首先体验的是 deepseek 的一个模型,翻译速度也很慢(我猜测是因为有推理思考过程),在网上多方了解,发现Hunyuan-MT-7B倒是一个不错的专为翻译而生的小模型,腾讯混元官方的 api 依然是收费的,还好硅基流动限免提供使用。最后我选择了三个体量都不大的小模型:Hunyuan-MT-7B、Qwen2.5-7B-Instruct和GLM-4-9b-chat。
Remember 2022, the golden era of the internet? Back then, a reel popped up on my Instagram feed. Security footage captures a man sitting on a Target-ball in front of a store. Target-ball? You know what I mean, those large red concrete balls spaced in front of Target’s glass entryway. They make smash-and-grab rammings nearly impossible. In the footage, an out-of-control car slams into the ball next to where the man sits. The ball caroms into the ball beneath the man, knocks it away like a billiard ball, and the new ball stops directly under the man’s butt. One ball replaced the other. The man barely reacts. He probably doesn’t even know he’s sitting on a new ball yet.
“Whoa!” I watch it again and again. I can’t see any evidence that it’s fake. Convinced, I share it with my family. Real life is better than fiction. I love stuff like this. Loved! In 2022, I could watch it all day. That’s in the past now. Today, anything unbelievable is assumed to be AI.
Last week, I saw a video on Facebook. A grizzly draped itself over a car. It inadvertently pushed the car down the road as it tried to find a good angle to lick spilled food off the car hood. Realistic? Horns blared, the bear huffed, the car beneath the bear struggled under the added weight. Even the camera angle looked believable, as if the photographer tried to keep an extra car between themselves and the bear. Everything appears legit, but the bear is too big, even for a grizzly. It’s as big as the car.
I’ve seen a grizzly up close. The Cabela’s Sporting Goods in Hamburg, Pennsylvania has a taxidermized grizzly standing upright in their store. When my kids were young, we often stopped there for bathroom breaks as we drove to visit their grandparents. Cabela’s had the grizzly, a couple of elk and several huge aquariums filled with lake trout. It’s a nice break after a few hours in the car. The grizzly is massive. It’s unbelievable, frightening, awe inspiring, but it’s not as big as a car. Probably.
Here’s the thing. Maybe the grizzly video could be authentic. Maybe grizzlies actually grow that big. What do I know. I’m a city-guy from Washington, DC. But it no longer matters. I can’t tell the difference between real and fake. And if it isn’t real, what’s the point. The video isn’t exciting if Google Veo 3 thought it up. It’s impressive that AI has advanced so far since Chat GPT’s splashy release a couple of years ago, but it has ruined the magic of the unbelievable. It put an end to wonder.
Yesterday, the New York Times published a test. They posted ten short videos. Readers watched the videos and then guessed whether the content was genuine or AI generated. I got seven out of ten correct, but not because I could spot AI, the videos all looked real to me. I just used basic psychology. The more outlandish the video seemed, the less likely I was to call it AI. The most mundane videos, a guy livestreaming as he walked down a dirt road, two news anchors introducing themselves, those were fake. A whimsical clip of a model releasing balloons into the sky while flapping birds surround her, that one is real.
This is my question, my fear. Will I ever be stunned by a photograph again? A list of some of the greatest, most recognizable photos in history: Charles Ebbets – Lunch atop a Skyscraper; Nick Ut – Napalm Girl*; Alfred Eisenstaedt – V-J Day in Times Square; Steve McCurry – Afghan Girl. The next time a world-changing photograph is published, will we even know if it’s real? Will an artist capture a unique and beautiful (or terrible) moment in time, or will a clever app simply generate something sure to stir those idiotic humans who keep the electricity running.
As if to put an exclamation point on this thought, just before bed last night, I saw a tornado reel on Facebook. Tornados fascinate me. The raw, focused power makes the ‘finger of God’ analogy I’ve heard since childhood the most appropriate descriptor. The tornado in this video was a monster, ever approaching the camera as it tore a swath across the barren countryside. It’s exactly the sort of video I would watch repeatedly, mesmerized by the awful beauty of nature. Instead, I gave my head a quick shake swiped to watch Anatoly prank another room full of muscleheads. Hmmm, I wonder if that one was real. It’s much easier to create these videos on a computer than find a group of weightlifters who haven’t heard of Anatoly.
*The attribution of Napalm Girl is currently in dispute. World Press Photo has determined that it’s possible Nick Ut, did not shoot the photo. “’Visual and technical’ evidence ‘leans toward’ an emerging theory that a Vietnamese freelance photographer, Nguyen Thanh Nghe, took the photo.”
强化学习入门:从试错中学习的艺术
Reinforcement Learning 101: The Art of Learning by Trial and Error
深度解析强化学习:Q-Learning算法详解
Deep Dive into Reinforcement Learning: Understanding the Q-Learning Algorithm
机器如何学会自己做决定?强化学习告诉你答案
How Do Machines Learn to Make Their Own Decisions? Reinforcement Learning Explained
从奖励中学习:人工智能的“试错智慧”
Learning from Rewards: The Trial-and-Error Intelligence Behind AI
第一次私校家长会: 原来家长比孩子还卷 前几天参加了娃的第一次家长会,和几位家长聊下来,真是个个都很厉害。不光孩子们卷,家长也一样卷,一眼望去基本都是 Dr/博士。娃还调侃我一句:“这有什么的,你不也是 Dr 吗?” 我心里默默想:还好没写学校名字,不然我这野鸡大学的头衔真拿不出手 😂。 私校里真是人才济济,乐器过 8 级的太常见了,卷得不得了。我还问过娃,是想当 big fish in a small pond...
给孩子第一台NUC小电脑 Next Unit of Computing Next Unit of Computing (NUC) is a line of small-form-factor computers...
和媳妇约个会: 剑桥的过桥米线 Dumpling Trees Dumpling Trees 是位于剑桥 Cherry Hilton 附近的一家中式餐厅,以云南特色的过桥米线闻名。店内环境宽敞整洁,菜品丰富,除了经典的米线,还有各类小吃、烧烤和炒饭,味道地道,分量十足。过桥米线的汤底鲜香,配料新鲜,包括鸡肉、鱿鱼、虾等食材,顾客可以自己下锅涮熟,既好吃又有趣。餐厅提供免费停车,但需在店内登记车牌,适合家庭聚餐或周末小聚。 剑桥 Cherry Hilton 那边有一家叫 Dumpling Trees 的过桥米线店,两三年前的冬天我们去吃过一次(剑桥 Dumpling Tree...
微信PC端程序占用了1.39 TB的空间! 快速清理微信占用空间 前两天我的 C 盘剩余空间突然变红了,我随手一查,竟然发现微信 PC 端程序居然占用了 1.39 TB 的空间,简直不可思议。在手机上,微信同样是名列前茅的“吞空间大户”,在 设置 → 通用 → 手机存储空间 里几乎稳居第一。 更离谱的是,这些空间大多并不是因为聊天记录,而是各种缓存文件、视频、图片和被动接收的文件所堆积起来的。平时我们只是点开看一眼,就算没保存下来,微信也会悄悄把它们留在本地,占据大量磁盘。尤其是群聊里转发的视频和文件,日积月累就成了一个“隐形黑洞”。...
C++的左值/lvalue, 右值/rvalue和右值引用/rvalue references C++ 左值(lvalue)、右值(rvalue)与右值引用(rvalue reference) 理解 C++ 中的左值、右值及其引用形式,是掌握现代 C++(尤其是 C++11 以后的移动语义/move和完美转发/perfect forwarding)必不可少的基础。 📌 什么是左值(lvalue) 左值指的是有名字、可寻址的对象,通常可以出现在赋值语句的左侧。 int x...
A beautiful young East Asian woman with a slim face and long dark hair sits on a wooden stool in a cozy Japanese izakaya. She wears a white button-up shirt tucked into high-waisted houndstooth shorts and beige pantyhose. She turns her upper body slightly to look back over her shoulder, smiling warmly at the camera with a soft and natural expression. The scene is warmly lit with dim yellow light, surrounded by other patrons and decorated with Japanese calligraphy posters, vintage wooden walls, and nostalgic details. The photo has a film-style grainy texture and shallow depth of field, evoking a retro, candid atmosphere.
第一次私校家长会: 原来家长比孩子还卷 前几天参加了娃的第一次家长会,和几位家长聊下来,真是个个都很厉害。不光孩子们卷,家长也一样卷,一眼望去基本都是 Dr/博士。娃还调侃我一句:“这有什么的,你不也是 Dr 吗?” 我心里默默想:还好没写学校名字,不然我这野鸡大学的头衔真拿不出手 😂。 私校里真是人才济济,乐器过 8 级的太常见了,卷得不得了。我还问过娃,是想当 big fish in a small pond...
再次换回 trae,居然发现它比 codebuddy 好用太多,它好像更能理解我的表述,分析代码的逻辑性也更强,修改一处代码后也会记得同时修改受到影响的其它模块。一开始它在修改代码后会在本地启用 npm run dev,但是它会开多个终端,然后发现端口一直被占用,后来就不用开发模式,而是直接在本地构建 npm run build,终端有报错的话它还会再次修改代码。
Master Universal Robot MCP server integration with this step-by-step tutorial. Learn installation methods, robot connection management, real-time status monitoring, joint and linear motion control, safety protocols, and simulation mode. Perfect for developers wanting to enable AI assistants to control Universal Robots through the Model Context Protocol interface.
Learn how to install and run Claude Code in Termux on Android, and seamlessly configure Claude Code Router to flexibly switch AI models for coding tasks. This tutorial covers setup, configuration, and practical tips for boosting development productivity with a powerful, customizable AI coding environment on your mobile device.