Discover new terms
before they trend
We find emerging keywords 7–30 days before they go mainstream — with competition data, monetization signals, and actionable next steps so you can move first.
Every Tuesday · 5–8 curated terms · Free forever
Latest Terms
Stage = days since the term's current meaning emerged, not since we detected it.
"Release Mythos AI" 指美国政府于 2026 年 6 月 26 日授权 Anthropic 将 [Claude Mythos 5](https://techcrunch.com/2026/06/26/trump-admin-releases-anthropic-mythos-to-be-used-by-more-than-100-us-companies-agencies/) 部…
A24 是美国一家独立电影公司。2026 年 6 月 22 日，Google DeepMind 宣布向其[投资 7500 万美元](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/deepmind-a24-research-partnership/)，双方将联合开发 AI 电影制作工具，也是 Goo…
Claude Tag 是 Anthropic 为 Slack 推出的多人协作 AI agent，2026 年 6 月 23 日上线。它把之前那个单用户的 Claude-in-Slack 集成彻底替换掉，给每个频道分配一个固定的 @Claude 身份，团队里任何人都能 @ 它、看着它干活、或者把任务接手过来。它不是每个人各用各的聊天机器人，而是整个团队共享的协作成员。…
The Coming Loop 是 Armin Ronacher 给「套在 agent 外层的调度系统」起的名字。它决定一项任务有没有做完、需不需要注入后续消息、要不要把任务转到另一台机器继续。这一层在 agent 内循环（模型调工具、读结果、继续执行）的上方，也高于 agent harness 本身。…
Language World Models (LWMs) 是一类专门用来模拟环境状态跳转的大模型：给定 agent 的历史操作记录，预测它下一步会观察到什么。它不负责决策「做什么」，只预测「会发生什么」，充当训练和测试 AI agent 的高保真模拟器，覆盖各类数字环境。…
Unlimited OCR 是百度开源的一个 30 亿参数模型，可以把多页文档在一次推理里扫完，不需要传统 OCR 管道那种逐页切片再拼接的做法。核心技术是 Reference Sliding Window Attention (R-SWA)，让 KV cache 的大小不随输出长度增长。…
Nanostack 是 IBM 发布的一种三维晶体管架构，把互补的 NMOS 和 PFET 纳米片在垂直方向上错层堆叠，用 z 轴绕开水平缩放的瓶颈，突破 1nm 限制，进入 IBM 所称的「埃米时代」。…
Qwen-AgentWorld 是阿里云 Qwen 团队推出的第一批原生语言世界模型 (Language World Models，LWM)。这类模型从头训练的目标只有一个：预测软件环境在 agent 执行某个动作后会怎样变化。生成文字、执行操作，都不是它的职责。…
Sakana Fugu 是一个多智能体编排模型，通过单一 OpenAI-compatible API 把任务动态分发到 Opus 4.8、GPT-5.5、Gemini 3.1 Pro 等主流大模型上。Fugu 本身是训练出来的模型，不是写死的工作流，它学会的是什么时候拆包分发、什么时候交叉验证、什么时候合并结果。…
Seedance 2.5 是字节跳动的新一代 AI 视频模型，能在一个镜头里原生输出 30 秒的连续画面，不靠拼接短片段，整个过程中场景切换、角色一致性、运动物理都保持稳定。…
We spotted these before they went mainstream.
Anthropic's background-jobs-for-LLMs category now has CrewAI, Devin, Claude Managed Agents, and a dozen Show HN launches.
Karpathy's one-line Feb 2025 tweet spawned a whole genre of tools and a debate over software craftsmanship.
The discipline of shaping LLM context windows went from practitioner slang to a named specialty in under six months.
Every term comes with a full research report
What the term means, with a canonical example and a memorable analogy a creator would quote.
Where the signal is going in 6 months, with historical analogs and a confidence rating.
Four forces, cited to primary sources, balanced on controversy when relevant.
Now / 3–6 mo / 6–12 mo commercial reads, with revenue signal rating.
Content pitches, websites to build, and product ideas with SERP-gap reasoning.
Variants × TLDs checked live via RDAP — grab the name before it's gone.
From signal to strategy in three steps
We scan Hacker News, GitHub, Readwise, and dev communities daily. A cascade pipeline filters mainstream noise and surfaces terms that are new, cross-source, and acquiring momentum.
Each term gets a full research report — definition, trend projection, why-now analysis, monetization outlook, story angles, and buildable product ideas. Human reviewed.
Published Tuesdays. Use the research to write content, pitch newsletters, build products, or register domains before the term trends on Google.
Human-curated, not auto-generated
Every report is written and checked by a human. No model-only autogen, no scraping spam content.
Every claim cites an external URL you can open — official docs, launch posts, HN threads, GitHub repos.
When a term is contested (license changes, safety claims, author disputes), we name the critics as well as the proponents.
Never miss a new term
5–8 emerging terms every Tuesday. Research, story angles, buildable ideas. Free while we're in early access.
Join the waitlist for issue #1. No spam, unsubscribe any time.