Yeemio

Good morning. I slept on it, and I still can’t believe it. To me, Jensen Huang is the soul of NVIDIA, just as Elon Musk is to Tesla, Bill Gates to Microsoft, Mark Zuckerberg to Meta, and Steve Jobs to Apple. The last time I saw a company push out the person who defined it, that person was Steve Jobs. And we all know how that comeback turned out. I hope it won’t be long before we see the man in the leather jacket back on top.

A hypothetical football research task shows how to distinguish a projected lineup from confirmation, block an unauthorized note upload, and keep the research moving locally.

@grok Are you overwhelmed?

I keep asking ChatGPT, “What else could be wrong?” It feels like opening Pandora’s box. More and more strange issues keep coming up, each sounding plausible but not quite right. Most of the time, I just keep saying, “Keep going. Keep going. Keep going.” I can’t seem to stop. Does anyone else do this? How do you decide when it’s time to stop?

有一说一，不是暴论 马斯克给大家发工资的准时性，超越了中国 99.1% 的企业。

I unfollowed Cat Wu. I had a lot of respect for her as a product manager who had worked at Anthropic and now works at OpenAI. In this AI era, that experience meant something to me. I thought she was very capable. But the last post I can see from her is dated November 5, and I haven’t seen her engage with the community since. I don’t think a product leader whose visible community engagement amounts to a joining announcement is doing enough to listen to the broader OpenAI user base and build products that serve them well. Please forgive me for using a GIF made with OpenAI’s image generation to express my respect for her.

Nineteen Independent Games Festival jurors say they won’t vote for games that used generative AI in any capacity. That’s their own voting decision, not an IGF ban. Big Walk’s developers say they used a chatbot to help solve a shadow bug that slowed the game at sunrise and sunset. After some dead ends, it suggested a fix. The team says they extensively modified the suggestion, ending up with about 10 lines of code. Treating that debugging help the same as every other use of generative AI is a lazy standard. And honestly, I can’t help wondering whether some of the loudest “AI is bad, full stop” takes are more about getting attention than seriously engaging with how the technology is used. teapawt.net/house_house_/status/20…

有没有最近去过故宫的小伙伴。 看看Fable5.2 对于故宫的还原度有多高？？ 我每次去北京都预约不到，已经有10来年没去过故宫了。 不知道现在变化大不大。

Another all-in-one software security issue has blown up. For the vast majority of affected workspaces, Notion says the AI chats involved in the Fireworks incident weren’t labeled with a user or workspace. On November 11, Notion also said it had revoked exposed Notion API keys found in affected users’ messages. Both can be true. A chat doesn’t need an account label to contain sensitive information. Notion says the incident affected fewer than 1% of its workspaces. The affected data included Notion AI chats sent to Fireworks for its Insights feature. Notion says its own systems weren’t breached. If you or your workspace received a notice, review the affected data using Notion’s instructions. If you pasted passwords or API keys into relevant Notion AI chats, rotate them. notion.com/en-gb/security/ad…

Tell your AI assistant to book a flight. You’ve approved the fare. But the airline may not let your assistant in. United Airlines says it is taking steps to restrict access by unauthorized third-party personal AI agents, both on its own platforms and through travel agencies and aggregators. The airline cites risks including exposed account credentials and unintended purchases. It also points to check-in, where a person must acknowledge restrictions on prohibited items. At an industry event the same day, JetBlue’s president said airlines should adapt if customers want AI booking tools, while warning that intermediaries could add costs. I don’t want an AI making that safety declaration for me. But if I approve the fare, I’d like to choose the assistant that helps me book it. reuters.com/legal/litigation…

Some comments pointed out that coordinating files isn’t enough if work stays stuck when a worker stops. I took that feedback, iterated, and published OwlRunKit 0.25.4 on npm with an explicit recovery path. Imagine a 20-hour task. A subagent crashes halfway through and can’t restart itself. OwlRunKit 0.25.4 gives another agent a way to pick up the saved work. First, register the task with an owner, a designated successor, and the files to preserve. As the worker runs, have it save checkpoints. Then prepare a skill or prompt in Codex or Claude Code so the host can bring in the successor and have it issue a recovery request using the observed task generation and revision. OwlRunKit checks the named successor and task state. If the original worker is still marked as running, it sends that worker a wake request and waits. A simple acknowledgement isn’t enough; a changed checkpoint linked to that request is evidence of progress. If progress is confirmed, the original keeps ownership. If the worker has exited or the wait ends without confirmed progress, the successor can continue from the last checkpoint in a new, isolated workspace. The old attempt can no longer submit checkpoints or write into that workspace. A timeout without confirmed progress is recorded as unconfirmed, not as proof of a crash. I verified this explicit recovery path on my OpenAI dot’s cloud computer using real processes running controlled worker scripts.

The crazy day brought by a cat is over, and I'm back to the dull, dry routine of posting technical articles. Occasionally grabbing coffee, occasionally venting. Hope you all show some support!

A hypothetical case shows how an agent should respond when access to a paid data service is revoked mid-task: block new requests, check whether sent requests can be canceled, and keep working with saved evidence and public sources.

Wow. This is astonishing. One-take shot. Forget which model made it for a moment. Just look at the animation itself. Whoever came up with this script is on another level. It plays like a miniature history of art and visual media, told through a pelican riding a bicycle. The same bird travels through paintings, moving images and screens. The closing overview numbers 21 scenes: the first 20 explore different styles or media, and the last delivers the punchline. Here is the sequence. The timestamps are approximate, and some visual references are easier to identify than others. Ancient art and traditional painting Around 0:05, prehistoric cave paintings Ochre-red cave walls, handprints and animals evoke Paleolithic cave art. The pelican rides a bicycle through a visual world that predates the bicycle by thousands of years. The absurdity begins right here. Around 0:10, ancient Egyptian murals Figures shown in profile, hieroglyphs and Nile plants evoke the wall paintings of ancient Egyptian tombs or temples. Around 0:18, Han dynasty stone reliefs White lines against a black background, horse-drawn vehicles and processions of figures resemble Han dynasty pictorial stone reliefs and the ink rubbings made from them. Around 0:24, a traditional Chinese figure handscroll Flowing robes and lightly coloured landscapes recall the Eastern Jin painting tradition associated with Gu Kaizhi. This may be a nod to Luoshen Fu Tu, commonly known in English as The Nymph of the Luo River, although I cannot confidently identify the exact work. Around 0:30, Northern Song landscape painting Towering mountains and tiny travellers at their feet strongly recall the composition of Fan Kuan’s Travelers among Mountains and Streams and the monumental landscape tradition. Around 0:35, a Northern Song city handscroll Bridges, waterways, shops and pedestrians bring to mind Zhang Zeduan’s Along the River During the Qingming Festival. Around 0:41, Maya-inspired imagery Borders of colourful glyphs and stepped temples resemble Maya imagery or codices. “Maya-inspired” is the safer description here; the brief sequence does not establish a specific source artwork. Around 0:51, Persian miniature painting Ornamental architecture, courtyards and intricately drawn figures evoke the Persian miniature tradition within the Islamic world. Around 0:55, late Ming and early Qing expressive ink painting Lotus leaves, generous empty space and a bird rendered with spare brushstrokes point toward Chinese xieyi, or freehand expressive painting. The image even includes “八大山人” (Bada Shanren) and “白眼向青天,” a phrase about turning a disdainful eye toward the heavens. This is a nod to the painting style of Zhu Da, better known as Bada Shanren. Around 1:00, Javanese shadow puppetry A warmly lit screen, intricately cut-out figures and pointed stage scenery evoke Wayang Kulit, the shadow-puppet theatre of Java, Indonesia. This represents a performing-arts tradition rather than a single historical date. Around 1:05, Edo-period Japanese ukiyo-e Mount Fuji, lines of rain, flattened waves and travellers on the road evoke the visual language of Edo-period woodblock prints. Modern art and digital media Around 1:12, Post-Impressionism The swirling blue night sky and stars clearly draw on the visual language of Van Gogh’s The Starry Night, placing this reference in the late 19th century. Around 1:18, Expressionism The distorted bridge and figures recall Edvard Munch’s The Scream, another reference rooted in the artistic world of the late 19th and early 20th centuries. Around 1:23, black-and-white silent cinema Huge shadows and stark contrasts between light and dark evoke early 20th-century German Expressionist film. There is something of Nosferatu about it, but I would not claim that it recreates a particular film. Around 1:29, electronic displays Green lines appear on a screen resembling an oscilloscope or radar display. The pelican moves from art on physical surfaces into electronic imagery. Around 1:35, 8-bit video games Pixel art and side-scrolling movement evoke the Famicom/NES era of video games, broadly recalling the 1980s. Around 1:40, early personal computers and the internet A classic Windows-style desktop and web forums recall the internet culture of the 1990s and early 2000s. Around 1:47, vaporwave Retro computers, neon colours and the word “VAPORWAVE” explicitly point to the internet aesthetic that emerged in the 2010s. Around 1:53, vertical short-form video The pelican enters a vertical phone screen and social-media-style imagery, bringing the journey into the era of mobile internet and short-form video. Around 1:59, generative AI Code and an image-generation interface bring us to the present, when software and generative models are used to create images and video. The final joke Around 2:12, the press conference An empty chair is labelled “鹈鹕（本人）,” or “Pelican (in person).” The subtitle says that the pelican itself has not yet commented. This is not another artistic era. It is the film’s final joke. What makes the whole thing work is that the pelican and bicycle stay much the same. What changes is how humans record, depict and circulate images. From cave walls to painted scrolls, cinema, pixel displays and AI generation, one absurd but continuous protagonist connects thousands of years of visual media. A note on the references: these are interpretations of the imagery, not an official scene-by-scene guide from the creator. The exact works behind the Chinese figure handscroll, Maya sequence and silent-film scene remain uncertain. ----------------------------------------------------------- 约 0:05，史前岩画 赭红色洞壁、手印和动物，模仿旧石器时代的洞穴绘画。鹈鹕骑着自行车，出现在远早于自行车发明的视觉世界中。荒诞感从这里开始。 约 0:10，古埃及壁画 侧身人物、象形文字和尼罗河植物，模仿古埃及墓室或神庙壁画。 约 0:18，中国汉代画像石 黑底白线、车马和人物队列，接近汉代画像石及其拓片的视觉效果。 约 0:24，中国古代人物长卷 人物衣带飘动，配上浅色山水，让人想到东晋顾恺之的绘画传统。这一幕可能在致意《洛神赋图》，但还不能确定具体对应哪一幅作品。 约 0:30，北宋山水 高耸山体与山脚下渺小的行旅，明显接近范宽《溪山行旅图》的构图，以及北宋巨碑式山水的传统。 约 0:35，北宋市井长卷 桥梁、河道、店铺和行人，让人想到张择端的《清明上河图》。 约 0:41，中美洲玛雅风格 彩色字形边框和阶梯形神庙，接近玛雅图像或抄本。这里说“玛雅风格”更稳妥，仅凭这段短镜头，还不足以确定具体的原作。 约 0:51，波斯细密画 装饰性建筑、庭院和刻画细密的人物，接近伊斯兰世界中的波斯细密画传统。 约 0:55，中国明末清初写意水墨 荷叶、大面积留白与简笔画出的鸟，指向中国写意绘画。画中还能看到“八大山人”和“白眼向青天”的文字，这是在致意八大山人朱耷的画风。 约 1:00，爪哇皮影戏 暖色幕布、精细镂空的人物和尖形舞台布景，接近印度尼西亚爪哇的 Wayang Kulit 皮影戏。它代表的是一种表演艺术传统，而不是某个单一的历史年份。 约 1:05，日本江户时代浮世绘 富士山、雨线、扁平的水波，以及路上的行人，接近江户时期木版画的视觉语言。 约 1:12，后印象派 旋转的蓝色夜空和星光，明显借用了梵高《星月夜》的视觉语言，对应的是 19 世纪末的艺术背景。 约 1:18，表现主义 扭曲的桥与人物，让人联想到爱德华·蒙克的《呐喊》，同样属于 19 世纪末至 20 世纪初的艺术语境。 约 1:23，黑白默片 巨大的影子和强烈的明暗对比，接近 20 世纪初德国表现主义电影。有《诺斯费拉图》一类作品的味道，但不宜断言它是在复刻某一部电影。 约 1:29，电子显示时代 绿色线条出现在类似示波器或雷达的屏幕上。鹈鹕从实体表面上的艺术，进入了电子图像的世界。 约 1:35，8-bit 电子游戏 像素画与横版移动，致意红白机时代的电子游戏视觉，大致让人想到 1980 年代。 约 1:40，早期个人电脑与互联网 经典 Windows 式桌面和论坛网页，接近 1990 年代至 2000 年代初的网络文化。 约 1:47，蒸汽波 复古电脑、霓虹色调，加上画面中直接出现的“VAPORWAVE”，明确指向 2010 年代兴起的网络美学。 约 1:53，竖屏短视频 鹈鹕进入手机竖屏和社交媒体式画面，来到移动互联网与短视频时代。 约 1:59，生成式 AI 代码与图像生成界面，把旅程带到当下：我们开始用程序和生成式模型创作图像与视频。 约 2:12，记者会结尾 一把空椅子标着“鹈鹕（本人）”，字幕说，鹈鹕本人尚未发表评论。 这不是另一个艺术时代，而是全片最后的玩笑。