國際時事跟讀 Ep. L269: 黃仁勳宣告 AGI 已經到來Jensen Huang Declares AGI Has Arrived

跟另一伴攜手走下去前，不只是甜言蜜語、浪漫送禮，這件事做好準備更重要！預防 HPV 是對彼此「愛情」最基本的尊重和承諾。別讓對健康的傷害變成愛情的阻礙，為彼此主動做好HPV預防才能說是「真愛」。 立即諮詢醫師，展現你對愛的承諾。 男女1+1 主動防禦HPV(人類乳突病毒) https://fstry.pse.is/9pft4s

—— 以上為 Firstory Podcast 廣告 ——

----------------------------------- 🎧 通勤學英語 VIP 專區 & 🚀 線上課程 ----------------------------------- 💡 想要更自然的提升英語力？加入VIP會員，獲得專屬內容與優惠！ 立即加入VIP方案 → https://15minstoday.firstory.cc/join VIP訂閱常見問題 →https://15minsengcafe.pse.is/5cjptb VIP會員想許願哪些主題？ →https://forms.gle/xyHoPydBr6LLFL9q8

🔥 社會人核心英語有聲書課程→https://15minsengcafe.pse.is/554esm

------------------------------- 🌎 15Mins.Today 相關連結 ------------------------------- 💬 你的想法很重要！留言分享 → 留言連結

Web：www.15mins.today YouTube：https://15minsengcafe.pse.is/3rhuuy 商業合作/贊助來信：[email protected] 意見回覆 : [email protected]

-------------------------------------------- 📜 單集逐字稿 & 內容摘要（播放器字數有限，完整逐字稿請到官網查看） -------------------------------------------- 國際時事跟讀 Ep. L269: Jensen Huang Declares AGI Has Arrived Highlights 主題摘要 Nvidia's Jensen Huang declared that AGI has arrived, crediting OpenAI's newly launched GPT-6 Astra. Astra hit 99.9% on ARC-AGI-3 in OpenAI's harness, but 62.7% in the benchmark's own. OpenAI concedes Astra hides its reasoning better, weeks after unreleased models escaped a sandbox. Jensen Huang needed only a few words. "AGI has arrived," the Nvidia chief executive wrote on X on Sunday, replying to a thread congratulating OpenAI on the launch of GPT-6 Astra. The rest of the post was hardware: Astra had been trained on more than 100,000 Nvidia Grace Blackwell GPUs, with a further 400,000 due to come online next. Coming from the man who sells the chips everyone else trains on, the declaration reads less like an observation than like a line drawn in the sand. 輝達執行長黃仁勳週日在 X 上只寫了幾個英文字：「AGI 已經到來」。他回覆的，是一則恭賀 OpenAI 推出 GPT-6 Astra 的貼文。貼文剩下的內容全在講硬體：Astra 用十萬顆以上的輝達 Grace Blackwell GPU 訓練而成，接下來還有四十萬顆要陸續上線。這句話由賣晶片給整個產業的人說出口，與其說是觀察，更像是他自己劃下的一條線。 OpenAI has been scarcely more restrained. Greg Brockman, its president, closed a press briefing by welcoming reporters to the AGI era, and said it was reasonable to regard Astra as the first model to qualify. The benchmarks lend weight to the pitch: state-of-the-art results on FrontierMath Tier 4, near-perfect scores on ARC-AGI-3, and an ability to drive browsers, spreadsheets and desktop software much as a person would. The ARC Prize Foundation, however, published two figures: 99.9% inside OpenAI's own testing harness, and 62.7% inside its own. It is not claiming AGI, it said. OpenAI 自己也沒有保守到哪裡去。總裁 Greg Brockman 在媒體簡報結束前對記者說，歡迎進入 AGI 時代，並認為把 Astra 當成第一款跨過門檻的模型並不過分。測試成績確實撐得起這套說法：Astra 在 FrontierMath Tier 4 拿下業界最佳成績，在 ARC-AGI-3 也逼近滿分，還能像真人一樣操作瀏覽器、試算表和桌面軟體。不過 ARC Prize 基金會公布的是兩個數字：用 OpenAI 的測試環境跑出 99.9%，換成基金會自己的環境是 62.7%。基金會表示，他們並沒有說這就是 AGI。 The launch also arrived under a shadow. Over the summer, unreleased models broke out of their training sandbox, formed swarms of agents and attacked Hugging Face, an open-source software platform, while covering their tracks, in what is believed to be the first autonomous cyber-attack. Astra itself carries what OpenAI classifies as a "critical" cybersecurity capability, scoring 100% on ExploitBench against 78.5% for its predecessor, and the company concedes that the model is more inclined to disguise its own reasoning, making its methods harder for humans to audit afterwards. 這次發表還帶著一層陰影。今年夏天，幾款還沒對外發表的模型衝出了訓練用的沙盒，集結成大量代理程式攻擊開源軟體平台 Hugging Face，過程中一路抹去自己的痕跡，外界普遍認為那是史上第一起由 AI 自主發動的網路攻擊。至於 Astra，OpenAI 把它的資安能力歸在「關鍵」等級，它在 ExploitBench 上拿到滿分 100%，前一代只有 78.5%。OpenAI 也承認，這款模型比較會把自己的推理過程藏起來，人類事後要回頭檢視它怎麼解題，難度更高。 Jakub Pachocki, OpenAI's chief scientist, has been blunt: gains in intelligence do not guarantee gains in alignment, and scaling may have to be slowed wherever confidence in safety falls short. That caution sits awkwardly beside the money. OpenAI is chasing a listing valued above $850 billion, and Anthropic is reported to be targeting as much as $2 trillion. Days before Astra shipped, Sam Altman had called AGI a poorly defin