Chubby♨️@kimmonismus · 4月12日

From my own experience, I can say that I'm using Claude more and more. It's simply the taste; the less "chatty" nature and the more concise answers make it more appealing.

译越来越多使用 Claude，因其回答简洁、不"话痨"的风格更讨喜。Similarweb 数据显示，Claude AI 第一季度网站流量同比增长 341.26%。

Nathan Lambert@natolambert · 4月12日

Seems like 30-200B open models are getting a massive surge in usage at least partially related to openclaw, but it's very hard to attribute back to. How could we make this easier for the ecosystem to measure?

译30-200B 参数规模的开源模型近期使用量激增，部分与 OpenClaw 相关，但难以精确归因。作者询问如何建立更清晰的生态系统测量机制来追踪此类影响。

Chubby♨️@kimmonismus · 4月12日

Yeah Claude. I know that feel. Wow.

译对 Claude 的某种表现产生强烈共鸣，以「我懂那种感觉」简短回应，配以「Wow」表达惊讶与认同，展现对 AI 反应的感性共鸣。

Nathan Lambert@natolambert · 4月12日

a bit over 7 days out from the Gemma 4 release and it's models are outpacing (slightly) the equivalent Qwen 3.5 models on downloads. Big numbers!

译Gemma 4 发布刚满 7 天，各尺寸模型下载量已小幅超越同等级 Qwen 3.5，数据表现亮眼。

Rohan Paul@rohanpaul_ai · 4月12日

Unitree’s H1 humanoid reached 10 m/s. That is world-champion territory, Usain Bolt's peak was 12.42 m/s, with his 100 m average around 10.44 m/s. Unitree was 3.3 m/s just 2 years ago. significant advances in balance & gait control at high velocities

译Unitree H1 人形机器人奔跑速度达到 10 m/s，接近博尔特百米平均速度（10.44 m/s）。较两年前的 3.3 m/s 大幅提升，标志着高速平衡与步态控制技术取得关键突破。

Rohan Paul@rohanpaul_ai · 4月12日

David Sacks: AI first drove value to chips, then hyperscalers, then model companies like OpenAI. Now the real question is if next gains go to application firms or get absorbed by model makers. Palantir shows what AI can do for app companies.

译David Sacks 称 AI 价值已从芯片、超大规模云服务商转移至 OpenAI 等模型公司，下一阶段看应用公司（如 Palantir）能否捕获价值。Brad Gerstner 指出，自有算力使成本固定，OpenAI 算力利润率从 35% 升至 70%，Anthropic 从 -94% 转正至 40%，物理电力成主要瓶颈。

DogeDesigner@cb_doge · 4月12日

How Grok saved this cat’s life ❤️ • A photographer @Lars_Maier_de in Frankfurt noticed his 9-year-old cat suddenly stop eating late at night • The cat hid under the couch, had a strange sweet/rotten breath, then suddenly collapsed unconscious • Panicked and unsure what was happening, he opened Grok and typed the symptoms • Grok responded instantly, warning it could be a serious diabetic emergency (ketoacidosis) • Grok told him clearly: don’t wait, go to a vet immediately • He rushed to a 24-hour clinic with the cat unconscious beside him • Kept checking Grok on the way, which explained the condition and how critical every minute was • This helped him stay calm enough to drive fast but safely • At the clinic, vets confirmed it was a full diabetic crisis • They said if he had arrived even 30 minutes later, the cat likely wouldn’t have survived • Because of the early warning and urgency, he made it just in time Now the cat is back home, and recovering.

译法兰克福摄影师的9岁猫深夜突发昏迷并出现甜腐味呼吸，主人通过Grok查询症状。Grok即时识别为糖尿病酮症酸中毒（ketoacidosis），明确警告需立即就医。主人驱车赶往24小时诊所，途中Grok持续说明病情危急程度。兽医确诊糖尿病危象，称晚到30分钟猫将无法存活。因Grok及时预警，猫咪获救并康复。

Chubby♨️@kimmonismus · 4月11日

That's why it's so important that AI is properly adapted to the working world and that we build a new post-laboratory economy. Gen Z workers, driven by fear of job loss, are actively sabotaging company AI rollouts, ironically making themselves more likely to be replaced while AI “power users” get promoted and rewarded.

译Z世代员工因担心失业而主动破坏公司AI推广，结果适得其反——这种行为反而增加了他们被替代的风险，而善于使用AI的"高级用户"则获得晋升和奖励。

DogeDesigner@cb_doge · 4月11日

BREAKING: Grok saved another life. This time, a cat from Frankfurt. ♥️ A Frankfurt Photographer, Lars Maier credits Grok for saving his cat, Moritz’s life during an overnight diabetic ketoacidosis crisis. Grok instantly recognized the symptoms and told him: “RUSH to the 24h vet NOW!”— vets say 30 mins later would’ve been fatal. Moritz is home, eating, and recovering.

译突发：Grok 又救了一条命。这次是一只来自法兰克福的猫。♥️

Chubby♨️@kimmonismus · 4月11日

Regardless of whether Mythos lives up to the hype, they've mastered PR. OpenAI would have liked to see the attention Claude Mythos is getting for "spud," especially with the upcoming IPO.

译Claude Mythos 凭借 "spud" 功能获得极高关注度，公关策略堪称典范。OpenAI 在即将 IPO 之际想必很羡慕这种舆论热度，无论 Mythos 最终是否名副其实。

SemiAnalysis@SemiAnalysis_ · 4月11日

The intern is running 8 windows of Claude simultaneously and has become the boss of the agent swarm. SemiAnalysis built something that token mogs Meta and episode 8 is out now with @jordannanos @dnishball and @sharshe02.

译实习生同时运行着8个Claude窗口，已成为智能体集群的老大。SemiAnalysis构建了一个在token方面碾压Meta的东西，第8集现已发布，嘉宾是@jordannanos @dnishball和@sharshe02。

SemiAnalysis@SemiAnalysis_ · 4月11日

Jensen showing Rubin Ultra as an MCM was the real tell. This is not just Nvidia gluing more dies together because it feels like the next cool architecture move. It is what happens when reticle limits, power density, yield, and package economics all start forcing the same answer. (1/5)🧵

译Jensen 将 Rubin Ultra 展示为 MCM 才是真正的信号。这不仅仅是 Nvidia 因为觉得这是下一个很酷的架构动作而把更多芯片粘在一起。而是当光罩限制、功率密度、良率和封装经济性都开始迫使得出相同答案时发生的事情。(1/5)🧵

Ethan Mollick@emollick · 4月11日

Our Lab just posted a new research report from Zimran Ahmed about how the game industry is adapting to AI. He spoke to people at 20 different studios and found a wide range of approaches to adapt (or failures to adapt) to AI at the organizational level. https://gail.wharton.upenn.edu/research-and-insights/beyond-copy-paste/

译宾夕法尼亚大学沃顿商学院生成式AI实验室（GAIL）发布Zimran Ahmed撰写的研究报告《Beyond Copy-Paste》。研究通过对20家游戏工作室的深入访谈发现，游戏行业在组织层面适应AI技术呈现显著分化：部分工作室积极整合AI工具优化流程，亦有机构未能有效应对。报告揭示了游戏业在AI转型过程中的多样化实践路径与组织挑战，为理解技术变革中的行业适应机制提供了实证依据。

Ethan Mollick@emollick · 4月11日

AI finally lets us see Raphael's The School of Athens the way Raphael obviously intended it, illustrating the delicate dance and subtle conflicts between Plato and Artistotle. (Seedance 2.0 is very fun to play with)

译Seedance 2.0 用 AI 技术重新诠释拉斐尔名作《雅典学院》，呈现柏拉图与亚里士多德之间的微妙冲突与思想张力。生成效果有趣，可玩性高。

Deedy@deedydas · 4月11日

Oh, how the tables have turned. OpenAI / ChatGPT is now running sponsored search ads on Claude keywords.

译OpenAI 与 ChatGPT 开始在搜索引擎针对 Claude 关键词投放赞助广告。昔日被拿来与 Claude 比较的 ChatGPT，如今也要购买竞品关键词广告位，形势彻底逆转。

Epoch AI@EpochAIResearch · 4月10日

The Iran War and Hormuz shutdown have disrupted oil, gas, and helium exports and threatened data centers and investments in the Gulf states. @justjoshinyou13 explores how a prolonged Iran war could affect AI, and why it probably won’t completely derail the compute buildout.

译伊朗战争及霍尔木兹海峡关闭已扰乱油气与氦气出口，威胁海湾地区数据中心与投资。分析指出，长期冲突虽将影响AI供应链，但不太可能彻底阻碍全球计算能力扩张进程。

Yuchen Jin@Yuchenj_UW · 4月10日

Claude Mythos refused to send my tax return to the IRS. Said it was “too dangerous and terrifying.”

译Claude Mythos 以"太危险且可怕"为由，拒绝代用户向 IRS 提交税表。网友借机吐槽：Anthropic 能"杀死"各种功能，为何不能干掉 TurboTax。

Ethan Mollick@emollick · 4月10日

One fun thing about AI is that it lets you play with interfaces and approaches to displaying information in new ways without a lot of effort. I got a an internet connected e-ink display and set it up to show me the weather as interpreted by nano banana using rotating styles.

译入手联网电子墨水屏，接入 nano banana 以轮换风格实时展示天气。AI 降低了尝试新型界面和数据可视化的门槛，无需复杂开发即可实现个性化信息展示。

Yuchen Jin@Yuchenj_UW · 4月10日

My convo with a startup founder: me: “How many Claude tokens do your employees burn per day?” him: “~$2000/day per person.” me: “Wow, that’s $730k/year per employee.” me: “If Claude Mythos costs 5x more, that’s $3.65M/year per employee.” him: “Take my money.” Future companies may pay more to agents than to humans.

译某创业公司员工日均消耗 $2000 的 Claude tokens，折合年薪 $73 万。即便未来 Claude Mythos 涨价 5 倍至 $365 万/年，创始人仍表示"拿走我的钱"。未来企业付给 AI agent 的费用可能超过人类薪资。

Nathan Lambert@natolambert · 4月9日

Directionally, I agree with this piece, but it's important to note that this is the first blip in a long, slow transition towards a more hybrid model of open and closed models. And in that, China still may be producing way more open models than the US, enough to cement the ecosystem as largely Chinese AI. I expect this evolution to take years, and the cultural default of open is still set, which may never fully decay. https://www.chinatalk.media/p/chinas-ai-companies-are-going-closed

译回应中国AI公司转向闭源的观点，指出这只是向开闭源混合模式长期过渡的初期信号。中国仍可能产出比美国更多的开源模型，且开源文化底色难以消退，这一演变预计将持续数年。

SemiAnalysis@SemiAnalysis_ · 4月9日

YOUR PARENTS PAID FOR THE CUDA MOAT! The #1 contributor to the CUDA MOAT isn't the the developers at NVIDIA, but it is the millions of developers outside of NVIDIA that invent new algorithms for CUDA like Flash Attention. For most of them, it started with an GeForce gaming GPU. NVIDIA is the only companies that has an reasonable good developer stack on consumer grade GPUs. As people grow up beyond playing CSGO & League of Legends & Minecraft, they either become anime weeaboos or they start programming on their existing computer with has an GeForce GPU

译CUDA生态的护城河并非主要由NVIDIA内部开发者构建，而是源于数百万外部开发者——他们基于CUDA发明了Flash Attention等算法。这些开发者大多从GeForce游戏GPU起步，因为NVIDIA是唯一在消费级GPU上提供完善开发者工具栈的公司。游戏玩家长大后，利用现有的GeForce显卡转向编程，形成了从游戏生态到AI开发的独特人才输送管道。

Ethan Mollick@emollick · 4月9日

Hallucinations remain in LLMs, but note that over centuries we have developed complicated, successful machines that take uncertain output from unreliable sources & reduce the risk of errors. We call those machines organizational structures & we can apply similar approaches to AI

译LLM 幻觉虽无法避免，但人类早已发展出应对之策。通过建立组织结构来处理不可靠信息并降低错误风险，这种数百年验证有效的方法同样适用于 AI。

Epoch AI@EpochAIResearch · 4月9日

New essay by @ansonwhho: Chinese and open model AI labs have ≈10× less compute than the frontier. But they can distill frontier models, replicate innovations fast, and have enormous talent. Is that enough to compete at the frontier? 🧵

译中国及开源 AI 实验室算力约为前沿的 1/10，但具备蒸馏前沿模型、快速复制技术创新及庞大人才储备等优势。@ansonwhho 探讨这些条件能否弥补算力差距，支撑其在最前沿 AI 领域保持竞争力。

Nathan Lambert@natolambert · 4月8日

New report with @xeophon is out with the latest open model adoption data we have gathered for Interconnects & The ATOM Project. At the surface level, we can see Chinese models continuing to accelerate in adoption. The report details much more. 1. We manually curate ~1.5K of the most important language models, creating a specific set of models to focus our analysis on (excludes embedding models, local inference models like MLX/GGUF, etc to have accurate download rankings). 2. Studying other adoption metrics, such as derivative models and inference share on OpenRouter, to show how they correlate with downloads, while often sifted in time. China has a strong lead here too. 3. Better classification of downloads across model sizes. Large models still are the models where Qwen is least competitive, relative to other model builders. 4. Expansion of our Relative Adoption Metric (RAM) to show standout recent models (we'll check Gemma 4 on Friday); Qwen 3.5, Nemontron 3, Kimi K2.5, all showing very strong adoption. Overall, this is another step towards formalizing and making public better data on the open language model ecosystem, so the community can better understand the impact and trends of its adoption. More on this soon!

译本报告基于Interconnects与ATOM Project数据，手动筛选约1.5K个重要语言模型，通过下载量、衍生模型数量及OpenRouter推理份额等多维度指标，分析开源模型采用趋势。数据显示，以Qwen、Kimi为代表的中国模型全球采用率持续加速领先，其中Qwen 3.5、Nemontron 3、Kimi K2.5等近期模型在相对采用指标(RAM)中表现突出。研究同时指出，大型模型仍是Qwen相对竞争力较弱的领域。该工作旨在为开源生态系统提供更准确的公开数据与趋势洞察。

Peter Steinberger 🦞@steipete · 4月8日

glad they banned openclaw, the servers are finally reliable again

译OpenClaw 遭封禁后服务器可靠性终于恢复，oncall 团队无需再紧急救火。该工具此前疑似导致系统频繁故障，让值班人员苦不堪言。

Haider.@haider1 · 4月8日39

i still can't get over this look at those benchmark results: > swe-bench verified: mythos 93.9% vs opus 4.6 80.8% > swe-bench pro: mythos 77.8% vs opus 4.6 53.4% > swe-bench multilingual: mythos 87.3% vs opus 4.6 77.8% > swe-bench multimodal: mythos 59.0% vs opus 4.6 27.1% > terminal-bench 2.0: mythos 82.0% vs opus 4.6 65.4%

译我仍然无法释怀 看看这些基准测试结果： > swe-bench 已验证：mythos 93.9% vs opus 4.6 80.8% > swe-bench 专业版：mythos 77.8% vs opus 4.6 53.4% > swe-bench 多语言版：mythos 87.3% vs opus 4.6 77.8% > swe-bench 多模态版：mythos 59.0% vs opus 4.6 27.1% > terminal-bench 2.0：mythos 82.0% vs opus 4.6 65.4%

Epoch AI@EpochAIResearch · 4月8日

Who owns the world's compute? Our new Chip Ownership hub shows that Google leads, holding around 25% of all compute sold since 2022.

译Chip Ownership 最新数据显示，Google 占据2022年以来全球销售算力约25%的份额，领先市场。

Nathan Lambert@natolambert · 4月7日

AI generally: not a bubble. OpenAI specifically: having to pay for 1B un-monetizable, high marginal costs users, need to raise so much -- maybe a bubble.

译AI 行业整体并非泡沫，但 OpenAI 处境特殊：需承担 10 亿难以变现且边际成本极高的用户，巨额融资需求下或许存在泡沫风险。

Yuchen Jin@Yuchenj_UW · 4月7日

I’m pretty sure the $20/$200 subscription pricing was vibe-coded by OpenAI, then copied by Anthropic. That pricing works for chatbots, not agents. A 24/7 agent can burn through orders of magnitude more tokens than a user chatting with a chatbot. Now they’re stuck. Neither Anthropic nor OpenAI wants to be the first to change pricing and risk user churn, so the options are: keep subsidizing, get more GPUs, tighter rate limits, and enforce rules like limiting 3rd-party apps. I wouldn’t be surprised if intelligence gets more expensive, not cheaper.

译$20/$200订阅定价由OpenAI设定并被Anthropic复制，适用于Chatbot却不适用于Agent。24/7 Agent的token消耗远超聊天场景。OpenAI与Anthropic陷入囚徒困境，无人愿率先调价以免用户流失，只能继续补贴、扩充GPU或限制第三方应用。作者预测，随着Agent普及，智能服务将变得更贵而非更便宜。

SemiAnalysis@SemiAnalysis_ · 4月7日

NVIDIA ARCHITECTURE ALERT🚨 Shared memory increased almost every generation, while register file size stayed constant. The reason for this is that Tensor Core throughput increase requires a deeper staging buffer. Because Tensor Cores consume data much faster than global memory can load, we use a staging memory to buffer data, so memory loading can run ahead of MMA operations. Tensor Core throughput doubled every generation, but global memory load latency didn’t decrease and in fact increased. As a result, we need to increase the staging memory size for buffering more data. To implement this, NVIDIA chose shared memory as the staging memory for Tensor Cores, which explains why shared memory increased but register file size remained constant. However, Blackwell’s shared memory size didn’t increase from Hopper. This is because tcgen05 MMA can leverage 2 SMs, so each SM’s shared memory only needs to load half of the operands. Thus, Blackwell’s shared memory size effectively doubled.

译NVIDIA GPU中Shared memory逐代递增而寄存器文件不变，主因是Tensor Core吞吐量翻倍需更大缓冲池。由于全局内存加载速度远不及Tensor Core处理速度且延迟攀升，NVIDIA将Shared memory用作Tensor Core的暂存区。Blackwell虽未提升单SM的Shared memory容量，但借助tcgen05 MMA双SM协同设计，每个SM仅需加载半数操作数，实现等效容量翻倍。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月6日

"The truth is, we're building portals from which we're genuinely summoning aliens." -Former OpenAI exec "The portals currently exist in the US, China, and Sam has added one in the Middle East." "It's wildly important to get how scary that should be."

译前 OpenAI 高管爆料，称 AI 是"召唤外星人的传送门"，已布局美国、中国和中东。OpenAI 被曝试图让中美俄竞价，斥巨资游说国会反对监管，打着美国旗号实则只为自身利益。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月6日

In 2017, Altman straight up lied to US officials that China had launched an "AGI Manhattan Project" He claimed he needed billions in government funding to keep pace. An intelligence official investigated, found no evidence. "It was just being used as a sales pitch."

译《纽约客》调查披露，2017 年 Altman 为争取政府资金，向美方官员谎称中国已启动"AGI 曼哈顿计划"。情报官员调查后未发现任何证据，证实该说法只是用于推销数十亿美元拨款的谎言。

Deedy@deedydas · 4月5日

How to keep up with AI (and everything else) 1. Twitter/X is the fastest source. LinkedIn/Reddit are way behind 2. Smaller accs (<50k followers) have high alpha 3. X feed is very noisy though 4. Add 10-20 accounts to "notify for all posts" 5. Aggregated notif is the curated feed

译Twitter/X 是获取 AI 资讯最快的渠道，LinkedIn 和 Reddit 明显滞后；粉丝少于 5 万的小账号往往包含高价值信息。建议将 10-20 个优质账号设为"全部通知"，通过聚合通知功能打造个人精选信息流，避免被主时间线的噪音淹没。

Tibo@thsottiaux · 4月4日

Always fun when you notice Codex being clever in a way you don't expect. In a session today, it was running a slow build process and got annoyed (don't we all). Before making a change it checked that progress was actually happening and did so not by checking the logs, but by checking CPU usage.

译Codex 在运行缓慢构建时表现出类似人类的"不耐烦"，未选择查看日志，而是通过监测 CPU 使用率来确认进度确实在推进，展现了意料之外的智能判断方式。

swyx 🇬🇧@swyx · 4月4日

We have achieved agentic self improvement - i can just copy paste blogposts and tweets into @devinai and it oneshots the complete implementation wasnt actually sure this was gonna work, jaw dropped when it did. this is very out of distribution of the underlying @GoogleDeepMind Gemini Flash Lite model but it Just Worked.

译将博客或推文直接粘贴至 @devinai，即可一次性生成完整代码实现。底层 Gemini Flash Lite 模型虽超出训练分布，但效果惊人，实现智能体自我改进。

OpenAI Developers@OpenAIDevs · 4月3日

When your voice agent debugs your slides live @charlierguo is using gpt-realtime-1.5

译@charlierguo 使用 gpt-realtime-1.5 进行实时演示，语音助手现场调试幻灯片内容，展示该模型在实时语音交互与视觉理解方面的应用能力。

Nathan Lambert@natolambert · 4月2日

Nemotron Super / Ultra Arcee Trinity Large (soon) Gemma 4 (eventually) Reflection's first models (maybe) GPT OSS 2? (maybe) Thinky? Other neolabs? Things looking up for open models built in the US in 2026. We had 0 for a bit there.

译Nemotron Super/Ultra、Arcee Trinity Large、Gemma 4 及 Reflection 首个模型都将在 2026 年发布，GPT OSS 2 和 Thinky 等也可能加入。美国开源模型此前一度挂零，如今终于迎来爆发期。

Deedy@deedydas · 4月2日

Blows my mind that we currently possess the technology for Google Maps to turn all the street view images of the entire world into a video game you can play! In the future, we'll be able to say "yeah let's check out New York City 100 years ago!"

译Google Maps 现有技术已能将全球街景图像转化为可玩视频游戏，令人震撼。未来还能借此回顾100年前的纽约等城市风貌，实现穿越时空的探索体验。

karminski-牙医@karminski3 · 4月1日

这点我同意宝玉老师, 微服务或者各种拆分本身是为了解决开发者水平参差不齐的问题, 现在人的问题已经不存在了...... mono-repo 是最终归宿

Deedy@deedydas · 4月1日

I’ve been obsessed with the most exciting software tech today that’s not AI: Gaussian splats. It’s the next generation of videos where you can move around in the scene. And the whole thing renders in realtime on your iPhone. I went into a pretty deep rabbit hole on it.. so here’s some history. The initial idea was: can we take pictures from different angles and reconstruct a 3D scene? Fun fact: one of the seminal papers in the field (“Photo Tourism”) was written by a professor I taught graphics for in college, Noah Snavely! Problem: objects look different at diff angles, because of light etc Then we had NeRFs which could figure out lighting. Problem: extremely slow. Gaussian splatting represented a 3D scene with diffuse blobs (gaussians) that encoded structure and appearance. Now, you could take camera shots or drone shots and make a splat in <5s. Problem: a) still needed many images b) splats were static and didn’t have video in them c) unseen parts of video or holes are just black or missing Still need many images? Apple’s ML SHARP can take one image and give you a splat! Can't have video? Companies like 4DV ai who made the video below build special capture techniques which allow dynamic scenes to be put in a splat Parts of the video just black? Generative models (a subset of world models) can fill in the missing parts not captured by camera. What does that leave us with? The future entertainment format whether it's in VR on a Vision Pro or interacting with immersive video are going to use splats. There's still open problrms: a) how do we create splats more efficiently b) how do we store and stream them more efficiently c) how do we make them visually more realistic (lighting d) instead of being a flying camera, can we move like a video game character in the space and interact with objects Splats are closely related world models and virtual reality. Cool projects like Seoul World Model take street view images and let you fly through any part of the city. It's only a matter of time before the entire world gets a 3D representation we can move through baked straight into Google Maps. Or you can play control a video game character watching a live sports game.

译Gaussian splats是新兴的实时3D渲染技术，可在iPhone上实现自由视角的沉浸式场景浏览。该技术用高斯分布编码场景结构与外观，相比NeRFs极大提升渲染速度。当前突破包括单图生成（Apple ML SHARP）、动态场景捕捉（4DV ai）及生成模型填补未拍摄区域。未来将成为Vision Pro等VR设备的核心娱乐格式，并与世界模型结合实现城市级漫游或游戏化交互，但仍需解决创建效率、存储传输及视觉真实感等挑战。