生成完视频还得手动导进剪辑软件?这个 Agent 直接帮你生成、拼装、用 Premiere 剪完,流程一步到位。
Latest
一位开发者把三个个人智能体工具全换成 Grok Bot 了,说 UI、功能、更新速度都到位,可以看看他为什么这么选。
Atmanity 发的 Atmee v2 挺有意思,一张照片加 10 秒录音就能做出能视频通话的数字人,做客服或家教类应用的开发者可以试试。
SambaNova 拉上 Supermicro 卖推理机,GPU 管预填充、自家的 SN50 芯片管解码,首批 1 万颗,专门冲着跑 agent 去的。
Krea 演示了从 Figma 分镜到 After Effects 关键帧全交给 Agent 的视频制作流程,做视频的可以看看这套玩法。
LMArena 搞了个 Agent Arena 排行榜,能看到各家智能体模型的排名,做 Agent 选型的可以对照着看。
开源编辑器 quiet.md 兼容飞书文档,Rust 写的,100 万字文档打字只要 5ms,本地 XML 文件还能让 Claude Code 直接改,推荐试试。
Perplexity 的 Computer 智能体帮一家龙虾商家管 3 万美元 Meta 广告,收入做到原来的 3.5 倍,还赢了他们的代理公司。
Browser-use in agent frameworks removes the need for site-specific APIs, so agents can automate tasks on any website a person can log into. Developers building task automation should look at LangChain's Managed Deep Agents to see how credential persistence across sessions works in practice. No regional angle appears in the demo itself.
Perplexity 的 Computer 现在能直接操作 Shopify 和 Klaviyo,帮你做广告、落地页、邮件营销,开店的朋友可以省不少活儿。
Perplexity 的 Computer 替 Get Maine Lobster 花 3 万美元投 Meta 广告,收入做到原来的 3.5 倍,代理公司都被比下去了,看看它怎么全自动跑的。
Pangram 刚出的图像检测模型,99.2% 的识别率加 0.27% 误报率,还直接装进 Chrome 扩展,刷网页就能查图是不是 AI 造的。
The release is a concrete counterexample to the assumption that large-scale AI-generated code is necessarily low quality, since every one of the 16M lines is machine-verified in Lean. Researchers and builders working on AI for mathematics can now treat the formalized Stacks Project as checked infrastructure rather than formalizing it themselves. Compared with prior piecemeal formalization efforts, this covers an entire interconnected corpus in one release.
Instead of forcing one-shot judgments on ambiguous cases, the framework makes model and user negotiate category boundaries iteratively, which addresses a real weakness in moderation pipelines. Teams building content moderation or annotation workflows can apply it to cut reviewer workload while tightening accuracy. The nearest comparison is standard rule-based or single-prompt VLM classification, which offers no such feedback loop.
Nace AI 在搞 120% 返现,改 3 个环境变量就能换到 Drex Max,跑 agent 任务只要 Claude Code 约 1/10 的钱。
FishAudio 的 Drama 3 语音模型要上阿里云了,能控制单个词的语气、句子中途换音色,做有声内容的朋友可以关注。
This is the strongest public counterexample to the claim that large-scale vibe coding produces only slop: a full mathematical field survived machine verification. Mathematicians and AI researchers can build new proofs on top of verified foundations rather than re-deriving them, with Lean catching any invalid step. It also shows a practical pattern for running agent swams with human-assigned roles on long-horizon tasks.
A single server layer that swaps between llama.cpp, vLLM, SGLang and MLX removes the usual lock-in to one inference stack, so the same machine can serve GGUF chat models and TensorRT-accelerated workloads. Developers building local agents can point Claude, Codex or OpenClaw at an OpenAI-compatible endpoint without rewriting code. The 3 million-model Hugging Face pull and free open-source licensing make it a direct competitor to Ollama and LM Studio.
Inference cost and latency increasingly hinge on compiler-level tuning rather than the choice of model itself, and having an agent handle that tuning continuously shifts what serving platforms compete on. Teams running high-volume inference can compare Together's 2.6x figure against their current vLLM or SGLang setups before committing. Together AI also offers ASEAN access via its API, making it a practical option for regional builders scaling production workloads.
Benchmarking by task completion time rather than tokens per second changes how verbose models rank: GPT-6 Astra beats faster rivals because it emits about 50k output tokens per task versus 120k–170k for competitors. Default-effort data shows medium effort cuts task time 3–5x at a modest intelligence cost, which directly informs cost-sensitive API choices. Developers comparing models for agentic workloads should re-read leaderboards with this metric in mind.
Fireworks AI 的 Nexus Auto 在 ValsAI Router Index 拿了第一,还是第一个过 60% 的路由器,可以直接试用看看效果。
用 Zed 跑 agent 的朋友可以升级了,v1.27 打开线程更省内存,干活时 CPU 和磁盘读取也更少。
Fireworks AI 的 Nexus Auto 用更低成本在 1222 项任务上跑赢了 Claude Opus 5.5,编码会话成本降 53%,团队版每月每席 10 美元。
Google 把 Gemini 的 computer use 工具转正了,做浏览器和桌面自动化 Agent 的可以放心上生产了。
NVIDIA 给 Agent Toolkit 加了企业数据接入和权限控制,Lila Sciences、Siemens 这些公司已经在拿来设计药物、排产了
一个做实时数字人的模型 Atmee v2,一张图加10秒音频就能生成会说话的角色,官方说比 Seedance 2.5 快50倍,适合做教学或客服视频智能体。
ElevenLabs 的 ElevenAgents 直接接进 Genesys、Amazon Connect 这些客服系统了,不用换平台就能给呼叫中心配上语音智能体,做客服自动化的可以看看。
Fireworks AI 出了个叫 Nexus 的工具,能自动帮你挑开源还是闭源模型来省钱,8 美元一个月就能用上。
Google 把科研全流程交给 AI 自动跑:选题、实验、评审、改稿一条龙,107 个顶会课题里 8 成跑赢人类水平,一篇论文只要 2.5 天。
Fireworks AI 搞了个 Frontier Ecosystem 生态,Pinecone 第一批加入。两家的组合是开源模型加向量数据库,跑在你自己的云上,做 RAG 应用的可以看看。
Sakana AI 在 Gartner 大会摆摊了,Marlin、Fugu、Namazu 三款都能现场上手玩,路过的可以去看看。
LinkedIn 把 OpenAI 的大模型蒸馏成小模型跑招聘任务,算力成本降了 67%,降本思路值得借鉴。
ClickUp 的 Brain 现在能替你操作网页:点按钮、填表单、下载文件,还能回放全过程,没有官方集成的网站也能用。
Living Brain 把记忆从 ChatGPT、Claude 里拆出来做成共享层,换模型记忆不丢,token 少 22.6 倍、成本降 95%。
Sourcegraph 的 Agentic Batch Changes 能连 Linear、Confluence、Buildkite 这些 MCP 服务器了,写代码的 agent 可以自己跑完整工作流,做批量代码变更的可以看看。
Only the latest 2000 items are listed; for older ones see the daily archive