 直到我看到實際的部署成本。 一個部署在企業...Wed, 11 Jun 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/patent-modular-ai/https://nilsliu.dev/de/insights/patent-modular-ai/Explodieren die Bereitstellungskosten für LLMs? Dieser Artikel zeigt, wie ein Bank-GenAI-Produktmanager eine modulare Architektur nutzte, um KI-Systeme nach...Wed, 11 Jun 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/patent-multi-agent/https://nilsliu.dev/en/insights/patent-multi-agent/Traditional DBAs manage databases on experience, but under high concurrency and complex loads, that's not enough. This article shares how a GenAI Product Ow...Sun, 01 Jun 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/patent-multi-agent/https://nilsliu.dev/zh/insights/patent-multi-agent/傳統 DBA 靠經驗管資料庫，但在高併發、複雜負載的環境下，這不夠用。本文分享一位 GenAI Product Owner 如何設計多 AI 代理協作的資料庫智慧優化系統，並取得新型專利。 資料庫優化是一門黑魔法。 資深 DBA 靠的是多年累積的直覺：看到某個查詢 pattern，知道要加什麼 ...Sun, 01 Jun 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/patent-multi-agent/https://nilsliu.dev/de/insights/patent-multi-agent/Traditionelle DBAs verwalten Datenbanken nach Erfahrung, aber bei hoher Nebenläufigkeit reicht das nicht aus. Dieser Artikel zeigt, wie ein GenAI Product Ow...Sun, 01 Jun 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/patent-pii-filter/https://nilsliu.dev/en/insights/patent-pii-filter/When introducing an AI knowledge base query system in a bank, how do you prevent PII leaks without sacrificing response quality? This article introduces a G...Sun, 01 Jun 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/patent-pii-filter/https://nilsliu.dev/zh/insights/patent-pii-filter/在銀行導入 AI 知識庫查詢系統時，如何防止個資外洩又不犧牲回覆品質？本文介紹一個 GenAI Product Owner 設計的敏感資訊過濾與替換專利架構。 想像這個場景： 你終於把 RAG 系統上線了，客服人員用它查詢客戶資料、業務規則、內部 SOP……效果很好，大家都很開心。 然後法遵部門...Sun, 01 Jun 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/patent-pii-filter/https://nilsliu.dev/de/insights/patent-pii-filter/Wie verhindert man bei der Einführung eines KI-Wissensdatenbanksystems in einer Bank Datenlecks, ohne die Antwortqualität zu beeinträchtigen? Dieser Artikel...Sun, 01 Jun 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/stargate-ai-infrastructure/https://nilsliu.dev/en/insights/stargate-ai-infrastructure/Stargate is a 24/7 round-the-clock server construction project. When Americans start running shifts like this, you know this is a race they don't intend to...Wed, 28 May 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/stargate-ai-infrastructure/https://nilsliu.dev/zh/insights/stargate-ai-infrastructure/星際之門現在是一個 24 小時輪班建置的伺服器中心專案，當美國人也開始 24 小時在卷的時候，你就知道這場仗是輸不起的。 星際之門現在是一個24小時輪班建置的伺服器中心專案，當美國人也開始24小時輪班在卷的時候，你就知道這場AI領導地位的仗是美國輸不起的一場戰役Wed, 28 May 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/stargate-ai-infrastructure/https://nilsliu.dev/de/insights/stargate-ai-infrastructure/Das Stargate ist nun ein 24-Stunden-Schichtprojekt für ein Serverzentrum. Wenn die Amerikaner anfangen, 24/7 zu kämpfen, weiß man, dies ist ein Krieg, den s...Wed, 28 May 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/google-io-one-for-all/https://nilsliu.dev/en/insights/google-io-one-for-all/Relentless. Product after product, packed into 32 minutes. Google I/O felt less like a keynote and more like being underwater with no room to breathe.Wed, 21 May 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/google-io-one-for-all/https://nilsliu.dev/zh/insights/google-io-one-for-all/Google 發佈會密集地發表新產品展現態勢，在推出新功能的速度上，OpenAI 還是一間小公司。 一即全，贏者全拿 這段影片是Google這次發佈會的精華32分鐘版本，密集地發表新產品，就展現了態勢。如果以前AI技術發展對我來說是眼花撩亂的話，那麽這次就是壓的人無法喘氣 相較之下，在推出新功能...Wed, 21 May 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/google-io-one-for-all/https://nilsliu.dev/de/insights/google-io-one-for-all/Die Google-Keynote war eine dichte Abfolge von Produktankündigungen, die deren Haltung zeigte. Im Vergleich zur Geschwindigkeit neuer Funktionen ist OpenAI...Wed, 21 May 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/kobe-imagination-ai-magic/https://nilsliu.dev/en/insights/kobe-imagination-ai-magic/Elon told Kobe: imagination matters more than knowledge. Frieren said magic is a world of imagination. To me, Transformers and generative AI are exactly tha...Thu, 15 May 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/kobe-imagination-ai-magic/https://nilsliu.dev/zh/insights/kobe-imagination-ai-magic/Elon 認為學習知識後最重要的是想像力。芙莉蓮說魔法是想像的世界——Transformer 和生成式 AI 就是這樣的魔法。 2017年曾有新聞採訪Kobe，詢問他對Elon Musk的看法。當時Kobe提到，Elon認為學習知識後最重要的是想像力；無論你學了多少東西、看了多少書，如果缺乏想像...Thu, 15 May 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/kobe-imagination-ai-magic/https://nilsliu.dev/de/insights/kobe-imagination-ai-magic/Elon glaubt, Fantasie sei nach dem Erlernen von Wissen das Wichtigste. Frieren sagt, Magie sei die Welt der Vorstellung – Transformers und GenAI sind diese...Thu, 15 May 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/patent-wealth-dashboard/https://nilsliu.dev/en/insights/patent-wealth-dashboard/What are the true pain points of Relationship Managers? How does GenAI help them generate real-time personalized investment advice in conversations? This ar...Sun, 11 May 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/patent-wealth-dashboard/https://nilsliu.dev/zh/insights/patent-wealth-dashboard/理財專員的痛點是什麼？GenAI 如何讓他們在對話中即時生成個人化投資建議？本文分享一位銀行 GenAI Product Manager 的金融 AI 產品設計歷程與專利思維。 在銀行做 AI 產品，我花了很多時間跟理財專員（RM）聊天。 他們不缺努力，也不缺對客戶的了解。 他們缺的是：在客戶面...Sun, 11 May 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/patent-wealth-dashboard/https://nilsliu.dev/de/insights/patent-wealth-dashboard/Was sind die wahren Schmerzpunkte von Kundenberatern? Wie hilft GenAI ihnen, in Echtzeit personalisierte Anlageempfehlungen zu generieren? Dieser Artikel te...Sun, 11 May 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/ai-team-recruitment-genai/https://nilsliu.dev/en/insights/ai-team-recruitment-genai/Our financial AI team is looking for DevOps and data science professionals passionate about deploying generative AI applications in real production environm...Thu, 01 May 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/ai-team-recruitment-genai/https://nilsliu.dev/zh/insights/ai-team-recruitment-genai/我們的金融 AI 科技團隊正在尋找對 DevOps 與資料科學充滿熱忱的專業人員，攜手開發並落實生成式 AI 應用。 歡迎加入團隊，一起積極創造AI場景 ——— 我們的金融 AI 科技團隊正在尋找對 DevOps 與資料科學充滿熱忱的專業人員，攜手開發並落實生成式 AI 應用 工作內容： 1....Thu, 01 May 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/ai-team-recruitment-genai/https://nilsliu.dev/de/insights/ai-team-recruitment-genai/Unser Financial AI Technology Team sucht leidenschaftliche Experten für DevOps und Data Science, um gemeinsam GenAI-Anwendungen zu entwickeln und umzusetzen.Thu, 01 May 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/genai-tech-interview-revolution/https://nilsliu.dev/en/insights/genai-tech-interview-revolution/We challenged intern candidates to build a static website combining LLM and front-end skills in 60 minutes. The results changed how I think about what a hir...Mon, 28 Apr 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/genai-tech-interview-revolution/https://nilsliu.dev/zh/insights/genai-tech-interview-revolution/要求候選人在 60 分鐘內結合 LLM 與前端開發能力完成靜態網站，從「知識背誦」轉向「工具落地」。 生成式 AI 驅動技術面試革新 最近，我們將實習生技術評估提升到新層次： 要求候選人在 60 分鐘內，結合 LLM 與前端開發能力，完成一個自我介紹＋OpenAI 提示工程指南的靜態網站 三位候...Mon, 28 Apr 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/genai-tech-interview-revolution/https://nilsliu.dev/de/insights/genai-tech-interview-revolution/Die Kandidaten müssen innerhalb von 60 Minuten mit LLMs und Frontend-Entwicklung eine statische Website fertigstellen, wodurch der Fokus vom 'Wissensgedächt...Mon, 28 Apr 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/ithome-devopsdays-2025/https://nilsliu.dev/en/insights/ithome-devopsdays-2025/This is the year of AI Agents. Join me at DevOpsDays on June 5–6 for a session on five agent behavior patterns and building the future DevOps ecosystem with AI.Wed, 02 Apr 2025 00:00:00 GMThttps://nilsliu.dev/zh/insights/ithome-devopsdays-2025/https://nilsliu.dev/zh/insights/ithome-devopsdays-2025/今年是 AI Agent 的年，6/5-6/6 DevOpsDays 見——五種 Agent 行為模式：用 AI 打造未來 DevOps 生態系。 A deep dive into the latest insights and developments.Wed, 02 Apr 2025 00:00:00 GMThttps://nilsliu.dev/de/insights/ithome-devopsdays-2025/https://nilsliu.dev/de/insights/ithome-devopsdays-2025/Dieses Jahr steht im Zeichen der KI-Agenten. Wir sehen uns auf den DevOpsDays am 5. und 6. Juni – Fünf Verhaltensmuster von KI-Agenten: Mit KI die Zukunft d...Wed, 02 Apr 2025 00:00:00 GMThttps://nilsliu.dev/en/insights/openai-devday-2024/https://nilsliu.dev/en/insights/openai-devday-2024/OpenAI showcased four major innovations: Vision Fine-Tuning, Realtime API, Model Distillation, and Prompt Caching — handing more creative control to develop...Wed, 02 Oct 2024 00:00:00 GMThttps://nilsliu.dev/zh/insights/openai-devday-2024/https://nilsliu.dev/zh/insights/openai-devday-2024/OpenAI 展示了四個重要創新：視覺微調、實時 API、模型蒸餾和提示快取，將更多的創新和控制權交到開發者手中。 OpenAI DevDay 2024: 一場屬於開發者的技術革新 就在今天，OpenAI 展示了四個重要創新：視覺微調（Vision Fine-Tuning）、實時 API（Rea...Wed, 02 Oct 2024 00:00:00 GMThttps://nilsliu.dev/de/insights/openai-devday-2024/https://nilsliu.dev/de/insights/openai-devday-2024/OpenAI zeigte vier große Innovationen: Vision Fine-Tuning, Realtime API, Model Distillation und Prompt Caching, wodurch mehr Kontrolle in die Hände der Entw...Wed, 02 Oct 2024 00:00:00 GMT