vikingmute 吐槽用 Opus 24 小时复刻游戏 demo 的风潮,拿杀戮尖塔和 Factorio 举例说明玩法设计才是好游戏的灵魂。
Latest
现代、LG 这些韩国大厂都在接 NVIDIA 的企业智能体方案,LG 还开放了自家模型的 API,想看企业落地案例的可以看看。
Hugging Face 把 MoE 训练的门道整理成了一份手册,从路由到专家分配都讲了,想自己训 MoE 模型的可以翻翻。
clone 下来一堆 npm install 却跑不起来的朋友可以试试,它会告诉你缺哪个 Node 版本、数据库怎么配,按你的系统出一份能直接敲的命令。
一位爸爸用 Claude Code 给儿子复刻收费 App,半小时就能用,还开了打字游戏源码,想给娃做工具的可以照着抄作业。
The practical shift here is that supervising agents demands the same skills as managing engineers — scoping, review and verification — rather than writing code yourself. Developers running long agent sessions can apply the agents.md persistence trick and automated verification loops immediately. Anyone frustrated by agents repeating fixed errors will find the lowest-friction fix in the third practice.
Hugging Face 花了 80 万 GPU 小时做实验,把 MoE 训练的坑全写成一本书,还带可跑的代码。想自己训 MoE 的直接收藏。
一个独立开发者做的可无限演变 Backrooms 游戏,36 小时就有 3.5 万人玩、峰值 1200 人同时在线,单人服务器架构也挺有意思。
The handbook turns a technique guarded by a handful of frontier labs into open material with runnable code, lowering the barrier for teams wanting to train sparse models themselves. Engineers and researchers doing model training can work through the 55 interactive figures and experiments directly. Anyone weighing dense versus MoE architectures for their next pretraining run gains a concrete reference based on 800,000 GPU hours of experiments.
一个新编程工具 T3 Code,界面挺好看,还能把 Claude 和 Codex 的历史会话直接搬进来接着用,想换工具的可以试试。
苹果居然比 AI 热潮前还少花 20% 的钱,而 Alphabet、微软、亚马逊、Meta 都烧了 5–8 倍,看看这两种路线谁扛得住。
教你怎么用 Claude Code 扫公开收入数据,找那些流量小但真在赚钱的网站,过程都贴出来了,照着跑一遍就行。
有人用 Claude Code 复刻了一个 4,800 美元单日收入的建站生意,4 家商户各付 400 美元,提示词都写在评论区了。
NVIDIA 在首尔办了场开发者活动,2000 多人到场,能现场玩智能体和机器人 demo,还有韩国开源模型的应用分享。
NVIDIA 给韩国四款开源模型出了 Dynamo 部署配方,B200 上推理快一倍,想跑这些模型的人直接抄作业。
几个中文 AI 圈的主播把播客搬上 X 了,最新一期还在聊 DeepSeek,听中文 AI 播客的可以去关注。
Gemini 产品负责人聊怎么做评测和产品取舍,60/20/20 那套分配方法对做 AI 产品的人挺实用。
Mistral 在巴黎大会上用自家 Voxtral 3 Transcribe 做实时转写,还靠 context biasing 提专有名词准确率,做活动直播字幕可以参考。
Simon Willison 分享的小技巧:把 AI 写的长篇 PR 描述塞进 <details> 标签折叠起来,reviewer 一眼看到重点,细节按需展开。
DeepMind 两位研究负责人同时放话未来几周有科学发现相关的新东西,做科研的朋友可以先关注起来。
The analysis quantifies the economics behind the 'diminishing returns' debate with a concrete figure: an 8x cost increase per capability gain over two years, based on Anthropic's own data. Investors and enterprise buyers deciding whether to budget for frontier-model upgrades can use this trend to pressure-test vendor claims. It contrasts with the scaling narratives that labs and GPU suppliers still promote.
Menlo Ventures把整个AI基础设施栈画成了一张图,代码沙盒一年内从11%涨到40%,做基础设施或智能体方向的人建议看看数据。
The 44% to 90% shift signals legacy cloud and infrastructure vendors are losing the AI budget battle to startups born for this stack, a reversal that reorders the vendor shortlist for anyone building AI products. Enterprise buyers and infrastructure founders should re-examine which layer their spending and roadmaps target, since agent tools are the fastest-growing segment. The adoption figures on sandboxes, search and browser tooling offer concrete benchmarks for sizing the agent tooling market.
The practical detail most coverage skips is her advice to rewrite accumulated workaround rules after every model upgrade — stale prompts can actively suppress new abilities. Product managers and team leads building on frontier models can start with her 5-20 prompt eval method to build their own quality bar instead of relying on leaderboards. The 60/20/20 framing offers a reusable allocation template for roadmap planning under model uncertainty.
有人用 Kling 4.0 完整做了支饮料广告,看看它产品细节、角色一致性到底行不行,做广告片的可以参考。
xAI 的 Grok Bot 用量给得很足,但最近幻觉变多,马斯克已经下场回应说当晚修复,用的人可以留意下效果变化。
归藏放出一个 GPT image 2.0 提示词,把你的咖啡照片变成满是杯子和小人的门店插画,提示词直接可抄。
The argument reframes the scaling debate: compute spending amplifies whatever generalization ability a method already has, so research effort should target methods with higher baseline generalization rather than relying on budget. Researchers choosing between architecture-driven and scaling-driven research agendas, and investors weighing AI infrastructure bets, can use this framing to evaluate long-term defensibility of current approaches.
levelsio 让 AI 10 分钟给 Hotelist 加了 9 种语言,彻底绕开 t() 翻译函数那套老办法,做独立产品的不妨参考。
有人把 WAN3 数字人视频做成了流水线,最便宜的模型跑解说视频,不用天天换工具,想省钱做视频的可以看看。
新加坡比价平台 MoneySmart 赚钱了,收入涨 25.6% 到 5010 万新元,但它也坦承 AI 搜索在抢走网站流量,看它怎么应对挺有意思。
Gary Marcus 发帖说生成式 AI 每一点改进的成本越来越高,他判断泡沫快撑不住了,看看他列举的几个理由。
Thinking Machines 把 Inkling 2 Small 放进了 Tinker,可以直接微调,还承诺用 Apache 2.0 放出全部权重,商用门槛很低。
Only the latest 2000 items are listed; for older ones see the daily archive