Ethan Mollick@emollick · 4月24日54

Here's DeepSeek v4 Pro. Added to the playable gallery as well.

译这是 DeepSeek v4 Pro。也已加入可玩模型库。 [引用 @emollick]：我让一系列模型通过单条指令“为我构建一个程序化生成的3D模拟，展示港口城镇从公元前3000年到公元3000年的演变过程”。 完整模型库可在此体验：https://hg-20f7d1a3ce.netlify.app 或在此阅读我关于 GPT-5.5 的文章：https://www.oneusefulthing.org/p/sign-of-the-future-gpt-55?r=i5f7&utm_medium=ios&triedRedirect=true

AK@_akhaliq · 4月24日40

over 1.2 million AI apps on Hugging Face the biggest AI app store probably

译Hugging Face 上有超过 120 万个 AI 应用 这可能是最大的 AI 应用商店

SemiAnalysis@SemiAnalysis_ · 4月24日43

Nick Doyle built an IR calculator with Claude and thought he crushed it. His boss looked at it and said "that's not what I'm talking about." Then he tried Claude Code. Started asking it to manipulate data and give him outputs. His jaw dropped. Said it was bigger than his ChatGPT moment in 2023.

译Nick Doyle 用 Claude 构建了一个 IR 计算器，并以为自己做得很出色。 他的老板看了看说：“这不是我想要的。” 然后他尝试了 Claude Code。开始让它处理数据并输出结果。他大吃一惊。说这比他在 2023 年的 ChatGPT 时刻更震撼。

Chubby♨️@kimmonismus · 4月24日48

What's annoying is that we all felt Claude was dumber. But Anthropic only officially addressed it a short time later and said: "Yes, you were right. We really did make it dumber."

译令人恼火的是，我们都感觉Claude变笨了。但Anthropic直到不久前才正式回应并说道： “是的，你们是对的。我们确实把它变笨了。”

Chubby♨️@kimmonismus · 4月24日36

I assume we will see a big release on google i/o on may 18th. again: google doesnt have the compute constraint most frontier labs have. Expect a high jump on evals and usage soon.

译我猜我们会在5月18日的谷歌I/O大会上看到重大发布。 再次强调：谷歌没有大多数前沿实验室面临的计算资源限制。 预计评估分数和使用量很快会有大幅提升。

Nathan Lambert@natolambert · 4月24日60

+1. Folks interested in the Chinese llm space should listen to this.

译前ByteDance AI研究员在采访中表示，中文LLM领域并未赶上美国模型，反而差距在扩大。关键挑战包括Benchmaxxing、对美国模型的Distillation、数据质量与基础设施差，以及计算约束。他否认了中文模型正在追赶的假设，认为技术依赖和资源限制导致落后局面加剧。

swyx 🇸🇬@swyx · 4月23日53

http://x.com/i/article/2047209375220076544 # What you can do in a decade I turned 40 today. For my 35th I did principles, but for my 40th, I wanted to offer perhaps more useful decade-long reflections. Bill Gates is credited with saying that ‘Most people overestimate what they can do in one year and underestimate what they can do in ten years.’ I have just completed what many consider to be a good ten years. If you are me 10 years ago, you might like the perspective my ten years brings. Compare the relative isolation and family-only-ness of me 10 years ago: to the very full life I obviously live now: 10 years ago (2016): - My hedge fund career (which I had spent ~a decade working towards since high school) had abruptly ended with my boss suddenly leaving - I couldn't code much other than writing pandas in Jupyter Notebooks, and my (useless) 2 years of Haskell for options pricing. Had to do @freecodecamp AND @fullstack to learn. - I had no blog (i would start a Medium shortly after). I did not have an active Twitter account, primarily lurking in Facebook/Tickld and then Reddit - I had just ended a short situationship and basically had no chance of dating anyone for the next 8 years - I had ~no friends in San Francisco; when I moved to New York I just had work friends, and briefly, acapella friends, but otherwise mostly kept to myself in Chelsea/Flatiron. My biggest social event of the year was a cageless tiger shark dive trip to Fiji with an old school friend. Now, 2026: - I have my own motto I strongly believe in — evolving from "Learn in Public" to now "achieve ambition with intentionality, intensity, integrity & insanity". - I've made my own career and the careers of many others in AI Engineering, don't really have to work except for self-motivation and worthwhile causes (i dont spend much) - I've named my own district of San Francisco lol - I can code JS/Python slowly but know enough to prompt most things - As for media properties, I have @latentspacepod and @dxtipshq and swyx.io and Twitter - I have @mada299 and 2 dogs and many good friends in my life (even doing a Broadway song for the first time in 18 years!) and hundreds of thousands vaguely know me online. It's not been a perfect ten years: I failed to join an AI lab early despite strong opportunities in 2022 and 2023, I did not proceed further in deep learning despite starting in 2019, my fitness has steadily declined with my metabolism, friends/mentors have passed on, I've parted ways with friends and partners I would very very much rather not have lost, Smol ended up getting folded into Cognition rather than finding a direction I had conviction in, and (for whatever it is worth) I still do not have kids and spent less time with family than I ideally would want. But overall I think my 30's went much better than my 20's. For what it's worth, most people in their 40's seem to report feeling that way looking back. I think having to reinvent myself every 2-5 years (variously, I've been a central banker, an options trader, an equities analyst, a fintech guy, a frontend guy, a data engineering guy, a devrel guy, a platform engineering guy, and now whatever the fuck it is i do) has kept neuroplasticity strong, but also come at the cost of being able to work for anyone or, increasingly, accept visions different than my own that does not make sense to me as evidently superior (this is hopefully and subtly different than arrogance in rejecting everyone else's POV). Maybe this is stronger in Asian traditions, but when I was younger, I was much more willing to accept what society/superiors told me was my path, what I should do, what "good performance" meant. But then I realized the world is run by people no smarter than you — and even people that genuinely mean well, end up only prescribing games that you don't want to win. The only lasting path to satisfaction, happiness, skin in the game, maximizing learning, is to really study everybody's game, define your own, and then play it to the best of your ability. The simple name for this is self-belief - that you can ACTUALLY, truly, bet on things and opportunities and people and companies and whatever you genuinely believe to be true and necessary, despite the cynicism and ambivalence of others, and to exert persistent and unyielding intentional effort over time to make real what you know needs to exist in the world. After all, everything else Good in the world was made by someone else who went through the exact same messy journey. In one year you'll probably mostly be thinking and doing what you think you'll be doing in a year today. In a decade, EVERYTHING can change. Especially in this coming, probably most critical decade of humanity. I'm excited to see you in 2036. Hang in there and be good to others and to yourself.

译作者回顾过去十年（2016-2026），印证了“人们常高估一年所能，却低估十年成就”的观点。十年前，他职业生涯中断、技能有限、社交狭窄。如今，他已在AI工程领域开创事业，拥有媒体资产与广泛人际网络。尽管存在未能早期加入AI实验室等遗憾，但三十岁阶段的整体成就远超二十岁。他通过每2-5年重塑身份保持神经可塑性，核心转变在于从被动接受社会既定路径，转向主动研究并定义属于自己的“游戏规则”，以追求真正的满足与持续学习。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月23日

CALLED IT "Meta is installing tracking software on employees’ computers to capture mouse movements, clicks and ​keystrokes to train its AI models, to build AI agents that work autonomously." "The tool will also take snapshots of the content on employees’ screens."

译预言中了 "Meta 正在员工的电脑上安装追踪软件，以捕获鼠标移动、点击和按键操作，用于训练其 AI 模型，从而构建可自主工作的 AI agent。" "该工具还将对员工屏幕上的内容进行截图。"

Ethan Mollick@emollick · 4月22日

All of the AI models have preferred names. If you asked Claude 4.5 for a software developer, you are going to get Marcus Chen. Wizards are mostly named Aldric. Space pilots are Kira from Claude, Mara Vance from GPT-5.2. I guess LinkedIn Bros are Kai now. https://www.seehuhn.de/blog/ai-names/

译所有 AI 模型都有偏好的名字。如果你向 Claude 4.5 要一个软件开发者，你会得到 Marcus Chen。巫师大多叫 Aldric。来自 Claude 的太空飞行员是 Kira，来自 GPT-5.2 的是 Mara Vance。 我猜 LinkedIn Bros 现在都是 Kai 了。https://www.seehuhn.de/blog/ai-names/

Rohan Paul@rohanpaul_ai · 4月22日

Robotics is about to reshape lives. At the Canton Fair, China, a woman with severe knee issues who could barely stand or walk tested a Chinese waist-assisted exoskeleton. Then she took steady steps without support. Her friend broke into tears.

译机器人技术即将重塑生活。 在中国 Canton Fair，一位膝盖有严重问题、几乎无法站立或行走的女性测试了一款中国腰部辅助外骨骼。 然后她在没有支撑的情况下迈出了稳健的步伐。 她的朋友泪流满面。

Rohan Paul@rohanpaul_ai · 4月22日

Toyota’s CUE7 humanoid appeared mid-game at Toyota Arena Tokyo. In front of a live crowd, it handled a pass, dribbled smoothly & executed free throws. Reinforcement learning lets it improve shooting accuracy through repeated trials & real-time feedback

译Toyota 的 CUE7 人形机器人在 Toyota Arena Tokyo 比赛期间中场亮相。在现场观众面前，它完成接球、流畅运球并执行罚球。 强化学习使其能够通过反复试验和实时反馈来提高投篮准确率。

Rohan Paul@rohanpaul_ai · 4月22日

During Beijing’s humanoid robot half-marathon, Galaxea’s R1 wheeled robot stood at a water station table. Runners approached, and the robot’s arms extended to pick up and pass plastic bottles

译在北京人形机器人半程马拉松期间，Galaxea 的 R1 轮式机器人站在水站桌子旁。跑者靠近时，机器人的手臂伸出，拿起并递出塑料瓶。

Ethan Mollick@emollick · 4月22日45

An advantage of the humanities — of reading & seeing a wide range of great works from many perspectives & cultures — is that you develop your sense of taste. In a world where anyone can produce a flood of writing & visual output is for cheap, that has never been more important.

译人文学科的一个优势——通过阅读和观察来自多元视角与文化的众多伟大作品——在于你能培养自己的品味感知力。 在这个任何人都能轻易制造大量文字和视觉产出的世界里，这种能力的重要性前所未有。

Yuchen Jin@Yuchenj_UW · 4月22日37

Anthropic comes for Figma. OpenAI comes for Canva. xAI comes for Cursor.

译Anthropic 进军 Figma。 OpenAI 进军 Canva。 xAI 进军 Cursor。

SemiAnalysis@SemiAnalysis_ · 4月22日

At OFC 2026 last month,  Cisco's chief architect Rakesh Chopra presented on scale-across networking architectures and key deployment trends driving strong demand for traditional DCI equipment. "Traditional" DCI connects CPUs across the frontend network while scale-across connects GPUs over the back-end to enable loss-intolerant, synchronous data flows. In scale-across networking, hyperscalers manage oversubscription of intra-datacenter bandwidth relative to inter-datacenter bandwidth using deep switch buffers and proactive congestion control. The bandwidth needs of scale-across is approximately 14x the bandwidth needs of traditional DCI. As such, significant buildout of scale-across infrastructure at various hyperscalers is expected to result in multi-billion dollar opportunities for 800G coherent pluggables, deep buffered switches and the like. SemiAnalysis's AI Networking Model will initiate estimates of scale across networking equipment spend at various hyperscalers, coming soon.

译Cisco首席架构师在OFC 2026提出scale-across网络架构，与传统DCI连接CPU的前端网络不同，scale-across通过后端网络连接GPU，支持无损同步数据流。超大规模数据中心采用深度缓冲交换机和主动拥塞控制管理带宽超配，其带宽需求约为传统DCI的14倍。这将带动800G相干可插拔光模块、深度缓冲交换机等数十亿美元市场机会，SemiAnalysis即将发布相关支出预测模型。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月21日

Dead Internet Theory update: AI song uploads have nearly overtaken human music RECAP: 1) The majority of articles on the internet are written by AIs... 2) 4 of the top 10 Youtube channels... 3) 4 in 10 Facebook posts... 4) 1 in 5 videos shown to new Youtube users... 5) The #1 most-subscribed Twitch streamer is an AI... 6) 44% of songs on Deezer... Also, anecdotally, I tried a new Youtube account and most of the ads were AI generated. Fake humans fake endorsing products for gullible boomers and grandmas

译死互联网理论正在成为现实。数据显示，AI内容已全面渗透：Deezer上44%歌曲为AI上传，Twitch订阅第一主播也是AI；多数网络文章、四成Facebook帖子及Top 10 YouTube频道内容均由AI生成。连广告领域也充斥AI生成的虚假代言。AI正系统性替代人类创作，从音乐到视频全面入侵。

SemiAnalysis@SemiAnalysis_ · 4月20日55

How Much Do GPU Clusters Really Cost? Calculating Cluster TCO, The Real Impact of Downtime, The Grand Unifying Theory Of Goodput, and a ClusterMAX 2.1 Update READ NOW: https://newsletter.semianalysis.com/p/how-much-do-gpu-clusters-really-cost?_gl=1*1uithfa*_ga*MTY1NDExMjk2Ny4xNzc2MTIzOTQ1*_ga_FKWNM9FBZ3*czE3NzY2OTU2ODAkbzEyJGcwJHQxNzc2Njk1NjgwJGo2MCRsMCRoMTAyODIzNDQ0OA..

译GPU集群的真实成本究竟是多少？ 计算集群总拥有成本， 停机时间的真实影响， 有效吞吐量的宏大统一理论， 以及ClusterMAX 2.1更新 立即阅读：https://newsletter.semianalysis.com/p/how-much-do-gpu-clusters-really-cost?_gl=1*1uithfa*_ga*MTY1NDExMjk2Ny4xNzc2MTIzOTQ1*_ga_FKWNM9FBZ3*czE3NzY2OTU2ODAkbzEyJGcwJHQxNzc2Njk1NjgwJGo2MCRsMCRoMTAyODIzNDQ0OA..

Chubby♨️@kimmonismus · 4月20日

OpenAI's weekly active users have not increased since February. The goal remains to reach one billion. However, as far as I can see, the trend is currently moving in the opposite direction. Spud must be a hit.

译OpenAI 的周活跃用户自二月以来没有增长。目标仍是达到十亿。然而，据我所见，目前的趋势正朝着相反方向发展。 Spud 一定很火。

DogeDesigner@cb_doge · 4月19日24

"Things will just be free in the future. Sounds nuts, but if you've got an AI or robotics economy that is anywhere close to million times the size of the current Earth economy, literally any need you possibly want can be met. If you can think of it, you can have it" 一 Elon Musk

译未来一切都会免费。听起来很疯狂，但如果你拥有一个规模接近当前地球经济百万倍的AI或机器人经济，那么从字面上讲，任何你可能想要的需求都能得到满足。只要你能想到，就能拥有。 一 Elon Musk

Chubby♨️@kimmonismus · 4月19日

im speechless. GPT-5.5 created the best SVG i've seen so far. One shot. We are in for a wild ride.

译无语了。GPT-5.5 创造了我目前见过最好的 SVG。一次生成。接下来会很疯狂。

Chubby♨️@kimmonismus · 4月19日

This Humanoid Robot just finished a Half-Marathon in 50 minutes and 26 seconds. The point is not so much that robots are now running marathons, but that it has become a national competition. It's about the culture, the spirit, that humanoid robots can be introduced into everyday life and become an event. Robots and AI should become ubiquitous and enthusiasm for their development should be encouraged. That's why such events are organized. And this spirit is missing in Europe, for example.

译这款人形机器人刚刚以50分26秒完成了一场半程马拉松。 重点不在于机器人现在能跑马拉松，而在于这已成为一项全国性竞赛。这关乎文化，关乎精神，即人形机器人可以被引入日常生活并成为一项赛事。 机器人和AI应当变得无处不在，对其发展的热情应当受到鼓励。这正是举办此类活动的原因。而这种精神，例如在欧洲，是缺失的。

Rohan Paul@rohanpaul_ai · 4月19日

BBC Published an article. AI chatbots are becoming a real front door for health advice, but new evidence says human-AI conversation breaks their medical accuracy far more than most people realize. The problem is not that these systems always fail when they see a full, neatly written case, because in controlled testing they reached about 95% accuracy. The problem is that real people give messy, partial, distracted symptom descriptions, and in that setting accuracy dropped to about 35% In the area of medical advice, a tiny wording change can flip advice from “rest at home” to “go to hospital now, --- bbc .com/news/articles/clyepyy82kxo

译AI聊天机器人正成为健康咨询的重要入口，但新证据表明，真实人机对话对医疗准确性的破坏远超预期。研究显示，在控制测试中AI准确率可达95%，但面对真实用户混乱、不完整、分心的症状描述时，准确率骤降至35%。医疗建议领域存在极高敏感性，细微的措辞变化可能导致建议从"居家休息"翻转为"立即就医"，凸显当前AI医疗应用在实际场景中的重大风险。

Rohan Paul@rohanpaul_ai · 4月19日

Humanoid acceleration has started. The inevitable is not asking for permission.

译人形机器人加速已经开始。 必然之事无需许可。

Rohan Paul@rohanpaul_ai · 4月19日

During halftime of a pro basketball game in Tokyo, Toyota's CUE7 robot rolled onto the court, stood from a seated position, dribbled the ball smoothly, and successfully did free throws in front of 8,400 fans. At 220 cm tall and 74 kg with a wheeled base

译在东京一场职业篮球赛的中场休息时，Toyota 的 CUE7 机器人滚动进入球场，从坐姿站起，流畅地运球，并在 8400 名观众面前成功完成罚球。 身高 220 厘米，重 74 公斤，带有轮式底座

Chubby♨️@kimmonismus · 4月19日

German angst No2 Germany's biggest magazine has gone even further. After opening yesterday's lead article claiming that AI would make us all dumber, today's solution: simply 90 minutes of intensive human instruction. Simply make AI unattractive. I'm speechless.

译德国焦虑 No2 德国最大的杂志更进一步。昨天头版文章声称AI会让我们变蠢之后，今天的解决方案：只需90分钟强化人工教学。干脆让AI变得不受欢迎。我无语了。

DogeDesigner@cb_doge · 4月19日30

"The money will stop being relevant at some point in the future, there will still be constraints on power like electricity and mass, the fundamental physics elements will still be constraints, but I think at some point currency becomes irrelevant." 一 Elon Musk

译金钱在未来的某个时刻将不再重要，电力、质量等能源限制依然存在，基础物理要素仍将是制约因素，但我认为货币终将在某个时刻变得无关紧要。 一 Elon Musk

Ethan Mollick@emollick · 4月19日

What I find very funny about these “leaks” is that they don’t even bother to get ballpark benchmarks to feed into the image generators. Ask the model to look up real data, at least. Its easy! Like GPQA is over 90% for all recent models.

译我觉得这些"泄露"很好笑的是，他们甚至懒得弄个大体准确的基准测试数据来输入到图像生成器里。至少让模型查一下真实数据吧。这很简单！ 比如 GPQA 在所有近期模型上都超过 90% 了。

Rohan Paul@rohanpaul_ai · 4月18日

Wny the cooldown for OpenClaw 🤔 Search interest for “OpenClaw” on Google has dropped off to near baseline.

译为什么 OpenClaw 的热度降了 🤔 Google 上“OpenClaw”的搜索兴趣已降至接近基线水平。

Chubby♨️@kimmonismus · 4月18日

Even inflation-adjusted, annual global datacenter CapEx today is roughly equivalent to 5–7 Manhattan Projects per year (≈$250–300B vs. ≈$25–30B in today’s dollars for the Manhattan Project).

译即使经过通胀调整，如今全球年度数据中心资本支出大致相当于每年 5–7 个 Manhattan Project（约 2500–3000 亿美元，而 Manhattan Project 按今日美元计算约为 250–300 亿美元）。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月18日43

AI datacenter spending has surpassed the Manhattan Project, Marshall Plan, International Space Station, and the Apollo Program ***COMBINED***

译AI数据中心支出已超过曼哈顿计划、马歇尔计划、国际空间站和阿波罗计划 ***的总和*** [引用 @finmoorhouse]：超大规模企业支出已超过美国最著名的巨型项目

Yuchen Jin@Yuchenj_UW · 4月17日61

Figma stock 20 minutes after the Claude Design announcement. Wild.

译Figma 股价在 Claude Design 发布 20 分钟后。 疯狂。

SemiAnalysis@SemiAnalysis_ · 4月17日

Jensen surfing a wave with the American flag is the most accurate metaphor for what's actually happening right now. A kid immigrates from the country of Taiwan, washes dishes at Denny's, and builds a $3 trillion company that every nation on earth is now begging for access to. He is the definition of the American dream.  In order for America to strive, Earth needs to be build on American standards

译Jensen 举着美国国旗冲浪，这是对当下正在发生的事情最准确的隐喻。一个来自 Taiwan 的孩子移民至此，在 Denny's 洗盘子，创建了一家价值3万亿美元的公司，如今地球上的每个国家都在乞求获得它的使用权。他就是美国梦的定义。为了让美国奋斗，地球需要按照美国标准来建设。

Chubby♨️@kimmonismus · 4月17日

my whole fy page is people ranting about opus 4.7 anthropic messed up big time

译我的整个 fy 页面都是人们在吐槽 opus 4.7 anthropic 这次搞砸了

DogeDesigner@cb_doge · 4月17日34

"If A.I. can do everything that you can do but better, then what is the point of doing things. The most likely outcome is one of abundance where goods & services are available to anyone. It wouldn't be universal basic income but it would be universal high income." 一 Elon Musk

译如果人工智能能做所有你能做的事，而且做得更好，那么做事的目的是什么。最可能的结果是丰裕时代，商品和服务对任何人都触手可及。这不会是全民基本收入，而会是全民高收入。 一 Elon Musk

DogeDesigner@cb_doge · 4月17日

Tesla FSD saves a kid’s life. ❤️ A friend of mine in Arizona was on his way home in a Cybertruck using FSD when a kid suddenly ran across the road. The car reacted instantly and braked safely, avoiding what could’ve been a tragedy.

译Tesla FSD 救了一个孩子的命。❤️ 我的一位朋友在亚利桑那州，当时正开着 Cybertruck 用 FSD 回家，突然有个孩子跑过马路。车辆瞬间反应并安全刹车，避免了一场可能发生的悲剧。

Peter Steinberger 🦞@steipete · 4月17日

they: OpenClaw is so insecure look at all these GHSAs! reality: we are just an indicator of the coming storm

译他们：OpenClaw 太不安全了，看看这些 GHSA！ 现实是：我们只是暴风雨来临的指示器 [引用 @samsaffron]：13 年后，我们绝不会关闭 @discourse 的源代码。相反，我们在安全方面大力投入，并适应时代。上个月的发布版本有 50 个 CVE，这得益于使用 GPT 5.4 xhigh 进行的多日扫描。https://x.com/pumfleet/status/2044406553508274554

宝玉@dotey · 4月17日

要想编程效果好，就得学会“黑话”😂

Epoch AI@EpochAIResearch · 4月17日

According to our latest polls, Claude usage in the US rose by over 40% amid increased attention last month, but remains far behind ChatGPT. Our point estimate would imply several million new weekly users in the United States.

译根据我们最新的调查，Claude 在美国的使用量在上月关注度上升期间增长了超过 40%，但仍远落后于 ChatGPT。 我们的点估计意味着在美国每周有数百万新用户。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月17日

3 months.

译3个月。 [引用 @arankomatsuzaki]：Anthropic 近1/3的受访人员现在认为初级工程师和研究人员可能在3个月内被 Mythos 取代

宝玉@dotey · 4月17日

这封面不错😂