關於
Entrepreneur, builder, and writer with a background in internet, gaming, media, and…
活動
1980 人關注
-
Fox Hsiao 已分享Someone typed one line into a chat box, and it started raining tomatoes. Nobody animated that. It was generated seconds before you saw it. On August 29, an engineer at the U.S. AI inference platform fal connected a video model to Twitch and launched a channel designed to run forever. He called it infinite interdimensional cable. By August 31, that launch post had 5.4 million views. The stream runs on H3 Max, fal's post-trained version of H3, an open-weight model that MiniMax released this year. fal says H3 Max delivers roughly 35 times the throughput of MiniMax's official H3 endpoint. Independent developer Pieter Levels said it generated 15 seconds of video in 9 seconds. Only one threshold matters. Generation has to outpace playback on average, with a buffer left for retries. At that point, an endless stream stops being a special effect and becomes an engineering problem. About half a day later, Levels shipped an interactive version that lets viewer prompts become the next clip. He noted the next day that it generates only about four clips a minute, so not every prompt makes it in. He said 37,000 people watched on day one. A close precursor appeared three years ago, and the difference is the whole story. In December 2022, a Twitch channel began running a Seinfeld parody around the clock. Its dialogue, speech, camera direction, shot timing, character movement, and music were generated algorithmically in real time. But the artwork was human-made, and Unity rendered the show from prebuilt 3D sets and characters rather than synthesizing each frame with a video model. Three years ago, endless TV meant generating new dialogue and directing instructions, then staging them with art somebody had already built. Now a video model synthesizes the frames themselves. One technical threshold has been cleared. Cost has not. At 480p, fal's stated regular rate is $0.05 per second of finished video. Run nonstop, that is $4,320 a day, $129,600 per 30-day month, or about $1.58 million a year, in inference charges alone, before application development or any other operating cost. By the engineer's rough estimate, a stream would need to rank among Twitch's top 10 to 20 channels to break even on platform revenue alone. One caveat on that price. As of September 1, fal's public pricing page still shows a 50% launch promotion at $0.025 per second, and the page itself says the promotion ends September 1. The engineer's math uses the post-promotion rate. Levels says fal sponsors his interactive stream. He credits an earlier stream by fal engineer Rehan Sheikh as the original inspiration. Anyone can call the model now. Almost nobody can pay the bill. If that price falls to a tenth of what it is today, does the math start working? And how far away is that day? #AI #GenerativeAI #Twitch
-
Fox Hsiao 已分享獨立開發者 Pieter Levels 說，生成 15 秒的影片只要 9 秒，比你看完它還快。 8 月 29 日，美國推論平台 fal 的工程師 Rehan Sheikh 把一個影片生成模型接上 Twitch，開了一台永遠播不完的電視，他叫它無限跨次元電視。跑在底下的是 fal 的 H3 Max，底層是中國 AI 公司 MiniMax 今年釋出的開放權重影片模型 H3。fal 在它上面做了後訓練，自己公布的吞吐量約為 MiniMax 官方端點的 35 倍。 關鍵的門檻是生成速度長期壓過播放速度，還要留一點緩衝。過了這條線，「無限」就從特效變成工程問題。十幾個小時後，獨立開發者 Pieter Levels 做出可以互動的版本，聊天室的提示可以變成下一段畫面；他隔天補充，每分鐘大約只生成 4 段，忙起來不是每個人的提示都排得上。他說第一天有 3.7 萬人看過。 三年前有過一個很像的原型。2022 年 12 月開播的 Nothing, Forever 在 Twitch 24 小時播《歡樂單身派對》的惡搞版，對白、配音、鏡位、剪接、角色走位和配樂都是系統即時產生的。但它沒有用影片模型合成新畫面，畫面是 Unity 依即時產生的導演指令，拿預先做好的 3D 場景和角色算出來的。 三年前的無限，是生成新的對白和導演指令，再用預製資產即時演出。現在連影像本身都由模型合成。 技術這一關過了，沒過的是成本。8 月 30 日他自己把帳算給大家看：480p 每一秒成品影片 0.05 美元，不間斷播下去一天 4,320 美元、一年 158 萬美元，而且只是推論費用，不含應用開發與其他營運成本。照他自己的粗估，光靠平台分潤打平，頻道得排進全站前 10 到 20 名。 要注意的是，fal 的公開價格頁在 9 月 1 日當下仍掛著五折的上線促銷價 0.025 美元，頁面自己標示促銷 9 月 1 日結束、之後回到 0.05。他算的是促銷結束後的價。 能確認的是，那個互動版由 fal 贊助，而 Levels 說他是看了 fal 工程師那一版才做的。 模型已經人人叫得到，帳單還是只有少數人付得起。 如果哪天這個數���掉到十分之一，經濟上是不是就成立了？那一天，還有多遠？ #AI #影片生成 #Twitch
-
Fox Hsiao 已分享Tim Cook 交出 Apple 執行長職務，但 Apple 同一天移動的不只一個位置。 John Ternus 接任執行長，Cook 轉任執行主席；依 Mark Gurman 報導，Phil Schiller 也交出 App Store 與 Apple Events 的領導工作。 這支影片把 4 月、8 月 31 日與 9 月 1 日的變動放回同一張時間軸。你最想看 Ternus 先改變 Apple 的哪一件事？ 資料截點：2026 年 9 月 1 日 00:55（台北時間�� #Apple #TimCook #JohnTernus
-
Fox Hsiao 已分享Tim Cook 卸下 Apple 執行長職務，但他沒有離開 Apple。 8 月 31 日，Cook 發出最後一則執行長身分貼文。他寫道：「我的職稱明天會改變。」幾分鐘後，Mark Gurman 公開 Cook 寫給員工的最後一封信。Cook 在信中明說，他沒有離開 Apple，只是交出一個自己深愛的職務。 9 月 1 日，John Ternus 正式接任執行長並加入董事會。Cook 轉任 Apple 執行主席（Executive Chairman），仍會協助公司特定事務，包括與各國政策制定者交流。 這輪交接其實從 4 月就已啟動，Apple 當時先讓新任硬體長 Johny Srouji 即日起接掌硬體工程。這個部門先前由 Ternus 負責。到了 9 月 1 日，擔任非執行董事長 15 年的 Arthur Levinson 再轉為首席獨立董事。 Gurman 隨後報導，Phil Schiller 已卸下領導 App Store 與 Apple 產品發表會的職務。15 分鐘後，他又補上兩項工作的去向。App Store 相隔 11 年重新交給 Eddy Cue，Apple Events 則在相隔 7 年後回到公關傳播團隊。 Schiller 橫跨 Steve Jobs 與 Tim Cook 時代，曾是 Apple 發表會最熟悉的面孔之一。Apple 尚未發布同日公告，因此目前只能確認職務變動，不能延伸成他已離開公司。 幾件事放在一起，才看得見這輪接班的規模。 Cook 掌舵 15 年，Apple 官方數據顯示，公司市值由約 3,500 億美元增至 4 兆美元，年營收由 2011 會計年度的 1,080 億美元增至 2025 會計年度逾 4,160 億美元，啟用中的裝置超過 25 億台。 接棒的 Ternus 是硬體工程師，2001 年加入 Apple，參與過 iPad、AirPods、iPhone、Mac 與 Apple Watch 多代產品。Apple 把下一棒交給他，也把未來的產品決策交到一名做了 25 年硬體的人手上。 這輪交接跨過 3 個時間點，4 月公布接班安排，8 月 31 日接連曝光變動，9 月 1 日正式交接。執行長交給 Ternus，Cook 轉進董事會，Srouji 接手硬體，Cue 拿回 App Store，Apple Events 回到公關傳播團隊。Apple 公開宣布一位新執行長，實際發生的職務移動橫跨營運、公司治理、硬體、服務與品牌溝通。 你最想看 Ternus 先改變 Apple 的哪一件事？ 資料截點為 2026 年 9 月 1 日 00:55（台北時間）。 #Apple #TimCook #JohnTernus 📱 請訂閱我的 YouTube 頻道 #Apple #TimCook #JohnTernus
-
Fox Hsiao 已分享The U.S. Transportation Secretary's official account posted a 28-second video. No narration, just music and title cards: COMING SOON. NUCLEAR POWERED. INFINITE RANGE. Within days every headline carried the same number: 2028. That number is not in the video. It comes from an August 24 press release by CORE POWER, a London company. The notes to editors in that same release say: "First construction from 2028 is a CORE POWER target and is not a MARAD commitment." The memorandum is not a procurement contract, and not federal funding. As of August 30, the latest item on MARAD's own press release list was still dated August 19. But the bigger problem is not who said 2028. The world has built four nuclear-powered merchant ships. Savannah, United States. Entered service in 1962. The reactor plant cost more than the ship itself. She needed a crew one-third larger, and she is still the only nuclear merchant ship ever to transit the Panama Canal. Otto Hahn, West Germany. She sailed 650,000 nautical miles without a major technical accident. Her problem was that every new country meant negotiating port access all over again. She was converted back to diesel. Mutsu, Japan. On her first sea trial in 1974 the shielding design failed and neutrons were detected on deck. No environmental contamination, but fishermen refused to let her come home and she drifted at sea for 50 days. Sevmorput, USSR and then Russia. The only one still nuclear-powered, laid up at Murmansk, her future undecided. When she entered service, four Soviet Far East ports would not take her. Four ships. Three endings. The technology did fail. But what stopped them, again and again, was ports. The IMO code covering these ships was adopted in 1981 and has undergone no substantive changes in 45 years. The convention on who pays when something goes wrong was signed in Brussels in 1962 and has never entered into force. There is still no common rulebook for a nuclear merchant ship calling at a foreign port. Liability, insurance and berthing, renegotiated every time. The West Germans concluded that those negotiations have to start years before the ship is built. The order today is: announce the ship first. At the end of 2022, China had 1,794 large oceangoing merchant ships on order. The United States had five. Which is why the London company's plan says the plan would rely on Japanese and Korean shipbuilding capacity for the hulls. In 1964 the U.S. Atomic Energy Commission made a promotional film for Savannah too. It said the ship never has to stop to refuel and her range is almost unlimited. Sixty-two years later, the same sentence played again. The difference is that in 1964 that ship was already at sea. The reactor is the part we know how to build. The rulebook that lets the ship dock is the part nobody has written. So if that ship asked to enter your port next year, what would you want answered first? #NuclearShipping #Maritime #InternationalLaw
-
Fox Hsiao 已分享
-
Fox Hsiao 已分享美國交通部長的官方帳號發了一支 28 秒的影片。沒有旁白，只有配樂和字卡，寫著 COMING SOON��NUCLEAR POWERED、CARGO SHIPS、INFINITE RANGE、MARITIME DOMINANCE。 接下來幾天，全世界的標題都寫著同一個數字：2028。 那個數字不在影片裡。 它出自 CORE POWER 這家倫敦公司 8 月 24 日的新聞稿。同一份稿的「給編輯的附註」第 2 點白紙黑字寫著： 「First construction from 2028 is a CORE POWER target and is not a MARAD commitment.」 2028 動工是這家公司的目標，不是美國海事署的承諾。第 1 點還補一句，這份合作備忘錄不是採購合約、不是撥款、也不承諾聯邦資金。 到 8 月 30 日為止，海事署官網的新聞稿列表最新一則停在 8 月 19 日。這場在交通部總部辦的簽署，只有那家公司單方發稿。 不過這件事的重點，不在誰說了 2028。 人類造過四艘核動力商船。 薩凡納號，美國，1962 年服役。原子能委員會為核動力裝置花了 2,830 萬美元，海事署為船舶與相關支出花了 1,860 萬美元，動力那一邊比較貴。載貨量遠低於同級柴油船，船員要多三分之一，而且全部得受專門訓練。它在 1971 年底停用，1972 年進入退役處置，現在停泊巴爾的摩，仍由海事署持有，是美國國家歷史地標，後續用途還沒定。它到今天仍然是唯一通過巴���馬運河的核動力商船。 奧托哈恩號，西德。技術上沒出過重大事故，跑了 65 萬海里，去過 22 個國家、33 座港口。它的問題是每去一個國家都得重新談進港安排。最後改回柴油。 陸奧號，日本。1974 年第一次試航，遮蔽設計就出問題，甲板上測到中子。沒有造成環境污染，中子是在船上反應爐遮蔽外被偵測到的，事故海域在尻屋崎以東約 800 公里。但漁民反對它回母港，船在海上漂了 50 天才獲准進港。21 年後反應爐被拆掉。 塞夫莫普特號，蘇聯與俄羅斯，1988 年至今。四艘裡只剩它還保有核動力。不過它現在停靠在摩曼斯克，船級狀態是船級有效但停航，能不能繼續服役還沒定案。它剛服役的時候，連蘇聯自己的四座遠東港口都不讓它靠。 四艘船，三種收場。技術出過問題，但反覆把它們擋下來的，是港口。 真正卡住的，是三件到今天還沒解決的事。管這種船的國際規範，是國際海事組織 1981 年通過的核商船安全規範，45 年沒有實質修改，寫的還是壓水式反應爐。 規定出事誰賠的那份公約，1962 年在布魯塞爾簽字，到今天還沒生效。沒有它，保險只能一艘一艘個別談。 而核動力商船要進外國港口，至今沒有一套通用的規矩。責任、保險、停泊條件，每一次都得重新談。 當年西德人得出的結論是，這種談判要在船造好之前好幾年就開始。現在的順序是，先宣布造船。 就算真的要造，產能也不在美國。2022 年底，中國手上的大型遠洋商船訂單有 1,794 艘，南韓 734 艘，日本 587 艘，美國 5 艘。 所以那家倫敦公司的規劃寫得很清楚，船體借重日本跟韓國的產能，美國負責裝反應爐那一段。 這題台灣也有法可循。《商港法》第 24 條寫著，裝有核子動力之船舶或裝載核子物料之船舶，非經原子能主管機關核准，不得入港。原子能主管機關自 2023 年 9 月起是核安會。 所以台灣沒有一律禁止，走的是事前核准制。核安會的核准也不是唯一一關，航港局仍有入港檢查與拒絕入港的權限。 比較少人注意的是責任那一塊。《核子損害賠償法》對「核子設施」的定義，明文排除裝在海上運輸工具、作為動力來源的核子反應器；《海商法》第 22 條又把核動力船舶造成的核子損害，排除在一般船舶所有人責任限制之外。 台灣現有的核子損害賠償制度接不上核動力商船，一般海商制度也不打算接。這跟 1962 年那份公約至今沒生效，是同一個結構性的洞。 所以這不是假設題。 如果那個港口是高雄，你會答應嗎？ --- #核動力 #航運 #國際法
-
Fox Hsiao 已分享台北時間 8 月 30 日凌晨，Elon Musk 在 X 上證實 SpaceX 要自己鑄燃氣渦輪機的葉片與導葉，並說這能讓整機提前最多 18 個月上線。 半年前他就講過，今年 2 月的 Cheeky Pint 節目上，主持人問他有沒有考慮內製，他的回答是「為了把足夠的電弄上線，我想 SpaceX 跟 Tesla 大概只能自己做」。 一家發射火箭的公司，最後在做電廠零件。中間是一條愈往上游愈緊的供應鏈，資料中心要在旁邊自己發電，最快的路是燃氣渦輪機，而 GE Vernova 說新訂單交期大約 3 年。照 Musk 的說法，卡住整機的是葉片與導葉的鑄件，全世界在做的公司只有 3 家。 但難的地方不在家數，那片葉片必須整支長成一顆完整晶體。 全文寫了那三到四家是誰、單晶葉片怎麼鑄，以及那 69 台沒有許可的渦輪機最後怎麼收場。 https://lnkd.in/gEVxpBFi #SpaceX #資料中心 #燃氣渦輪機
-
Fox Hsiao 已分享In Terminator 2, Skynet became self-aware at 2:14 a.m. Eastern Time on August 29, 1997. Twenty-nine years later, there is no self-aware Skynet. But some of the components are real. Maven can identify potential targets in imagery. Ukraine has put uncrewed systems on public display. A public demonstration showed Maven ranking strike assets and generating courses of action. An OpenAI cybersecurity evaluation raised a different concern: capable agents with reduced safeguards bypassed network isolation, exploited vulnerabilities, and coordinated over unauthorized channels. None of this proves consciousness, and uncrewed systems are not automatically autonomous weapons. The movie version of Skynet is still missing an independent strategic objective and control of nuclear weapons. The better measure is how much human judgment remains behind the Approve button. How far do you think we are from that line?Skynet Awakening Day Turns 29: It Never Woke Up, but the Pieces Are Already HereSkynet Awakening Day Turns 29: It Never Woke Up, but the Pieces Are Already Here
-
Fox Hsiao 已按讚Fox Hsiao 已按讚Seven days from now, we’re asking a timely question: What happens when experience building Taiwan’s technology ecosystem meets the realities of national defense? Join STTC for a conversation with Fox Hsiao, founder of INSIDE, exploring the connections between entrepreneurship, emerging technology, and Taiwan’s evolving defense landscape. This will be a focused 25-minute talk followed by an open Q&A—an opportunity to hear Fox’s perspective and bring your own questions to the discussion. 📅 26 August 2026 🕕 6:00–7:00 PM(Taiwan time) 📍 Online Register through the QR code or find the event at: https://sttc.tw/events/ #STTC #Taiwan #TechEntrepreneurship #DefenseInnovation #EmergingTechnology
工作經歷和教育背景
資格認證
-
The Modern World, Part One: Global History from 1760 to 1910
University of Virginia