全部 AI 动态

How can the math community forge the best path forward in the AI age? We're committing $15M to workshops and initiatives — plus free AI computing resources for foundation-supported researchers — through new partnerships with @AnthropicAI and @OpenAI: https://bit.ly/4zpfIbW #math #science #AI

恭喜 StepFun 发布 Step 5 Preview 🙌 这是他们规模最大的开放权重发布。 这个 600B MoE 模型专为处理长达 24 小时的 agent 任务而构建。 该模型在 NVIDIA GB300 上使用 NVFP4，运行速度接近 8K tokens/sec/GPU。 NVP4 checkpoint：https://huggingface.co/stepfun-ai/Step-5-Preview-NVFP4 使用 @SGL_project 部署 - https://docs.sglang.io/cookbook/autoregressive/StepFun/Step-5-Preview 或 @vLLM_project - https://recipes.vllm.ai/stepfun-ai/Step-5-Preview 使用 NeMo Automodel 进行后训练：https://github.com/NVIDIA-NeMo/Automodel/blob/main/docs/model-coverage/llm/stepfun-ai/step-5-preview.mdx 使用 NVIDIA NIM 通过单条命令运行它：https://catalog.ngc.nvidia.com/orgs/nim/nvidia/containers/vllm-model-free-nim-experimental/-

Chat models return messages. Decision models return decisions. Workflow models execute workflows. TypeLLM introduces the first workflow model. Run an entire workflow in a single API call. Simpler, faster, cheaper. Try it now: https://typellm.ai/examples/ticket-workflow

SGLang has Day-0 support for Ling-3.1-flash from @AntLingAGI! A 560B MoE hybrid reasoning model, 25B active, 1M context, built for real work: it turns a rough brief into a page, mobile UI, or interactive prototype. ✅ MXFP4 + block-FP8 quantization cuts weights from ~534GB to ~292GB, fitting on 4 Hopper GPUs ✅ Strong on real-world agent, office, and coding tasks, including GDPVal-AA, FrontierSWE, and Terminal-Bench 4.0. Get started with SGLang:

🎉 Day-0 support for Ling-3.1-flash is here in vLLM! ~560B total parameters, ~25B active per token, and up to 1M context. From a rough brief to websites, mobile UIs, and interactive prototypes—serve it on your own infrastructure with vLLM. Congrats to @AntLingAGI!

Day-zero drop: Ling-3.1-flash from @AntLingAGI is now live on DeepInfra ⚡ Built for long-horizon coding, multi-step analysis, and tool-using agents with 560B total parameters and 25B active per token. Per 1M tokens: $0.30 input · $0.90 output · $0.06 cached. Try it, compare it, and tell us what you build.

https://x.com/i/article/2110601994578866176

We’re bringing natural-language control to any scientific generator. We are very excited to introduce LLM-to-Flow Distillation (LFD), a method to adapt scientific generators to high-level natural-language discovery goals. Starting from textual goals like “design peptides with promising anticancer activity” or “design small molecules binding to the D2 dopamine receptor”, LFD distills high-level goal-related knowledge from LLMs into scientific generators through iterative DPO on LLM-generated sub-goals. 🔥 LFD does not need task-specific proxy rewards, works with any frontier LLM via API-access, and outperforms state-of-the-art reward-guided methods such as FlowGRPO and DiffusionNFT on held-out computational evaluations. (1/6)

> Anthropic第三季度营收为168亿美元，较第二季度增长约45%，较第一季度到第二季度超过100%的增幅大幅回落。 ~ 金融时报 营收在第一季度到第二季度之间翻了一倍多，从约48亿美元增至115亿美元，但第三季度的增速落后于此前水平。 > 该公司告诉投资者，每在算力上花费1美元，就能赚取1.70美元，这意味着算力成本约占营收的59%，并暗示其85%的调整后毛利率并未计入其中很大一部分支出。 > 竞争对手实验室进行的蒸馏——即用Claude的输出训练它们自己的模型——曾占Anthropic营收的5%至7%，直到该公司在第三季度大幅削减了这一部分。 > Anthropic连续第二个季度报告调整后盈利，调整后营业利润率为11%，不过高管警告称，在算力和生物医学研究上的巨额支出可能会改变这一状况。 > SemiAnalysis预计其年化营收约为740亿美元，而一些投资者在夏季曾预测，到年底这一数字将突破1000亿美元。 > Meta和Microsoft这两个主要客户已转向自家AI工具，而针对OpenAI的降价以及更便宜的中国模型，压缩了每个客户支付的金额。 > 银行家们已提出以约2万亿美元估值进行首次公开募股，募资超过1000亿美元，这将超过SpaceX在6月以1.77万亿美元估值上市的交易，成为史上最大规模的IPO。 > 一份早期的S-1草稿将近三分之一的篇幅用于风险因素，包括AI模型表现不可预测或带有恶意，以及Anthropic所描述的对人类可能存在的生存风险。

FT: The OpenAI engineer charged with bugging an executive conference room had booked one-way flights to Saudi Arabia, prosecutors say. Newly unsealed filings show Abdulrahman Alfozan, his wife and a third relative held tickets for October 12, 3 days after his arrest. Prosecutors used those bookings to get the complaint sealed on October 8, arguing he had the means and incentive to flee if he learned of the case first.

MiniCPM-V 4.7 is now open source on ModelScope: a 35B-A3B MoE multimodal model built to run on-device without compromising performance. Apache 2.0. 🤖 https://modelscope.ai/collections/OpenBMB/MiniCPM-V-47 📱 First open-source MoE with smooth, lossless on-device inference, running at 13+ tokens/s on iPhone 17 Pro Max. 🔥 New Pareto frontier: in Instruct mode, it reaches a higher overall score in less inference time than Qwen3.6 35B-A3B. ⚡ Mixed 4x/16x visual token compression lets developers balance visual detail and inference cost per task. 🛠️ Two sizes (35B-A3B and 1B) with FP8, GPTQ, AWQ, BNB, and GGUF variants. Works with SGLang, vLLM, llama.cpp, Ollama, SWIFT, and LLaMA-Factory. 💻 Runs on iOS, Android, and HarmonyOS, with day-0 support for Rockchip, Houmo, Intel, AMD, and Arm.

FT: The OpenAI engineer charged with bugging an executive conference room had booked one-way flights to Saudi Arabia, prosecutors say. Newly unsealed filings show Abdulrahman Alfozan, his wife and a third relative held tickets for October 12, 3 days after his arrest. Prosecutors used those bookings to get the complaint sealed on October 8, arguing he had the means and incentive to flee if he learned of the case first.

Bloomberg: OpenAI has put Uday Ruddarraju, its compute CTO, in charge of the infrastructure business Sachin Katti ran. Katti will leave the role over the next few weeks and stay on as an adviser, about a year after joining from Intel in November 2025. Ruddarraju came to OpenAI in July 2025 from xAI, where he led infrastructure engineering for both xAI and X, and was named CTO of compute in July 2026.

让不同平台的个人 agent 相互通讯的极简方法：电子邮件！Grok Bot 先行一步，你可以给自己的 Bot 配置一个邮箱 xxx@mail.grokbot.com 这种，例如我现在要让我的 Grok Bot 和我的 Muse Agent 之间相互发指令，可以拿 Gmail 当中介，设置一个 Gmail 的别名（xxx+muse@gmail.com）给 Muse 使用，然后通过邮箱和签名来验证，最后让你的 agent 信任这个邮件地址，发消息不需要通过你确认就可以了。 现在 Gmail 就是没法 push 消息，需要每 15 分钟检查一遍新邮件，所以说消息没法即时，其它都很方便！如果想用更强大即时的方案，可以用 @turingou 同学做的服务 http://yobiru.com 👀

My dear friends, I’d like to show you my latest toy project Introducing Handwired: turn your regular JavaScript code into a neural net! - no training - exact outputs vs the JS program - fully differentiable All runnable in a browser, without GPU

This happened when I tried to get my Grok Bot to order stuff for me on Amazon. “Continued access by an unauthorized AI agent violates Amazon's Conditions of Use, to which our customers have agreed.” My bots can no longer access Amazon for orders or anything.

Grok @bot booked a flight for me today the entire process was really really easy…..I send a photo of my credit card and it did the entire process…. Tbh, last time I booked, the website was way too confusing…. It’s not an easy website, and Grok Bot did perfectly well and booked a flight for me…. And I got a message for a booking confirmation….going to board it soon