Latest

Every published item, in order · 29234 items
Nov 11
Chinese contentExecuTorch 推出 WebGPU 后端,PyTorch 模型可在浏览器 GPU 上直接运行

在浏览器里跑 PyTorch 模型不用再转 ONNX 或 TFJS 了,导出一个 .pte 文件就能在 Chrome 和 Safari 的 GPU 上跑 Llama 3.2、Whisper 这些模型,还能在本地做 LoRA 微调。做 Web 端推理或关注隐私部署的开发者应该试试官方的 Playground。

Official accountPyTorch@PyTorchSource
IndustryBuildersMulti-source03:50
Chinese contentswyx:本地写代码该停了,改用云端 Agent 编程

swyx、Tailwind 作者 Adam Wathan 和 Cognition CEO Scott Wu 都在推云端 Agent 写代码,认为本地手写已经落后。和 Cursor 这类本地工具思路不同,任务直接丢给云上异步跑。如果你还在自己敲代码,可以看看他们怎么安排工作流。

Event
swyx (AI Engineer)@swyxSource
Chinese contentHedra 推出 Model Factory,为机器人等行业提供定制模型训练服务

Hedra 把自动跑实验的智能体和 GPU 打包成服务,制造和机器人公司拿传感器数据来就能换回专属模型,部署后还会继续收集数据迭代。比起自己签 GPU 长约或养研究团队,这条路便宜不少,做具身智能和工厂自动化的团队可以关注。

rohanpaul_ai@rohanpaul_aiSource
Chinese contentLatent Space 开启模型实测系列,首期评测 Naive-N0.5-Flash 编程模型

Latent Space 新开了实测栏目,首期把 Naive-N0.5-Flash 和 DeepSeek、Z 的模型掰开比速度。这模型主打开放权重、专攻 AI 编程、追求 token 每秒吞吐。如果你在给 coding agent 挑推理引擎,这篇的速度数据很实用。

Latent.Space@latentspacepodSource
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
Chinese contentLovable 集成 ElevenLabs,应用可加入语音与音乐功能

用 Lovable 做应用的朋友注意了,现在不用自己接 API 就能加上 ElevenLabs 的语音代理、实时转写和音效生成。相比单独去 ElevenLabs 申请密钥写代码,直接在 Lovable 里配置省事不少。做语音交互类应用的开发者可以先试试。

Lovable@lovable_devSource
Chinese contentLenny Rachitsky 报告:工程岗位增长停滞,PM 岗位在 AI 时代逆势上升

Lenny Rachitsky 追踪五年的招聘数据第一次出现工程岗停滞:增速从 +29% 掉到 +2.5%,设计岗创新低,PM 反而在涨。打算找技术工作或做招聘规划的人,这份报告里的分岗位趋势很值得对照看。

Lenny Rachitsky@lennysanSource
ProductsBuildersSelected03:26
Microsoft Research Builds Synthetic Enterprise Environments to Test Copilot

The non-obvious takeaway is that evaluation data, not model quality, is emerging as the enterprise AI bottleneck, and synthetic environments are one workable answer. Enterprise IT teams and compliance officers evaluating Copilot or rival assistants such as Gemini for Workspace can borrow this pattern to run safe pilots on simulated tenants instead of production data. It also sets a benchmark for vendors claiming domain-specific accuracy without third-party evidence.

Official accountMicrosoft Research@MSFTResearchSource
ProductsBuildersMulti-source74°03:26
Chinese contentAnthropic 在 Claude 内推出 Whiteboards、Sheets、Meetings 等办公工具

Anthropic 直接把 Claude 变成办公套件了:能进 Google Meet 会中干活、按一句话生成带公式的表格、把头脑风暴变成看板和幻灯片。用 Team 或 Enterprise 版的团队可以少开好几个软件,跟 Microsoft Copilot 正面碰上了,值得升级试试。

Event
The Rundown AI@therundownaiSource
ProductsUS vs ChinaMulti-source84°03:25
Chinese contentAnthropic 推出办公套件,微软股价一度跌超 3.5%

Anthropic 把表格、白板、仪表盘直接搬进 Claude,说一句话就能生成可协作编辑的电子表格,还不用买 Office 订阅。微软股价当场跌了 3.5%,说明资本市场真的觉得这会动到 M365 的生意。经常做报表和数据看板的企业用户可以先试试 Claude Data。

Event
rohanpaul_ai@rohanpaul_aiSource
ProductsBuildersMulti-sourceSelected03:20
Claude Mythos 5.2 lands in Devin's managed security scans

The 16.6-point jump on Cognition's own cyber benchmark suggests agentic security scanning is now sensitive to frontier-model upgrades, so teams running Devin scans should evaluate Mythos 5.2 against their current configuration. Enterprises with large codebases and limited security headcount are the direct beneficiaries of the US$100M credits programme. Compared with running vulnerability triage manually or with generic coding assistants, this pairs an autonomous agent with a model tuned for the task — useful reference for security engineering teams weighing agentic tooling.

Event
Cognition@cognition_labsSource
ProductsBuildersMulti-source03:20
Chinese contentChatGPT Business 用户可认领商家并搭建 Business Agent 接待客户咨询

OpenAI 在 ChatGPT 里给美国商家开放了认领入口,认领后能建一个 Business Agent 自动回复客户咨询。配合今年早些时候的 ChatGPT 广告投放自定义 Agent 的能力,等于把流量到转化的链路做完了。做本地生意或营销的朋友可以关注怎么在 ChatGPT 里占位。

Event
TestingCatalog@testingcatalogSource
ProductsBuildersMulti-source03:14
Chinese contentOpenAI 推出 Business Agent,面向美国电商及在线服务企业

商家现在可以在 ChatGPT 里开一个自己的导购对话窗口,标注清楚是企业提供的。跟 Perplexity 和 Amazon Rufus 那种平台主导的模式不同,这次是企业自己上传资料来配置。做美国电商或在线服务的朋友可以看看是否符合开通条件。

Event
Official accountOpenAI@OpenAISource
ProductsBuildersMulti-source03:14
Chinese contentOpenAI 推出 Business Hub,商家可管理 ChatGPT 中的商业信息

OpenAI 给商家开了个后台,认领后能在 ChatGPT 里放产品价格、库存,还能看被提及和点击数据,有点像 AI 版的 Google 商家资料。做本地生意或电商的朋友可以早点认领占位,别让自己的信息被模型写错。

Event
Official accountOpenAI@OpenAISource
ProductsBuildersMulti-source03:08
Chinese contentFish Audio 发布全双工语音模型 Live 3

做语音智能体的朋友可以看看 Fish Audio 新出的 Live 3,能听懂"嗯嗯"这类反馈,也不会在用户话说一半时插嘴。相比传统串行语音助手,全双工让对话自然很多,企业还能免费申请定制,开发者可以直接上手试。

Event
Fish Audio@FishAudioSource
IndustryPolicyMulti-source74°03:07
Chinese contentAnthropic 二次修改 S-1:确认越过负责任扩展政策门槛,IPO 定价 70-78 美元

上市文件里直接写明自己越过了 AI 加速 AI 研究的安全门槛,这在公司招股书里挺少见。定价 70-78 美元也一并公开了。想看头部实验室怎么把安全政策和上市披露绑在一起的,这份文件值得翻一翻。

Event
rohanpaul_ai@rohanpaul_aiSource
ModelsBuildersSelected03:03
ElevenLabs releases Scribe v3 with live diarization

Turn detection based on whole-conversation context rather than silence addresses a persistent failure point in live voice agents, and a 10,000-term library makes specialized vocabularies practical without fine-tuning. Teams building call-center, legal or medical transcription pipelines should evaluate Scribe v3 against Deepgram and AssemblyAI on their own accented and code-switching audio. The Eneco result, 40% more first-try correct Dutch postal codes, is a useful signal for multilingual deployments in European and ASEAN markets.

Official accountElevenLabs@elevenlabsioSource
ProductsBuildersMulti-source03:03
Chinese contentElevenLabs 发布 Scribe v3 与 Scribe v3 Live 语音识别模型

ElevenLabs 把语音识别更新成了 Scribe v3,还加了实时版本,直接内置进 ElevenAgents 智能体。做语音对话或转写应用的开发者可以试试 ElevenAPI 的调用方式,跟 Deepgram、AssemblyAI 比一下延迟和准确率再选。

Event
Official accountElevenLabs@elevenlabsioSource
ProductsBuildersMulti-sourceSelected03:02
ElevenLabs launches Scribe v3 and Scribe v3 Live speech-to-text models

Noisy multi-speaker accuracy and sentence-boundary detection address the failure points that make transcripts from Whisper or basic APIs need heavy cleanup, reducing post-editing work. Developers building meeting notetakers, call-centre analytics or live captioning should evaluate Scribe v3 Live against AssemblyAI and Deepgram on their own audio. ElevenLabs is a US-based lab, so this widens vendor choice for teams globally without regional pricing or availability constraints announced.

Event
Official accountElevenLabs@elevenlabsioSource
Chinese contentLegora 自家法务用自家 AI,5 个月完成 1 轮融资和 4 起收购尽调

法务 AI 公司 Legora 拿自家法务当案例:5 个月做完整尽调,省下 10 万美元外部律师费,合同改写从 5 小时降到 1 小时。对想用 AI 精简法务团队的创业公司和企业法务,这份一手复盘比供应商宣传更有参考价值。

WeAreLegora@WeAreLegoraSource
ModelsUS vs ChinaMulti-source02:55
Chinese contentGemini 4 Carbon 与 Fable 5.2 对比测试:Gru 3D 体素重建

有人拿两个模型做了同题对比:从同一段《神偷奶爸》视频重建 3D 体素 Gru。Gemini 4 Carbon 花 $24.11,Fable 5.2 花 $53.34,但只有后者的 Gru 像电影里的。做视频转 3D 或游戏素材生成的可以看看两边差距到底在哪。

Event
@atomic_chat_hq@atomic_chat_hqSource
ProductsBuildersMulti-sourceSelected02:54
Hedra launches Model Factory for autonomous model post-training

Post-training as a managed service with autonomous research loops changes who can run serious model improvement — vertical teams no longer need in-house RL infrastructure. Companies in robotics, manufacturing or defense building simulators should evaluate it against hiring their own post-training team. Hedra's internal Physics-IQ gain gives a concrete, if self-reported, benchmark to scrutinize.

Event
elvis@omarsar0Source
PapersBuildersSelected02:53
LingBot-Touch: tactile foundation model transfers across robot sensors

A frozen tactile model matching a sensor-specific adapter is the notable result — it suggests touch data can become reusable infrastructure rather than per-robot sensor data, similar to what ImageNet-style pretraining did for vision. Robotics teams building dexterous hands should evaluate it as a drop-in feature extractor before collecting sensor-specific datasets. Weights and code are open, so benchmarks can be replicated directly on any tactile hardware.

elvis@omarsar0Source

Only the latest 2000 items are listed; for older ones see the daily archive