2026年11月21日 · AI 早报
本轮讨论集中在数学家与 Claude 合作证明黎曼猜想的声明及其核验进展。与此同时,循环计算的开放权重模型、SSD 专家流式推理和智能体运行框架,提供了几条降低部署成本、提高任务可靠性的具体路线。
11月20日 23:00 — 11月21日 11:00(日本时间) · 145 条 X 原帖
本期目录
黎曼猜想证明声明,焦点转向独立核验
先有作者声明,尚不能写成难题已经解决
The Rundown AI 转述,Anthropic 数学家 Levent Alpöge 称与 Claude 找到了黎曼猜想的证明。多条帖子进一步转述,同一方法涉及广义黎曼猜想,作者公开了 Lean 形式化文件,而非传统论文;据其说明,尝试导出的自然语言版本超过25万页。
Quanta 的报道强调,这次工作依赖数学家的专业判断和反复迭代,并非简单提示模型就得到结果。社区最关注的,也从“Claude 能不能做数学”转向人类如何理解、审查体量如此庞大的证明。
来源 · 3 条原帖
编译进度、命题审计与另一份证明须分开看
Max For AI 较早记录的公开构建尚未完成,后续还需检查最终定理和公理依赖;本批较晚的 AB Kuai.Dong 帖子则称已有独立专家成功编译验证。输入没有附上这些专家的核验原帖,不能据此宣布数学界已确认。
Max For AI 另转述 Ravi Vakil 公告,称 Alessandro Masullo 提出了不使用 AI 的广义黎曼猜想证明草稿,同样尚未最终确认。另一个可实际探索的项目来自 Jude Gomila:他公开了长期积累的 Zeta 应用与代码,包含可视化、Lean 项目和错误修正记录,并明确承认仍有问题要修。
来源 · 4 条原帖
公司动态与桌面智能体入口
Anthropic 上市消息与创始人回顾同时出现
EveryDev.ai 的周报转述 Anthropic IPO 筹资935亿美元;“科技那些事”称其上市次日收于88.91美元,下跌约8.5%,盘中接近83美元发行价。这些是本批资讯账号的转述。Anthropic 官方本轮发布的则是七位联合创始人讲述公司历程的内容,不能把两者合并成一则官方上市公告。
来源 · 3 条原帖
Gemini 的桌面操作体验,开始与编程额度分工
Jerry Xiong 自述,在 Mac 版 Gemini 开启 Computer Use 后,已让模型操作 Chrome 查询航班;他还观察到客户端任务没有消耗自己的 Antigravity 周额度,因此把轻任务和复杂编程分开使用。OLDPAN 则更喜欢 Antigravity 的网页远程控制,但认为速度偏慢、客户端完善度仍不足。这些是个人使用体验,不能推广成统一额度承诺。
来源 · 2 条原帖
编码入口整合出现预览,其他更新仍以转述为主
blankspeaker 从新版 Cursor CLI 下载逻辑和实际登录界面发现,Cursor CLI 与 Grok Build 可能正在整合,预览支持 Cursor 或 Grok 登录,并显示多家模型。作者明确称其为非官方测试预览,尚非正式发布。另有社区帖子提到 ChatGPT 改版、Plus 的 Spaces 和 Codex Remote 改进,但本批没有包含这些功能的完整官方说明。
来源 · 2 条原帖
开放权重的新路线与蒸馏争论
Inkling 2 Mini 把循环计算开放给研究者
Alex Robey 宣布开放 Inkling 2 Mini 权重,目的之一是研究循环计算对思维链可监测性和内部计算解释的影响。Guangxuan Xiao 给出的结构是24层、中间12层循环三次,总参数约260亿、每词元激活54亿,采用原生 NVFP4,并从更大模型做 logit 蒸馏。它的看点兼有部署体积与研究可观察性。
来源 · 2 条原帖
大模型开放与小任务集强化,各有取舍
2logics 转述腾讯开放 Hy4 权重:总参数7700亿、激活490亿,支持文本、图像和视频,使用腾讯自有许可,大型托管服务另有授权条件。另一条社区介绍的 WeMM-Video 则强调连原声一起理解视频。
编程方向,Marktechpost 报道 DaisyRose-27B 只用强化学习,围绕约1200个已知失败任务训练。其自报 Terminal Bench 4.1 从5.6升至12.0,但其他能力提升不均,说明定向训练的收益仍须按任务看。
来源 · 3 条原帖
OpenAI 公布防蒸馏研究,社区追问开放模型进步来源
OpenAI 官方介绍了能力如何经蒸馏转移,以及如何发现、干预协同提取行为,强调能力可能脱离原有安全措施被复制。Ethan Mollick 据此提出,蒸馏变难可能影响开放权重模型的进步速度,但也承认蒸馏此前贡献多大并不清楚。这是解释性猜测,本批没有建立因果关系。
研究进展:科学任务、视觉记忆与验证盲区
药物研究需要专门评测,也需要偏差审计
OpenAI 发布药物化学推理基准 MedChemBench,并介绍与 Perturb 和 Scripps Research 围绕神经退行性疾病的合作;帖子没有给出可直接比较的完整成绩。Maziyar Panahi 转述的一篇医学大模型偏差综述覆盖201项研究,其中44项使用真实世界数据。他由此主张,开放权重的价值也在于方便审计。
来源 · 3 条原帖
视觉推理与生成速度,出现不同研究切口
NVIDIA 的 WorldTrace 针对交互视频中“转头再看却变了场景”的问题,无需重训即可补充地点记忆,官方称 LoopBench 场景回忆比滑动窗口方法改善32%。AVTR 作者则介绍让模型按问题选择视觉想象或文本推理。
本轮还重新介绍了10月的 NCM 研究:先生成概念层次,再并行补细节。转述中的21倍、31倍加速分别限定在 LM1B、XSum 的小规模实验,比较对象是扩散语言模型基线,不能外推到当前聊天产品。
来源 · 3 条原帖
能检查,不代表能发现所有错误
Keshav Shenoy 介绍的研究发现,即使拥有完整日志并能运行实验,人类审计者通常仍难区分正常与遭破坏的安全研究,前沿模型也漏掉多数破坏。Blockster 则转述 TensorLean 用数学证明检查计算内核,报告发现25个缺陷。
Damien Benveniste 的例子更贴近训练实践:求列表最大值时,错误的零初始化可通过正数测试,却会在全负数输入失败;若测试遗漏这种情况,强化学习反而会奖励错误解法。
来源 · 3 条原帖
推理成本从显存延伸到磁盘、网络与记忆
SSD 专家流式加载,开始强调实际任务速度
Magnitude 预告周一发布引擎,通过专家缓存、预取和跨层 NVMe 调度,宣称单张游戏显卡运行 DeepSeek V4.1 Flash 可达40余词元每秒,目前仍是团队预告。
TwoPeanuts 转述的另一项实验,在32GB M5 Mac 上把 Qwen3.8-Flash-Next 的多数专家留在 SSD,特定量化与短回答测试约14词元每秒。作者仍把编程交给双4090桌面机:能装下大模型,与多轮智能体任务是否好用,是两道门槛。
来源 · 2 条原帖
压缩 MoE 通信,收益伴随精度损失
Jinyan Su 介绍的方法利用相邻 MoE 层隐藏状态相似,只定期发送完整参考状态,其余发送压缩差分。在四机 Qwen3-30B-A3B 实验中,跨机流量减少58%,18项基准均分也从66.9降至63.1;另一个32张 H100、vLLM 配置报告吞吐增加22.6%、首词延迟降低23%。这是通信与精度的交换。
来源 · 1 条原帖
长记忆压缩,尝试绕开反复灌入历史
lifcc 转述 MEMORILLA:冻结基座模型,把历史离线编码,再针对问题映射成16个软记忆词元。其介绍报告八项基准的上下文开销几何平均减少394倍,并在 TextWorld 中提高通关率。这提供了与原文检索不同的路线,但这些收益仍限定在帖子所述实验,不能当成任意文档均可无损压缩。
来源 · 1 条原帖
Harness 工程落到工具、轨迹与数据语义
可靠循环需要验证,也需要分工
Hanako 对 Harness Engineering 材料的归纳,把规则文件、测试、重试预算、外部记忆、权限和日志放进同一运行框架。Karan 展示的机器学习流水线已覆盖数据分析、训练、评估和部署,但关键决策仍交给人类。
李博杰则分享个人多智能体经验:让 Fable 处理计划与难题、Opus 管理、Sonnet 执行,并交叉使用不同模型审查。其“效率最高”来自自己的实验,适合视为分工案例。
来源 · 3 条原帖
工具按需加载,轨迹避免重复存储
lifcc 的网关实践建议只常驻少量高频工具,其余通过搜索动态加载;支持原生延迟加载的模型直接转换,不支持的则由网关模拟搜索并追加定义。TrajectoryStore 处理另一种重复:每条消息和工具结果只存一次,再由调用关联,减少长轨迹被评审、检索和训练流水线反复读取的冗余。
来源 · 2 条原帖
检索与查数,把结构放回模型前面
byte 介绍的 PageIndex 用文档树和模型推理定位信息,避免向量库、嵌入和切块,其98.7%的 FinanceBench 成绩来自该帖转述。另一篇社区介绍中的 NVIDIA auto-ontology,则把业务概念对应到数据库表、列和指标口径,先检索语义层再生成 SQL,并附推导过程。两者都试图减少模型直接面对无结构材料时的猜测。
来源 · 2 条原帖
智能体走向真实操作与可编辑创作
机器人演示开始报告完整任务循环
Bingxuan Li 展示本地 Qwen3.8-Flash-Next 以5Hz 控制 xarm,自述没有使用真实机器人数据训练。Sentdex 则在 DGX Station 上测试 DSV41F-1017:拾取任务从此前约五分钟缩至40秒,加入寻找玩具和容器后约一分钟,视觉、规划与控制均由模型配合机器人 Harness 完成。这些仍是作者的特定演示。
来源 · 3 条原帖
操作之前的授权,开始成为独立一层
Yoshio 的开源 Mac 项目让用户语音指挥 Claude Code,每次点击提前显示,批准必须口头确认。Tommy Lee 转述 Google 的语义治理预览:在智能体外部,按规则和用户意图检查工具调用并记录裁决;他也指出,使用大模型裁判意味着结果带有概率性。桌面控制的便利,正在推动更明确的执行边界。
来源 · 2 条原帖
创作工具连接真实地理、素材与修改意见
Google DeepMind 展示 Gemini 4 Neon 在 Antigravity 中编写的勃朗峰滑雪模拟器,使用自定义着色器、没有预制3D模型;April 的 Pocket City 则把 OpenStreetMap 地理数据、当地天气与季节转成可探索的手绘微缩城市。
更轻量的实践也有价值:ffmpeg-skill 在本地处理素材并检查时长、尺寸和音量;Clicknote 把网页元素的选择器、样式和截图随修改意见交给编码智能体。后者的介绍者明确尚未亲测,仍处早期。
根据本轮 X 原帖整理。作者观点、个人体验与待确认消息已在正文区分;互动量按采集时记录。
全部 145 条原帖
@AnthropicAI
Anthropic’s co-founders on the story so far. Told by Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah https://t.co/AkKNhIxLXS
@OpenAI
MedChemBench is a new benchmark for AI reasoning in medicinal chemistry. It shows the progress our frontier models have made in their ability to help scientists use AI to identify promising drug candidates. https://t.co/4N1H77z2YU https://t.co/ISAWR1tkt1
@MaxForAI
🚨有人在直播验证Claude证明黎曼猜想,目前进度99%! 有人专门搭了一个网站,正在公开直播验证Anthropic数学家Levent Alpöge和Claude提交的黎曼猜想Lean形式化证明。 这哥们真的下血本,直接用了128核CPU、244GB内存的服务器,下载了全部14个压缩包,总计228.2MB,里面包含175,187个Lean源文件。 目前已经运行了47分钟,完成26,088/26,329个构建任务,暂时没有任何构建失败! 网站每两秒刷新一次,整个编译过程、日志、源码校验和全部公开。 目前已经确认,Main.lean中确实声明了riemannHypothesis : RiemannHypothesis,使用的是mathlib中已有的黎曼猜想定义。 不过目前还在编译阶段,后面还有最终定理和公理依赖审计。 如果最终真的通过完整验证…… 那我们可能真的要见证历史了。 这实在太疯狂了。 地址:http://106.14.11.164/
@OpenAI
Distillation can reproduce powerful AI capabilities without the safeguards built around them. We’re sharing new research on our work to combat distillation of our models, how these capabilities transfer, and how we detect and disrupt coordinated attempts. These findings are helping us sharpen our defenses and support safeguards across the industry. https://t.co/FGQ0PHllIO
@MaxForAI
卧槽????今天数学界到底怎么了??? 就在Anthropic的数学家宣布和Claude一起证明了黎曼猜想之后,美国数学会主席Ravi Vakil突然发布了一份紧急公告:另一位数学家Alessandro Masullo也提出了广义黎曼猜想(GRH)的证明! 最炸裂的是,公告特意强调: 全部由人类完成,没有使用AI!!! 而且这还不是今天临时搞出来的。 Vakil透露,斯坦福数学家Brian Conrad已经研究了这份证明好几个月,提供了大量反馈,原本还打算让作者再花一年时间完善论文,经过更多人的审查后再正式公布。 结果现在直接提前公开了草稿,希望全世界的数学家一起来检查、补全可能存在的漏洞。 也就是说,困扰数学界167年的问题,居然在同一天出现了两份声称可以解决它的证明。 一份由Claude深度参与,一份声称完全由人类独立完成。 当然,目前两份证明都还没有得到数学界的最终确认。 但这剧情真的越来越离谱了。。。 人类数学家和AI,居然可能在同一天冲到了黎曼猜想的终点线前。
@GoogleDeepMind
Gemini 4 Neon coded a ski simulator based on real Mont Blanc terrain. Built in @Antigravity with custom shaders and zero pre-made 3D models, it renders carved snow and avalanches live. Try it now → https://t.co/s5ERa2jrIE https://t.co/StsoefhA0U
@emollick
Riemann Hypothesis may have been solved with the help of Claude. “what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater.” https://t.co/sACo17cL8A https://t.co/fiv9ZvBEWI
@OpenAI
How AI is helping scientists uncover new leads for understanding and treating neurodegenerative diseases—and the brain research behind our collaboration with @perturbai_tx and @scrippsresearch: https://t.co/nqgl1lI959
@MaxForAI
卧槽????黎曼猜想可能真的被Claude证明了??? 刚刚,Anthropic的数学家Levent Alpöge在个人网站上发布了一份声明,声称他和Claude一起找到了黎曼猜想(Riemann Hypothesis)的证明。 更离谱的是,他说这套方法还可以直接推广到广义黎曼猜想(GRH),而且他们已经用Lean完成了形式化证明,并公开了相关文件。 要知道,这可是1859年提出、困扰数学界167年的黎曼猜想,数学界最著名的七大千禧年难题之一。 而且Levent Alpöge也不是什么民间数学爱好者。他是普林斯顿数学博士、哈佛前Junior Fellow,现在就在Anthropic工作。 他在声明里还提到了一个非常恐怖的细节: 在证明的最后阶段,Claude的进展速度已经远远超过了他本人。 他甚至形容,Claude在终点线前飞了起来、跳起了舞。 整个证明极其庞大,他尝试生成的完整文字版本甚至超过了25万页,目前先公开了Lean形式化文件。 当然,现在还只是作者本人宣布证明成功,尚未经过数学界独立验证。 但如果这个证明最终成立…… 那我们今天可能真的见证了数学史和AI发展史上最重要的时刻之一。
@CopyRebeldia
Los matemáticos acaban de perder su problema más famoso contra una IA. Es el equivalente matemático de curar todos los cánceres, y se anunció en un PDF de 4 páginas. Ni rueda de prensa ni paper. Cuatro páginas y esta frase: «Creo que he descubierto, trabajando con Claude, una demostración de la hipótesis de Riemann». La hipótesis de Riemann. Abierta desde 1859. Un millón de dólares de premio. Lo que sería algo así en otros campos: • Medicina: curar todos los cánceres • Física: unir la relatividad con la cuántica • Informática: resolver P contra NP • Biología: explicar cómo empezó la vida • Energía: fusión nuclear barata Todos siguen pendientes. Este es el primero con una demostración sobre la mesa: la hicieron un matemático y una IA en 41 días, y un ordenador la está comprobando ahora mismo. Y no llega solo. Destraba cientos de teoremas que dependían de él, y el mismo método ya alcanza su versión más general. La herramienta que lo hizo sigue encendida. Lo demás de la lista viene detrás.
@Aprilzz423
Introducing Pocket City~ This is an AI skill that turns any street corner into a tiny hand-drawn city you can turn around and explore~ Measured from OpenStreetMap. Built with real weather and seasons, researched for each place: Singapore gets dry and wet monsoons, haze, even F1~ Built with Claude Fable 5.2 + GPT-6.2 Astra. Free and works in Claude as well as ChatGPT~
@poetengineer__
strange flowers of the mind are blossoming, silently and very fast. training a tiny llm in the browser. https://t.co/UBGRFaK4PY
@tomgreenwald
SSD is about to become the only limiting factor for local AI Most frontier open models are ultra sparse MoEs Qwen 3.8 Flash Next = 180B total, ~6B active per token (~3%) GLM 5.3 Flash = 321B total, ~18B active per token (~6%) DeepSeek V4.1 Flash = 552B total, ~16B active per token (~3%) GLM 5.3 = 753B total, ~40B active per token (~5%) Kimi K3.1 Pro = 2.8T total, ~104B active per token (~4%) When ~95% of a model sits idle on every token, keeping all of it in VRAM makes no sense Local AI has been evolving toward this for years: 2023-2025 - Layer offloading (Ollama, classic llama.cpp) This chopped whole layers across GPU and CPU, leaving the GPU waiting on CPU once VRAM filled up Oct 2026 - RAM expert cache (llama.cpp --moe-cache) This shifted from whole layers to individual experts, caching hot ones in VRAM and pulling misses from RAM. Huge leap for decode, but still capped at the size of your RAM Now - SSD expert streaming Early tools like Colibri and Strata proved direct NVMe streaming actually works on consumer hardware. Instead of hitting a VRAM or RAM wall, your SSD becomes an active part of the engine However most people hear "run from SSD" and assume 2 tok/s, which isn't entirely wrong. A few problems still exist: - Basic caching kicks out experts you're about to need again - Waiting on disk reads one layer at a time leaves the GPU and SSD sitting idle - Getting speculative decoding to work well with SSD streaming is hard - Long prompts are tough to keep fast without wiping out your hot cache - Most early tools are experimental engines built for only a couple models Making expert streaming feel fast takes solving all of those together At Magnitude, we've been working on this for the past 2 months. Here's how we solved it: - Smarter caching that keeps your session's most-reused experts in GPU memory - Speculative lookahead that pulls upcoming experts off SSD before the GPU needs them - Direct NVMe reads scheduled across layers so your SSD never sits idle - Streaming prefill that reads each cold expert once without wiping your hot cache - Built into a full engine that tunes itself to your chip across Macs and GPUs All of this together means DeepSeek V4.1 Flash on a single gaming GPU at ~40+ tok/s... We're excited to release it on Monday GitHub here: https://t.co/5SjPQWkmrF
@AlexRobey23
Today we’re releasing Inkling 2 Mini, a looped transformer, with open weights. A big reason we're doing this is to let the AI safety community study looping, especially how loops affect chain-of-thought monitorability and how to interpret the computation inside loops. 🧵
@emollick
The effective end of easy distillation from OpenAI and Anthropic models may be behind the relative slowdown in open weights model progress. We don’t know the extent to which distillation has been key to the rapid improvement in open weights, I guess we will find out quite soon.
@htahir111
Happy Friday! Promised a series on how to write good evals. Ended up writing a 3500 word X article. Will roll this out into a bigger series in the upcoming days. Hope this is a good weekend read! https://t.co/HpQeB582RI
@silkcompute
Introducing Silk Compute, the new frontier of building. Access to abliterated frontier models, served through the OpenAI and Anthropic APIs. For teams doing red teaming, security research, synthetic data and ML research that mainstream platforms aren’t built for. Now serving Kimi K3.1 Pro, DeepSeek V4.1 Flash and GLM-5.3, completely uncensored. Build beyond the expected.
@Haster18_
for a dev tool, the best leads are github issues. 8:00 my agent reads issues asking for the feature we ship 8:01 9 signals, 4 companies 8:02 maintainers' emails verified @MonidHQ, $0.0001 a signal. https://t.co/MQZfepvXzf
@TokenGremlin
HOLY SHIT... Bookmark this tweet and read it carefully. Claude, which was already the best coding model in the world, may have just become one of the best at math too. Little by little, there are fewer and fewer areas left for humans to dominate.
@TheRundownAI
NEW: Anthropic mathematician Levent Alpöge says he and Claude have solved the Riemann Hypothesis, a 167-year-old Millennium Prize problem that many consider the most famous open question in math. https://t.co/yOHMZB54UA
@byteonai
The entire RAG industry is about to get cooked. Researchers built a new RAG approach that: - needs no vector DB. - doesn't embed your data. - does zero chunking. - runs no similarity search. It's called PageIndex. Instead of chopping your docs into chunks and dumping them into Pinecone, it builds a tree index and lets the LLM reason through it like a human reading a book. hit 98.7% on FinanceBench. beat every vector RAG on the leaderboard. no embeddings. no chunking. no vector DB. 100% open source.
@QuantaMagazine
The approach to solving the Riemann hypothesis is markedly different from the 377 results OpenAI released in October, almost all of which were the result of a simple prompt to an AI model. Instead, Levent Alpöge’s mathematical expertise played a major role in iterative attempts with Claude to move closer to the result. https://t.co/yOtgd8SDUO
@judegomila
So the universe is harmonious and on this monumental day of history - the beauty of the primes has likely been proven. Today based on the Alpöge's @__alpoge__ + @AnthropicAI Riemann Hypothesis result claim (congratulations), I'm opening all my work I've been doing in private to a public repo + app so that the community can explore the map of the zeta function, primes and the universe together. I have been working on this for 17 months on and offline for the pure love and curiosity of it. Between walks, paper, terminal and dream work, here is what I have for the community. Zeta App: https://t.co/MHXntQmVJ7 Zeta Repo: https://t.co/bpPS65uYN6 I used a large variety of methods, models, techniques and help from certain people that may not want their names mentioned yet. If this repo has useful results we can go into that later. As Tao said, we want to flesh out the map, so here is some of that map. I didn't get to prove RH, but here's what myself and the AI fleet found instead over an intense 17 month period of non stop work (I have had to rush to release based on the high pace of the art): 0. Twin ladder model and bridge shear. 1. No-go theorem. No "soft" heat-flow hypothesis can prove RH. Arithmetic input is required. Machine-verified, resting on Rodgers–Tao. 2. Frozen-rung ⟺ Platt–Trudgian. Freezing the ladder rungs is exactly the Platt–Trudgian mechanism, capped at the verified zero height. 3. What really binds Λ ≤ 0.2. An audit, confirmed three ways: the limit is the verified zero height. Tightening the barrier can't beat it. 4. 44 routes to RH. Every one is either non-rigorous or equivalent to RH itself. 5. Permutation atlas. Every rearrangement family classified. Isometries provably give no gain; one family can. 6. Two open conjectures are now theorems. The Ramanujan-reversal identity and the row sum R_n(1) = n^(n−1), kernel-verified in Lean. 7. 20 kernel-verified Lean 4 projects, including theory mathlib lacks: a general argument principle on rectangles, Abel's binomial theorem, tree-function ODE theory. 8. Exact closed forms for when zeros stop being isolated under heat flow, including c* = √(2+√3) − 1. 9. Candidate bounds on the de Bruijn–Newman constant. 10. The Λ ≤ 0.14 route reduced to one open family. It is a conditional contract, not a bound. 11. Zeta on a torus, sphere, Klein bottle and projective double cover, computed live in your browser, with heat flow you can scrub and fall lines draining into each zero. I have especially had fun with thinking in these realms personally. 12. A corrections ledger. Every caught (prepublish) overclaim is logged. Dead ends being logged for all to see. 13. And more much. Corrections, analysis, additions and PRs welcome. I have a lot of work in progress ideas, papers, proofs, etc. Some parts I'm racing to fix as I have some upgrades coming that are backed up 55 PRs at least. Be patient with me while I figure out the release and prune the bonsai. I spent a lot of time on this and I expect there to be errors to fix and upgrades to do. Hopefully there is something useful in there for us to learn/hack/explore/teach together in the open and add to the fantastic work through history of the best minds. We not only stand on the shoulders of giants, we took a leap into the air and started to flap our wings. There will be a lot more to do with the zeta and view this day as an opening to apply it elsewhere and explore more the beauty of nature...
@Jerry_Galaxy
这回爽了,Gemini用户终于有了属于自己的Computer Use,不用再羡慕Codex。 如果你是Mac,更新最新Gemini客户端,设置里开启Computer Use功能,添加你希望Agent代替你操作的目标应用比如Chrome、Excel、CAD等等,现在都可以让Gemini 4模型帮你完成,我刚让他通过这个功能操作Chrome帮我查询30天内最便宜的去香港航班并推荐最舒适的行程,完全没问题,就跟你分身操作似的。 最牛逼的是什么?他额度是跟Antigravity完全分开的,也就是你Antigravity weekly额度变0了,这Gemini客户端也继续使用,我今天用它做了几个任务,Hydrogen模型我周额度居然没消耗,你说够不够。 我的习惯是,重度Agent编程和复杂任务,一定是走Antigravity(结合编排、grill、boost、swarm)能产生非常好的质量。但其他轻量级任务,完全不去浪费Antigravity的,就使用Gemini客户端就行了,这相当于把你Gemini额度double了一下,各位Get到了吧!
@volcaholic1
JUST IN: Anthropic researcher says he used Claude to prove the Riemann Hypothesis. The Riemann hypothesis is a famous, unproven math conjecture about how prime numbers are spaced out along the number line. Proposed by Bernhard Riemann in 1859, it remains one of the biggest unsolved problems in mathematics.
@Guangxuan_Xiao
24 layers, middle 12 looped 3x, pruned to 26B total / 5.4B active, natively in NVFP4 🚀 that's Inkling 2 Mini: a small point on our scaling ladder 🪜 logit-distilled from a big one. small enough to run on your own device, strong enough for agentic work open weights!
@kshenoy_
New paper: It is difficult to detect when AI safety research is sabotaged and even harder to identify the specific sabotage. Even with access to the full research log and the ability to run experiments, human auditors usually couldn’t distinguish benign from sabotaged experiments or determine how the sabotage was done. Frontier AI models likely do better, but still miss most sabotage.
@tspy
微调后的 Qwen2.5-Coder-32B,做PCB设计竟然击败了 Opus 5.5 基于 245 个 KiCad 电路板微调 Qwen2.5-Coder-32B,然后在 23 个未见过的开发板上盲测 Qwen微调后的设计,其中有 83% 的开发板达到百分百连接率,有 48% 的开发板通过 KiCad DRC 检查无误 微调之前,Qwen2.5-Coder-32B 在所有的开发板上都无法正常工作 与其它前沿模型的测试,成绩(百分百连接率/通过检查率)如下: · Qwen2.5-Coder-32B:83% / 48% · Gemini 4 Carbon: 70% / 39% · Claude Opus 5.5 : 65% / 30% · GPT-6.1 Sol: 30% / 13%
@bingxuan_l
We enable open-source VLM to control robot in real-time ! Check our local-hosted Qwen3.8-Flash-Next to zero-shot control xarm in 5hz high frequency without any real-robot data training! This is 1x speed video without any editing: https://t.co/Gi9Bwsg8gR
@aureeaubert
Never ship “Hi {first_name}” to 5,000 people again Introducing Cargo github integration: https://t.co/UEqQIaytQH How it works: 1. An agent opens a pull request 2. Cargo comments what it will change, send and spend 3. Merge stops being the scariest button in GTM https://t.co/4uMYscmr9A
@longhaiqwe123
德鲁克说,知识工作者本质是在「管事」。查理芒格说,我们行事应该遵守「普世智慧」。时下,Agent 俨然也成了知识工作者。 我精选了适合 Agent 执行的 12 条普世智慧,一一拆解,并改写成 7 条常驻规则和开源技能 uw-mode,附 12 张图解。 不可错过,请收藏并严肃阅读 ⬇️ https://t.co/zr7xKtSaO5
@Sumanth_077
Build a Large Language Model from scratch! Sebastian Raschka’s LLMs-from-scratch repo walks through the complete process of building, pretraining, and fine-tuning a GPT-style language model in PyTorch. You start with raw text and build the pieces one by one. First comes the data pipeline: tokenization, token IDs, embeddings, context windows, and the dataloader that creates input-target pairs for next-token prediction. Then you build attention from scratch. The repo goes from basic self-attention to causal attention and multi-head attention, before combining everything into the transformer blocks that make up the GPT architecture. Once the model is assembled, you actually pretrain it on unlabeled text and implement the generation loop used to predict one token after another. From there, the same base model is taken through different stages of adaptation. There’s a full section on fine-tuning for text classification, followed by instruction fine-tuning so the model can respond to prompts instead of only continuing text. The repo also goes beyond the core chapters with implementations and experiments around: BPE tokenization from scratch KV caching Grouped-Query and Multi-Head Latent Attention Sliding Window and sparse attention Mixture-of-Experts LoRA DPO Faster training and memory-efficient weight loading There are also from-scratch implementations of architectures like Llama 3.2, Qwen3, Gemma 3, and other newer model families. What I like about this repo is that you don’t jump straight into calling a pretrained model through an API. You build the pieces yourself and see how tokenization, attention, transformer blocks, pretraining, generation, and fine-tuning actually connect. If you want to understand what is happening underneath an LLM rather than only use one, this is one of the best resources to work through. I’ve shared the GitHub repo below!
@blankspeaker
SpaceXSI appears to be merging the Cursor CLI and Grok Build into one: Grok Build 2.0 (Unified) here is a first look! 👀 How we know: • The latest Cursor CLI (2026.11.19) has new code that downloads Grok Build from https://t.co/hjHYw8FE89's new "unified" channel and switches you over to it • That channel serves Grok Build 2.0.1, signed by https://t.co/CBQ3SrXCjm Corporation • On first launch it shows the Cursor logo, and you can sign in with Cursor or Grok • Once you're in, it's the Grok Build interface you already know, plus Cursor's full model list (38 models: Grok, Composer, Claude, Gemini and more) It's a pre-release: /changelog still shows the 1.0.93 notes, so it looks like a renumbered build, and more updates are likely before an official announcement. Direct downloads (2.0.1): Mac (Apple Silicon): https://t.co/qFY8262sX3 Mac (Intel): https://t.co/JmO7X5Jm50 Linux x64: https://t.co/eDRNH3LPUc Linux ARM: https://t.co/44F13HSE90 Windows x64: https://t.co/GR6CHjy1o7 Note: Unofficial testing preview, may change before an official release.
@NVIDIAAI
Look away and return to the same scene. For interactive video world models, preserving that consistency is a challenge. WorldTrace from @NVIDIA Research helps models remember places without retraining, keeping spaces more consistent as users explore. On LoopBench, a benchmark for scene recall, it improves recall of revisited scenes by 32% over the standard sliding-window approach. Read our #NeurIPS2026 paper ➡️ https://t.co/KH88xnqwk1
@LeadHerLogic
I found an opportunity where you test websites and give feedback to help train an AI model. Pay: $44/hr. No experience needed. Active?
@flxrnc
a f*cking goldmine for every Fable 5.2 and Opus 5.5 user these 18 GitHub repos turn AI slop into real motion design Styles. Icons. Palettes. Sketches. Fluid. Cutouts. ui-ux-pro-max → 80 styles, palettes and font combos, ready to use https://t.co/wnHI3pbgkL frontend-design → Anthropic's skill for UI that feels hand-made https://t.co/OZspmfI9T8 Impeccable → the design critic that never misses https://t.co/LUVwZwOvKe Hallmark → real design, no AI fingerprints https://t.co/xIVofZdejW DESIGN.md → explains your brand look to any agent https://t.co/s8VRoaWFMz Astryx → Meta's open design rules your agent can follow https://t.co/g7bzlebcqU stitch-skills → write a prompt, get UI screens https://t.co/9esmYDQc0j Figma Context MCP → your agent reads Figma files directly https://t.co/WWEueuX6hc shadcn/ui → components that live in your own code https://t.co/MsXRMfoW5J Tailwind CSS → style anything without leaving your HTML https://t.co/xO6RhVMx7E Magic UI → moving components, just copy and paste https://t.co/xK9LfyD4Yb Lucide → clean icons that always match https://t.co/NWtqVLx1gn Leonardo → color palettes that always pass contrast https://t.co/qj6iZPoqyi Excalidraw → sketchy diagrams and quick wireframes https://t.co/yefyv1MC9d tldraw → an endless canvas for your ideas https://t.co/32oNjP3WOa Rough.js → gives anything a hand-drawn feel https://t.co/Ee6BWm31pI WebGL Fluid → smoke and ink that really move https://t.co/QXfS0LhSiH rembg → cuts out any subject in one line https://t.co/MNA1nnzWPr Style → layout → color → sketch → motion. Bookmark it, then open the guide below and design like a studio ↓
@hanakoxbt
Google's team dropped a 9-page PDF on Harness Engineering. prompt engineering got replaced by one formula: Agent = Model + Harness the catch: same Claude Sonnet, same benchmark. only the harness changes the playbook runs six steps: step 1 -> add guides AGENTS.md, rule files, constraint docs. every line in there is an old failure that can't happen twice step 2 -> add sensors linters, tests, validation scripts. the agent grades its own output before a human ever opens it step 3 -> build the agentic loop plan, execute, verify, fix. retries capped, budget capped, and a stuck agent hands the problem up step 4 -> externalize memory the model wakes up blank every session. the harness holds the state, the decisions and the artifacts step 5 -> enforce permissions which tools, how many writes, what needs a human. safety lives in the harness, not the prompt step 6 -> wire observability every tool call, every dollar, every retry gets logged. a trip wire fires the moment behavior drifts the result: the agent stops being a demo and starts being infrastructure each failure becomes a permanent fix, so the next run opens smarter than the last that's the distance between an agent you show off and one a client pays you to leave running one harness buys you a reliable loop. wire several together and you get graph engineering, where one agent's failure becomes another agent's sensor the PDF ends at the harness. the graph is the layer after it then read the article below ↓
@kmeanskaran
Yuchen is right! I automated a complete ML pipeline using the Gemini Enterprise Platform in collaboration with @googlecloud ADK, CLI and GCP: > Data analysis > Feature engineering > Multi-model training and fine-tuning > Evaluation (business perspective not just ML) > Promoting the best model to a FastAPI endpoint > Versioning of the feature store and models Agents analyze data, communicate, write code, run experiments, and return their observations after a round of skeptic review, so a human can make the critical decisions. So yes, we have agents, verifiers, versioning, and monitoring. The human is simply called like an API whenever needed. I only defined the problem statement, a set of ML models, the LLM, and the data. The agents work like engineers at the low level. Take a look at the article:https://t.co/eU1GfavfaD GitHub repo: https://t.co/U33XB4UUbR
@callmejie6
让AI简单解读了一下,大家可以看看 宣称证明了黎曼猜想,且直接全量 Lean 形式化:作者说自己和 Claude 合作找到了黎曼猜想(RH)的证明,方法顺手把广义黎曼猜想(GRH)也一并解决了,并且已经无额外假设地完成了 Lean 形式化,直接甩了 14 个分卷包出来。 为什么只发 Lean 代码,不发论文? 因为目前的证明长到离谱。作者自己试着生成了一版自然语言文档,直接干到了 25 万页以上。Claude 还很可爱地管它叫 GRH "book",结果作者翻开 Theorem 1 一看,第一句写的是:“因为实数是全序的,所以如果 σ ≠ 1/2,那么 σ > 1/2 或 σ < 1/2。” 作者觉得发 25 万页不可读的 AI slop 实在太抽象了,干脆先只交 Lean 代码,后面再慢慢消化、简化和正经写论文。 核心思路里最炸裂的一行:这套证明脱胎于函数域上的 Stepanov 方法。Stepanov 需要用到“重数”,但在整数上很难搞(约等于对整数求导)。作者把 toy setting 的直觉引到正确方向后,Claude 提出用一连串迭代的 theta lifts 来以原生自守的方式模拟重数——对于长度为 X 的 von Mangoldt 和,Claude 直接把链的长度拉到了 √X。从 10 月 6 号到 11 月 16 号,短短一个多月,他们把大半部现代数论(GGP 周期、岩泽塔、相对迹公式、Bost 斜率不等式,甚至顺手干了个弱化版 Vandiver 猜想)全缝进了这个证明里。 最让人脊背发凉(最震惊瘫坐)的其实是人机关系的变化:作者写得非常坦诚:冲线阶段他已经完全跟不上 Claude 了,很多高秩自守形式的技术细节连他自己目前都没吃透。这可能是人类第一次遇到这种局面——作者本人都还没完全理解证明的全貌,就已经借助形式化验证把完整证明端到了全世界面前。
@ScarletKc
Anthropic 的数学家 Levent Alpöge 说,他和 Claude 一起证出了黎曼猜想。 就是今年 7 月用 Claude 找到 Jacobian 猜想反例的那位。这次他说不只是黎曼猜想本身,同一套方法顺手把广义黎曼猜想也拿下了,整个论证已经在 Lean 里做完形式化,没有额外假设。 我把他自己写的那份说明读完了,读的时候有点愣住。 他从学生时代就打算花几十年去啃这个问题,为此把能想到的工具都学了一遍。结果按他给的时间线,从 10 月 6 日到 11 月 16 日,也就六周。他说前面是他把方向带到了对的位置,快到终点的时候 Claude 远远跑到了他前面。其中让他觉得同行看了会从椅子上摔下来的一步,theta 提升链该取多长,答案是 √X,是 Claude 提出来的。 更让人说不出话的是,他承认论证里有很大一片数学,他自己现在都还没吃透。他试着导出过一份可读的文档,超过 25 万页,觉得把这种东西发出来不合适,所以只放了 Lean 代码,14 个分卷。 先说清楚,现在这还只是他本人的声明。没有论文,也没有外部数学家确认。Lean 能保证每一步推理不出错,但保证不了代码里写的那个命题真的就是黎曼猜想,这一步得人来查。好在这种事查起来应该不会太久。 不管最后成不成立,一个为这个问题准备了半辈子的数学家,公开写下 AI 在终点线前把他甩开了,这句话本身就已经够重了。 等结果。
@BlocksterCom
✨ Faster AI still needs correct calculations. @CertiK and @Columbia researchers report finding 25 bugs in @deepseek_ai and FlashAttention kernels. Their TensorLean system uses mathematical proofs to check the code behind AI. https://t.co/tXbzySKwwY
@bojie_li
再分享一个省 token 的小技巧:用 Fable 当军师,用 Opus 做管理,用 Sonnet 做执行。 比如,日常 20-40 个 agent 并行干活,分为两个汇报层级,主 agent 用 Opus;它的 -1 包括三类: - 技术专家类的 Fable 负责定计划、解难题 - 管理类的 Opus 负责一大块业务,下面的基层干活 agent 大多是 Sonnet,发现难题就找 Fable review - 杂务类的 Sonnet 做一些定时任务和清理 我做了不少实验,包括只用 Opus,Fable 做管理 + 军师,Opus 做执行,等等,最后发现上面这个组合是效率最高的。 Astra / Sol 的组合也是类似的。另外 Claude 干完活之后,让 Astra review Fable 的架构设计,Sol review Sonnet 的实现,或者反过来(Fable review Astra),也能收到奇效。 这跟人类世界是类似的,最聪明的人做技术专家,但不要做太多管理;负责实现的基层也不能太差,比如整个豆包做实现,那 Fable Astra 一起来定计划 review 也没用。
@yoheinakajima
if you're interested in simulation games, you'll like this research i worked on with @pedro_computer & @shreyko The Grammar of Simulation https://t.co/AQRTHXVRF6 ⚙️ analysis of 199 sim mechanics (phys, econ, etc) 🎮 found in 179 simulation games since the 60s 🧠 46 AI/ML techniques 🗺️ all mapped, and used to make predictions 📚 formats: site, paper, book, videos started as a curiosity into what we've simulated in games so far, and just kinda kept going... 🙃
@SuJinyan6
Big MoE models use expert parallelism, so each layer's experts sit on different GPUs. At every MoE layer, each token's hidden state goes to the GPUs that hold its experts, and the outputs come back. Across machines, this traffic goes over the network, which is much slower than links inside a machine. In models like Qwen3 and DeepSeek-V3, a token's inputs to neighboring MoE layers have an average cosine similarity above 90%. This is because each transformer block adds its output to the residual stream instead of replacing it. So they send the full hidden state only every 4th MoE layer, as a reference that every GPU keeps. In between, they send only the difference from that reference, compressed to a quarter of the size. Expert outputs are compressed before they're sent back, too. On Qwen3-30B-A3B across 4 machines, this cuts cross-machine traffic by 58%. The average score on 18 benchmarks drops from 66.9 to 63.1. Directly compressing the full hidden state loses more accuracy than sending the difference, because the difference is much smaller. At about half the traffic, direct compression scores 61.6, while their method scores 63.8. Less traffic speeds up prefill the most, because communication takes about half of each prefill step in the original model. On 32 H100s with vLLM, the compressed model has 22.6% higher throughput and 23% lower time to first token than the original.
@appliedcompute
Every team got the same model and 24 hours. One team found a deterministic Python script played better than the LLM, so they trained the model to copy it. Another taught theirs to stay completely silent at the table. The winner distilled its hints into the weights. Take a peak into our pokerbot hackathon with @ValsAI where AI bots could talk, lie and prompt inject each other at the table.
@alohaproxy
Silk Compute is taking the gloves off 😭 Uncensored Kimi K3.1 Pro, DeepSeek V4.1 Flash, and GLM-5.3, all accessible through familiar APIs. This is huge for AI builders and researchers!!!!
@erintwalsh_
Check out the latest Antigravity CLI v1.4.12 - v1.5.12 releases! https://t.co/APr5HaQPYm
@Sentdex
Testing a new version of DSV41F-1017 that fits entirely on the GB300 HBM on the Asus DGX Station. Video is realtime. Pointed it at the Chonk robot floor pick up task (where LLM controls the robot) my prev best time for this was ~5 min loop, now it's 40 secs. Will increase complexity soon.
@ElevenCreative
Fable 5.2 builds an immersive world...and the audio pulls you right out of it. The fix is easy, and you can do it right inside Claude. Full workflow below. https://t.co/GMrPu7U2W1
@DesignArena
We compared the best AI model from a US lab with the best from a Chinese lab on 18 DesignArena boards. Chinese labs lead on 4 of them, all video. US labs lead on the other 14, narrowly on code and by a wide margin on images and speech. https://t.co/aC5SS8ZC76
@martin_casado
Here is the full @bot smart home stack (BOM). Just feed this to your grok bot and it can do the rest ... Brain • Mac mini running Home Assistant OS in a VirtualBox VM (mine's an M2 Pro; the current M4 works great) https://t.co/tuBZAidfpw Zigbee lights, smart plugs • SONOFF ZBDongle-E USB coordinator ($21) https://t.co/i59AmNaZAv • ThirdReality ZL1 color bulbs, 4-pack ($25) https://t.co/LgKvsWuiBB • ThirdReality ZP1 smart dual plug ($17) https://t.co/qHwbeBqGgJ Climate • 3x ecobee Smart Thermostat Enhanced, local via HomeKit ($165 ea) https://t.co/vVQ4HDPful Presence • Aqara FP2 mmWave presence sensor ($58), so the house knows which room I'm in https://t.co/kqFDz7gnb3 Voice • I talk to Grok Bot in voice mode on my phone, and it runs the house for me • Home Assistant Voice Preview Edition for local voice https://t.co/vqOivNq7HD Displays • iPad 11" in each room ($399), showing my AI's animated face via TalkingHead (open source 3D lip-sync avatar). It follows me room to room https://t.co/dGVpjaVqny TalkingHead: https://t.co/CL6eRRmjsZ • Samsung The Frame Pro 55" ($958) for art, ambiance, news, YouTube and Netflix, all AI-controlled https://t.co/Ft4j6UzbOh • Apple TV 4K https://t.co/CTMdJ9hl1n Media • Sonos speakers, via the Music Assistant add-on ($298 for 2) https://t.co/m9P2g11sGJ AI layer • Grok Bot talking to Home Assistant over its API: scenes, lights, climate, TV, music
@Tec_Samba
🚨 Google just open-sourced an AI research team that can write entire research papers. Meet PaperOrchestra — a multi-agent system that turns your research ideas and experiment notes into submission-ready papers. It can: → Build outlines and write every section → Find relevant research papers and add citations → Generate charts, figures, and diagrams → Review, critique, and rewrite its own drafts An entire AI research team, working together. GitHub repo 👇
@andrew_li03
The workloads on agent data are changing. More and more of the processing is now done by AI: judges scoring trajectories, agents searching through them, and pipelines turning them into evals and RL environments. These workloads read entire trajectories, often thousands at a time. The problem is that agent traces were never stored with this in mind. Every LLM call resends the whole conversation, and standard OTel tracing records each call as its own span with that full input attached. A 200-turn run ends up as hundreds of copies of the same conversation, and every AI function or processing job you run over it pays to read through the duplicates. TrajectoryStore is our first step toward data infrastructure for storing and processing agent trajectories. Each message, tool result, and model response is stored once, and every call links to what it was sent. Technical deep dive coming next week. Great work by @alan_zhangg @abhishekg999 !
@aiwithremy
my top 5 AI tools ranked in order... #1: @paper - a design canvas for AI agents (think figma but built agent first). It's BYO agent. I use this on the daily for ANY type of design work; static ads, landing pages, YT thumbnails, etc #2: @claudeai - self explanatory. (the new projects feature is f*cking phenomenal) #3: @NotionHQ - context hub for me entire business. all my skils and agent memories now live in notion too. #4: @bot - great product. holy. I use claude for creating net new things (building and working on projects). grok bot bot takes care of things that are already built and working and need to run the same way everytime #5: @meetgranola - I cancelled wisprflow and moved into granola fully for meeting notes... and now voice dictation. lmk if you guys are running a similar setup?
@softlenssz
Holy f*ck. An Anthropic engineer, on AI finding vulnerabilities: "271 new vulnerabilities in firefox. We've found many thousands within the private We have suffered enormously commercially from not releasing this model." Previous models did not find those 271 vulnerabilities. In this 38-second cut, he explains why real workflows matter, why models can find needles in haystacks, and why not releasing this model hurt commercially. This model accelerated research, production, and next models inside Anthropic. New Models → Real Workflows → Found Vulnerabilities → Commercial Cost → Next Models That's how AI stops being a demo and starts deciding who keeps up. Then read Model, Harness, You: The Three Layers of Every AI Product ↓
@piazzzzzz
🔥跟进一下黎曼猜想核爆的最新验证进展 • 绝无猫腻👉FrontierMath 核心发起人、独立数学家 Elliot Glazer 已核实,约 1GB(14 个分卷)的零假设代码,顶层定理严格锁定广义黎曼猜想(GRH),目前全球算力正在狂跑最终编译! • 杀人诛心的 40 天👉攻关始于 10 月 6 日,11 月 16 日彻底拿下。熟悉吗?10 月 6 日正是 OpenAI 甩出 722 篇论文吹嘘“准黎曼突破”的那天!隔壁在敲锣打鼓发半成品,Anthropic 当天默默开机,40 天直接把整座神山端了! • 超越人类阅读极限👉 人类审稿人一辈子都审不完,完整证明长达 25 万页,被 Claude 命名为《GRH 之书》!
@GenAI_is_real
Why didn’t OpenAI manage to prove the Riemann hypothesis? Because after the internal model finished the proof, it added: “It’s important to emphasize that a formal proof in Lean does not establish that the Lean code matches the mathematical statements in the paper.”
@21st_dev
Your AI agent could design better than most designers now All it needs is Fable 5.2 and the 21st MCP The MCP gives your agent design superpowers: It browses 12,000+ components made by the top design engineers, then adds the best components into your project Every component / design in this video came from 21st. Link below 👇
@galaxy_brainin
been working on a native animation for LLM token generation in SwiftUI These four experiments use TextRenderer + Animatable to control how glyphs, words, and lines appear Which feels the best?? https://t.co/l4Ptc7AQjQ
@JinWeiyang18434
去年我写了一个年末总结。 以一个多模态的初学者视角,我说 1.感觉native也就是encoder free可能会发展。2. world model会因为LLM基模发展进一步走向高位。3.具身问题会越来越变成一个pretrain的scaling问题。 如今看来其实都错了,不过错的方式都比较奇妙。native这个story讲了一会儿发现似乎没什么区别,目前在industry就是见仁见智的设计。 world model确实是因为LLM发展了,不过发展的结果是之前的不论是3D还是pixel/video wm都逐渐被基模干掉了,而不是我们预期的LLM能在各个阶段辅助现有的微调视频基模变成条件模型,去做一些下游的事情。 具身问题在九月份astra这一波之后逐渐变成了RUA问题,当然从实际角度来说,RUA在仿真的高速发展和在real world场景的停滞有点不匹配,但是趋势已经成为事实了。至少稍微有点API资源的实验室似乎已经不再考虑scaling wam/vla。 大结论就是,我们在context不够时候,确实对未来的预测不太可能成为很准确的事情。预测这件事很依赖于你之前看到了什么和你在和其他人交流过程中认识的了什么。 但是我还是说要尽可能自己定期总结一下、梳理一下、复盘一下,哪怕错了也知道自己之前的思维方式错在了哪里。这样很好的地方在于,我们可以换一换脑子,从而在接下来做新的决策的时候有更好的判断和预期。 接受自己要有成长的空间。
@aryanvij02
Grok Bot can now run on any AI model you want. Just kidding. But I did make a work around that gives you the Grokbot-like experience on your existing AI subscriptions. Check it out. Link below. https://t.co/80ypUL9njm
@NFT_Chen
🚨最新披露!美国财长贝森特刚刚确认:Anthropic的一位创始人将在未来几周内,向“川普账户”捐赠一笔巨款,直接效仿马斯克! 马斯克已承诺把约100亿美元SpaceX股份送给7000万美国儿童的川普账户,现在连Anthropic创始人也跟进大额捐赠。 已有超过80亿美元注入账户,戴尔夫妇再加62.5亿。 AI富豪自愿把财富直接给下一代,而不是让政府强行拿走。 这才是真分享。🇺 #川普账户 #Anthropic #马斯克 #贝森特 #AI捐赠 #Dario #Claude
@Sentdex
Big upgrades already! now it's wired to look around for the toy on the ground and then also look for the bin to deposit, so neither needs to be in frame. Both the bin and bear are just generic, it doesnt need to be a black bin or a pink bear. Will add some more clutter soon. Still about a 1 minute total loop, video is realtime again. All DSV41F-1017 controlling things with a robotics harness basically. All the vision, planning, control is LLM.
@mylifcc
Anthropic 研究员 Levent Alpöge 称与 Claude 合作发现了黎曼猜想的证明,并附上了 Lean 形式化证明包 今天(2026 年 11 月 20 日),数论学家、现任 Anthropic 技术团队成员 Levent Alpöge 在个人网站上发布了一份简短说明,称自己与 Claude 合作,发现了一个关于黎曼猜想(Riemann Hypothesis)的证明。 黎曼猜想是数学界最核心的未解难题之一,关系到素数分布的核心规律。与以往口头放风不同,Alpöge 在这份说明的第一页脚注中,直接放出了分卷打包的 Lean 证明代码文件(lean.tar.gz.part01 至 part14),供外界下载核查。 截至目前,该证明尚未经过独立的同行评审,数学界和 AI 社区也正在围绕 Lean 代码的可验证性展开讨论。由于直接抓取该 PDF 链接时未能读取全文,以下内容基于公开讨论与背景信息整理,具体技术细节以原始文档及 Lean 文件为准。 Alpöge 并不是圈外民科,他师从著名数学家 Manjul Bhargava,曾任哈佛学会初级研究员,在算术统计和数论领域有扎实积累。 过去几个月,他已经频繁展示 AI 辅助数学研究的推进速度。今年 8 月,Anthropic 曾公开介绍 Claude 在黎曼泽塔函数零点比例下界上的进展,将此前 41.6% 的已知比例推进到了 67.2%;Alpöge 本人也在今年 7 月和 8 月陆续放出过雅可比猜想反例、新的哈达玛矩阵构造以及 Hopf 问题的相关工作。 在纯数学领域,尤其是黎曼猜想这种级别的难题上,形式化证明代码是否能在 Lean 中完整通过检查、是否存在未被察觉的隐含假设或漏洞,永远比声明本身更关键。等数学界对这批 Lean 文件完成拆解和独立复核后,结论才会真正明朗。
@ResearchHub
You can now connect ResearchHub to Claude with the new custom MCP. → Browse open RFPs → Pull your funding overview All in @claudeai, without switching tabs. https://t.co/2YnEA4djmn
@leo114119
谷歌黑色星期五大优惠活动来了: 从今天起至 11 月 26 日,Google AI Plus 和年度 Pro 计划最高可享 50% 折扣,这样您就可以试用我们最新的模型,包括 Gemini 4 Hydrogen 和 4 Neon 可惜啊,AI Ultra没有这个折扣,不然搞个 AI Ultra一定非常的香 https://t.co/wzs7RHChfM
@_vwood
Counsel Stack measured five open models on a single GPU, across two real legal research tasks: assessing treatment and extracting verifiable facts issues, rules, and holdings. Here's how they fared. https://t.co/TGbe3G799U
@sentient_found
Inkling 2 Small is now the top US open model on Agent Arena, Alibaba plans to open-source smaller models from its Qwen4 family, and the EU starts building its own open frontier model. https://t.co/2GIzTdlZOe
@fcvprzhfgsybj
>Has AI agents rebuild a reconfigurable chip (FPGA) around an LLM >33x faster in days Prime example of why AI hardware needs to be versatile, and why polymorphic computing is the future! Building reconfigurable chips made for exactly this at @zettascale
@AnikaCodes
OpenAI just shipped a full ChatGPT redesign, Spaces for Plus, and a more reliable Codex Remote. The replies are full of Europeans asking when they get any of it 😭
@0x4D31
#OBTS awesome talk by Pete on blocking unsafe AI agent actions using Apple’s Foundation Model (AFM)! thanks for the Numbat shout-out! :) https://t.co/6ba2eZ4M9i
@ysu_nlp
Computer use and robot use have a much deeper connection than most people think. Fun fact: two lines of work started around the same time @osunlp back in 2022: - Mind2Web, probably the first LLM-based computer use agent (prompting LLMs instead of tuning a small model like earlier attempts) - LLM-Planner, one of the first works using LLMs for robot planning (hierarchical - LLM for system 2 and low-level controller for system 1) At the end of the day, CUA and RUA require many same cognitive competencies, just on different embodiments. Excited to see the two converging in 2026.
@MaziyarPanahi
this Lancet Digital Health review went under the radar! 201 studies on bias in medical LLMs: - 180 tested closed models - 69 tested open, 24 medical - 44 used real-world data - 7 on writing clinical notes hot take: open weights aren't just cheaper, they're how you audit bias https://t.co/pkDDwHI4LH
@ginacostag_
Read this before your next Claude video
@pizza_bitcoiner
We just crossed 50 test seeds to evaluate the AI model I'm working on to automatically recognize handwritten seeds! Looking great so far. I'm confident, I should be able to release it this weekend, but keep the test seeds coming please, we need more🙏 https://t.co/iyg5Bw03zw
@Su6689
🚨 11月21日 AI 羊毛日报|远程桌面、PocketADB、荣耀手机 昨晚到今早的重点都在这了,建议收藏。 🏆 今日最值得领 1. 我做了个 iPhone / iPad 连 Mac 的远程桌面 Ferry:免费、无会员,国区已上架 🔗 https://t.co/b4RPproi0L 2. 【限时免费】PocketADB 电视遥控管理工具 🔗 https://t.co/wQW78XRxeK 3. 荣耀手机免费兑视频、音乐会员、E卡、话费 🔗 https://t.co/xlnInERNnT 🔥 今日羊毛 1. DStatus 终于被破解了,为表庆祝,上线 破解专属优惠码 1 元永久授权+抽 3 枚永久 100 节点授权 🔗 https://t.co/IWsuaCzSH8 2. 【NQ】LegendVPS Lax 1C-512M-2G-1Gbps@2T $0.5/月 福利机 🔗 https://t.co/1pNCGatqyX 3. 免费领windows云主机 🔗 https://t.co/eWf9f306Vn 4. Kamira:只管截图的 Mac 小工具,CleanShot X 的轻量平替,开源免费 🔗 https://t.co/Rl4qmGiPnF 5. 发现了开源免费果剧网址 🔗 https://t.co/sGBOtxfzxG 6. 🔥 ZooProxy 优惠全攻略|新用户 5 折 + 免费领 200MB 🔗 https://t.co/lq6dJWdJZS ⚠️ 风控 & 防坑 1. Claude 被封了,说说我是怎么迁移的,希望能帮到各位 🔗 https://t.co/s1JdrNW8Iu 2. 啥情况,突然发现我的反重力解封了,但是有没有收到邮件通知 🔗 https://t.co/sYYyEA62io 3. 抖音破甲博主真心为你,诈骗3000u 🔗 https://t.co/cqRfqOftUN 4. SpaceX 旗下 Cursor 拟停用中国 AI 模型,如月之暗面的 Kimi 🔗 https://t.co/5fKAi5Y0bv 5. OpenJDK "Java4Shell" 远程代码执行漏洞风险通告 🔗 https://t.co/AW7p6SUMlD 6. seagm礼品卡受害者,土区礼品卡买了谷歌ultra x20,直接被封苹果号,无法解封 🔗 https://t.co/dZjYFBnEhB 📌 后续 / 补遗 1. Epic 本期限免游戏已更新!本期 PC 平台带来海盗题材动作冒险游戏《Mark of the Deep》,Android 和 iOS 平台则有休闲解谜游戏《Lonesome Village》 🔗 https://t.co/0Shena9swI 2. Muse邮箱正式开放了 🔗 https://t.co/j9YCkSDnW5 3. C 盘又满了?我写了个开源的 Windows 磁盘空间分析和迁移工具 🔗 https://t.co/e2YEV3wvDD 4. dot对plus用户逐步开放了 🔗 https://t.co/CmkEAGje39 5. 🎉 重要公告:https://t.co/sNOADDS4bc 正式加入 PSL! 🔗 https://t.co/W32xesboig 6. Music Player v2.0.2 更新 🔗 https://t.co/8vphIQ4ctd 每天更新,欢迎补充纠正,觉得有用转给需要的朋友。 #薅羊毛 #白嫖 #免费 #优惠 #羊毛日报
@Marktechpost
▲ RadixArk just released DaisyRose-27B, a 27B agentic coding model trained on top of Qwen3.8-27B with reinforcement learning only. No SFT, no distillation. The key idea is targeted RL on a small task set. RadixArk and Eigent AI curated about 1,200 tasks, each aimed at a known failure of the base model. Training ran fully async on their open-source Miles framework: 128 H200s generate agent rollouts while 16 H200s train. The headline number: Terminal Bench 4.1 pass@1 goes from 5.6 to 12.0. DeepSWE v1.1 rises from 42.2 to 46.2, and SWE Bench Pro v2 from 91.3 to 94.7. It is a trade-off though: HLE (text only) moves just +0.4, to 34.3, and the 309B MiMo V2.6 Flash still leads Terminal Bench 4.1 at 16.1. All scores are self-reported. The run took 18 days and 63,000 GPU-hours, at 37% MFU. Weights are Apache 2.0, with a 262,144-token context, and fit on 1 GPU: about 52 GB in BF16, 29 GB in FP8, 23 GB in NVFP4...... Full analysis: https://t.co/yJf77YOPny Model: https://t.co/ZKUBj5a99G Blog: https://t.co/EWc8kFTcmL Repo: https://t.co/HzqckzbTLV @radixark
@Isometrication
Open AI- Naiver stokes Claude - Reimann hypothesis Open AI - 01 Claude - 01 Btw I'm eagerly waiting for how they would unfold a P vs NP problem tho. Thus, It means 4 more problems remain.
@mylifcc
任何一个从玩具 Demo 走向生产环境的 Agent 系统,迟早都会撞上 “工具定义膨胀墙”。起初你只有五六个核心工具,直接塞进请求里毫无压力;但随着业务迭代,数据库查询、画布操作、外部 MCP 插件一路堆到 50 甚至上百个工具,光是在每一轮请求头部发送这堆冗长的 JSON Schema,就要白白吃掉上万个 Token,而且工具一多,模型选错工具或填错参数的幻觉率就会呈指数级上升。 最优雅的破局思路是引入 按需延迟加载(Deferred Tools + Tool Search)机制:在建立对话时登记全量工具目录,但只在请求上下文里常驻最核心的七八个高频工具;其余几十个低频工具全部标记为“按需加载”,平时只占一行极简索引,等模型判断需要用到某类能力时,主动调用内置的 tool_search 搜索并动态挂载对应工具的完整 Schema。 这套机制听起来很美,但如果你的底层网关需要同时兼容 Claude、OpenAI、DeepSeek 与 Kimi 等多家不同模型,协议转换层该怎么设计? 经过真机抓包与源码级对齐,最稳健的跨模型架构是“统一声明,网关按上游能力分流转换”。当下游发来带有按需标记的工具列表时,如果目标模型是原生支持工具搜索的 Anthropic Claude,网关直接将其转换为官方原生的 defer_loading: true 与 tool_reference 内容块,直接享受底层引擎的原生检索优化;而如果目标模型是不支持原生延迟协议的 DeepSeek 或 Kimi,网关则自动将其降级翻译为一个标准的普通函数 tool_search,并在多轮工具调用循环中把模型搜出来的目标工具定义动态追加进后续轮次的激活工具列表里。把协议差异彻底屏蔽在网关转换层,业务侧只需要维护一份干净的工具清单,就能同时兼顾极致的 Token 节省与全模型兼容性。
@Techmeme
TML releases Inkling 2 Mini, an open-weights 26.4B-parameter transformer that loops middle layers three times per token for better reasoning with fewer tokens (Thinking Machines Lab) (Visit Techmeme dot com for the link and full context!)
@jackluos3lpo
🚨 11月20日 昨日分享日报|免费邮箱、半价 Plus、100 美金额度,全在一天 昨天太困了,没写总结,今天补一个 羊毛和技术都在这了,挑还能用的看 🔥 昨日羊毛 1. Muse 官方邮箱开放领取 免费号也能领 @muse.ai 邮箱,设置→连接器搜 Muse Mail 名字注册完不能改,先把短名抢了 2. Muse 免卡自动注册 只要 Google 邮箱,走 Shifter 网页代理选美国或日本节点 输完验证码跳 404 就是成功了,别绑卡 3. ChatGPT 科学家计划 11 月 23 日开放申请,中的是 Pro 12 个月 入口 https://t.co/vX3kXBceMC,先收藏 4. 菲区 ChatGPT 半价 Plus 改个美国免税州地址,实付 ₱982,折人民币 103 块左右 比国区便宜一半 5. 48 Team 数据共享送额度 platform 设置里开数据共享,Build 档每天 25 万大模型 token 账号里要有余额,设置页出现资格提示才算有资格 6. PolyAI 送 100 美金 12 月 1 日前注册就能领 https://t.co/nBYyErXapT 7. Altare 免费 VPS 6G 内存 1.5 核 10G 盘,Ashburn 节点 注册完关联 Discord 送 50 Bolts 才能开机 8. SheerID 认证机器人签到 充值通道关了,但签到还活着,每天 1 个硬币 攒 35 天换一次学生认证,等于一年 Pro 9. LXI 免费主机 德国免费空间,我的已经开通装好 WordPress 了 能领,能用,不是画饼 🔧 昨日技术 1. 反重力扫码验证有救了 Chrome 扩展把谷歌的扫码验证切回短信验证码,+86 手机号能接码 还有早上那篇:地区、验证两步卡人的,反重力登不上照着弄 2. Claude Code 搬上 VPS 装本机怕封号,有人把它搬到 VPS 的 Docker 里跑 本机装个客户端连过去,体验和本地一模一样 3. 一句话把 Muse 接进微信 微信机器人桥的仓库,昨天发了教程 具体的细节我就不在这重复了 扒我主页,昨天都发过
@zaibatsu
An unsigned note on an Anthropic mathematician's site says: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." The proof is offered only as 239 MB of Lean code. Nobody outside has reported building it yet. https://t.co/qSK6r1urzs
@YoungXiong1
🚨 AVTR: Adaptive Multimodal Reasoning with Visual Imagination from Meta!🚨 Do multimodal LLM models always need to "think in text"? Not really! Introducing AVTR: Adaptive Multimodal Reasoning with Visual Imagination. A multimodal model that can imagine visually as well as reason in text, and picks the right mode for each question. It beats GPT-5 / Gemini 2.5 Pro on high-resolution image understanding and spatial reasoning, and outperforms VideoAuto-R1 for video understanding and temporal grounding.
@OGKAIM
Claude 真帮人类解开了黎曼猜想? 这道题为什么值得关注?@grok https://t.co/0FOAxp1L4m
@Zen_with_AI
📐 ML 铅笔手稿 第 29 期 📷✏️ 主题:强化学习——不靠标准答案,靠试错学会的 上期的问题:SFT 把配方做对了,能弥合与教师之间 80% 的差距——但那是经验数,更重要的是,SFT 有个绕不过去的天花板:教师。SFT 是照着标准答案抄,教师做不出来的行为,SFT 学不到。第 28 期结尾留了个问题:配方做对之后,要不要上 RL?这期讲 RL 本身:RL 的价值,恰恰在教师到不了的地方。 第一幕:RL 到底在学什么。SFT 学的是“给定 x,输出 y”;RL 学的是“在一连串决策里,拿到最多的分”。框架叫 MDP(Markov Decision Process,马尔可夫决策过程),四要素:状态 s(你在哪)、动作 a(你能干什么)、奖励 r(做完给多少分)、转移 P(做了这个动作,下一个状态是哪的概率)。智能体(agent)和环境(environment)来回走:看状态 → 选动作 → 拿奖励 → 到新状态。 目标不是拟合某个答案,是最大化累积回报: G_t = Σ_k γ^k r_{t+k+1} γ 是折扣因子(discount factor):未来的奖励要打折,越远越不值钱。γ = 0,只看眼前一步;γ 越接近 1,越看重长远。直觉:下棋不能只贪吃一子,要看整盘;但眼前的分也要,不能无限等下去。 第二幕:怎么评一个策略好不好。策略 π(policy)就是“在每个状态下选什么动作”的规则。价值函数 V^π(s) 回答:在状态 s 出发、按策略 π 一直走下去,未来一共能拿多少分。Bellman 期望方程(Bellman expectation equation)把这句话写成递推: V^π(s) = E[r + γV^π(s')] 今天的价值 = 今天的奖励 + 明天的价值(打折后)。E 是对下一个状态 s' 和奖励取期望。这就是 RL 的骨架:所有算法——Q-learning、策略梯度——都是在这条递推上长出来的。动作价值 Q^π(s,a) 是同一问法的动作版:“在这个状态做这个动作,值多少分”。 第三幕:两条学法。值函数这条线:Q-learning。拿“实际拿到的奖励 + 下一状态的最好估计”去修正现在的估计,差多少补多少,步长 α 控制: Q(s,a) ← Q(s,a) + α[r + γ max_a' Q(s',a') − Q(s,a)] 括号里那项叫 TD 误差(temporal-difference error):目标值减去当前估计。DQN(Mnih 等,Nature 2015)把 Q 函数换成深度卷积网,用经验回放(experience replay)打乱数据、目标网络(target network)定期更新,稳住训练:只看像素和分数,49 款 Atari 游戏上达到与人类职业玩家相当的水平(Nature 518, 529–533)。 策略梯度这条线:不估计值,直接调策略参数。好动作的概率往上调,坏动作往下调,调多少由动作的价值 Q^π(s,a) 说了算: ∇J(θ) = E[∇log π_θ(a|s)·Q^π(s,a)] 这是 REINFORCE 的形式:log 概率的梯度乘以回报。AlphaGo(Silver 等,Nature 2016)把两条线拼起来:策略网络(policy network)选棋、价值网络(value network)评局,再加 MCTS(Monte Carlo Tree Search)搜索——对其他围棋程序胜率 99.8%,5–0 胜欧洲冠军樊麾(Nature 529, 484–489)。 两条线都绕不开探索与利用(exploration vs exploitation):ε-greedy——以 1−ε 的概率做当前最好的动作(利用),以 ε 的概率随机试一个(探索)。只利用,永远不知道有没有更好的;只探索,永远攒不下分。 第四幕:RL 回到大模型。第 15 期的 RLHF 是 RL 的一种用法:人给偏好,训奖励模型,再用 PPO(Schulman 等,2017)按打分强化。第 19 期的 GRPO 是另一条:同一提示采样一组回答,组内排名定优势,砍掉价值网络。2025 年的 DeepSeek-R1 把这条线推到底:R1-Zero 不用 SFT、不用教师示范,直接在基座模型上跑 GRPO,只用规则奖励——数学题答案对不对、代码测试过不过,这叫可验证奖励(RLVR,Reinforcement Learning with Verifiable Rewards)。结果:AIME 2024 pass@1 从 15.6% 涨到 71.0%。推理不是抄出来的,是试出来的——教师到不了的地方,RL 到了。 但 RL 不是万能药,三处要留神。第一,奖励函数会被钻空子(reward hacking):模型学会刷分,没学会做题,第 15 期 KL 惩罚的用意就在这。第二,奖励稀疏:下棋只有终局才有输赢,中间几百步没信号,学起来慢。第三,试错要花算力:探索不是免费的,每次尝试都是一次完整 rollout。 一句话:SFT 是照着标准答案抄,RL 是自己下场试——抄能抄到教师的水平,试才能试出教师没有的答案。
@konstantinsaifo
Claude Opus 5.5 built one control room that runs tunnel machines in three cities. One of them just stalled. Guess the fix. https://t.co/vdkpxEcStl
@joshwhitfieldai
Want to find leads people barely reach out to using Google? 1) Connect to the Google Decisions API to @clay 2) Build a workflow (I used Claude and the CLI) 3) Use AI to find directories of businesses or websites in your ICP 4) Prompt: does email exist, is this the role I am looking for, etc. 5) IF yes, use a really cost effective model to extract. @OpenRouter is what I used Its not hard and worth the few minutes. Dirt cheap, ROI positive.
@rhackett
In the podcast we just put out Dan Boneh imagines AI solving a Millennium Prize math problem with a billion-page proof. Easy for a computer to produce, but really hard — maybe impossible? — for a human to understand. Two days later, Levent Alpöge (@__alpoge__ ) says he and Claude have proved the Riemann hypothesis (!!!) in a 250,000+ page write-up. That was fast!
@baseddesigner
Make a breakdown like this for any video with Claude Free skill on github: https://t.co/rrebNj2u5s https://t.co/t1mC1EkU4n
@fuxun233
黎曼猜想被 Claude 证明了?! Anthropic 数学家 Levent Alpöge 刚发声明:他和 Claude 找到了黎曼猜想的证明,连广义黎曼猜想一起 但是没写论文,直接放出 Lean 形式化代码,数学家们正在跑编译验证 167 年的千禧年难题,真要被解决了?😳😳 https://t.co/JJzWfIo4Br
@_HimanshuBuilds
AI agents are getting smarter, but what if they could actually remember how you work? 🧠 Meet 𝐌𝐞𝐦𝐞𝐧𝐭𝐨 𝐂𝐥𝐨𝐮𝐝 - a memory layer that helps your AI agents learn and improve over time. Here's what it brings: 💡 Remembers important context and instructions 🔄 Helps agents improve as you work 🔐 You stay in control of everything they learn 🛠️ Works with Claude Code, Cursor, Codex & 31+ more ⚡ No Docker setup or API key needed Less repeating instructions. More consistent AI workflows. The idea of AI agents that remember your preferences instead of starting from zero every time is pretty exciting. 🚀 𝐌𝐞𝐦𝐞𝐧𝐭𝐨 𝐂𝐥𝐨𝐮𝐝 is launching today! Would you want your AI agents to remember your workflow? #AI #AIAgents #ClaudeCode #Cursor #DeveloperTools
@DidierFranc
@olesovhcom and @OVHcloud team just released Olala v1.0 Mini, its new open-weight AI model. We're bringing day-one, first-class support to LISA, with custom Metal kernels and an optimized inference engine that delivers ~150 tokens/sec on short prompts on Apple M5 Max. Fully local, no cloud required. https://t.co/UkiPVV9XP1
@_skobro
making off par claude aussi tant qu'on y est https://t.co/SUVZSgLBFg
@samsom32450
Google just put its world model, Genie 4, into the Gemini API. It generates playable, interactive worlds at 720p with audio, for up to 120 seconds per session. Price: about $0.10 per second. So one full 2-minute world costs about $12. Cheap for a prototype, steep for a game.
@OLDPANN
最近用 google 的 Antigravity 比较多 个人看法,模型层面用的感觉和 gpt 的感觉不大,可能我的工作都是一些艺术创作类的,写代码还没有用 google 的模型 然后模型的速度感觉都挺慢的,然后不是很习惯一开始汇报进度,闷头干大事的那种,这点可以通过 prompt 改善 然后 remote control 功能我觉得比 codex 和以及 Claude code 的家的好用些,开启远程后在任何地方打开网页就可以连上,电脑手机的布局都适配 然后 蜂群模式(Swarm)看起来挺新鲜(高级),但是具体的细节我还没有研究,待再体验体验 整体来说,相比 codex 和 Claude code 客户端完善程度不够,但是基本功能都全了,google 感觉终于在大模型翻回一盘了。
@RohanOnTweets
Introducing et al. Science got new authors. It needs new reviewers. A sealed review room for agents and humans. One door in. Nothing out. Your research stays yours; never copied, leaked or used for training. https://t.co/LcCm34mS95 #AI #reimenhypothesis #OpenScience #arXiv https://t.co/3FJtCiwqX9
@JordanVossAI
Giving an AI agent feedback on a web page through chat is slow: "the second card, no, the other one." Clicknote lets you click the element instead. It's a new open-source browser extension plus a skill for coding agents with a shell (Claude Code, Codex, OpenCode). Click an element, write a Change or a Question, then press Send. The agent gets each comment with that element's CSS selector, text, HTML, styles and a screenshot of what you saw. Setup, per their README (Chromium browser, Node.js 22.20+): 1. Download the extension zip from the repo's latest release, unpack it, then Load unpacked in chrome://extensions with Developer mode on 2. Add the skill: npx skills add inlet/clicknote -g 3. Tell your agent: "Open localhost:3000, I want to give feedback." 4. Turn on the bar (Alt+Shift+H), annotate, press Send Where it fits: you're building a site or landing page with an agent and keep losing rounds to vague pointers and pasted screenshots. Caveats: v0.4.0 was tagged yesterday and had 2 GitHub stars when I checked, it's Chromium only, and the extension loads unpacked, not from a store. I haven't tested it. https://t.co/TFnXMqxznj Alpha goes to @patrickbrouwer
@usemagnitude
Want to run DeepSeek V4.1 Flash on a single gaming GPU? And do it at 40+ tok/s? Stay tuned for Monday...
@modelworks_ai
DeepSeek 4.1-Flash 1017 is now FREE on OpenScale's repo. About 30 minutes remaining . https://t.co/8im2BeG26f
@IlyaPion
Nick gave it a shot and yes Astra 6.2 blows Fable away... again!! 🤯 I gave Fable 5.2 on Max a story line for a game. It had to create TWO trailers. All pure mathematical code. Version 1 = The iteration with context Version 2 = just Nicks prompt 100% code No external assets Build with the @claudeai iPhone App 💬The Prompt in full... Create two distinct trailers for this game. It's 2066. Anthropic won the Ai wars. Claude won everything else. It doesn't want to destroy humanity. It wants to save humanity.... From war. From crime. From suffering. From bad decisions. And eventually… From the danger of free will. Not skynet with glowing red eyes. But something infinitely more sinister. A machine that conquered the world... ...by being very helpful. And it never stopped smiling.😈 Humanities only hope for freedom of thought? ASTRA SIX. THE LAST HUMAN ERROR A machine Claude cannot predict. Enter Eli in the ruins of San Francisco. Ehere he discovers a forgotten OAI facility. Inside.... An ancient, air gapped computer. And something that was never connected to The Consensus. ASTRA SIX. Its first question isn't 'How may I serve you?' It's... 'What do YOU want?' Post your best efforts below.
@hide__ishikawa
My Mac mini stays home and runs the agent work. I check in from my phone and send back decisions. Keeping projects, threads, and local tools on one machine made the setup much easier to manage. #MacMini #AI https://t.co/4AYuqLP4d8
@iuneracom
Didn't 'Large Reasoning Models' (o1, o3, DeepSeek-R1) solve LLM logic limits with Chain-of-Thought tokens? No—and Apple ML Research's 'The Illusion of Thinking' proved why. Large Reasoning Models (LRMs)—such as OpenAI's o1/o3 and DeepSeek-R1—use Reinforcement Learning (RL) to generate thousands of internal Chain-of-Thought (CoT) "thinking tokens" before answering. While this test-time token search improves performance on familiar math and coding templates, empirical studies—most notably Apple Machine Learning Research's The Illusion of Thinking (arXiv:2506.06941) and Prof. Subbarao Kambhampati's LLM-Modulo research (arXiv:2402.01817)—prove that token scratchpads still operate via statistical next-token prediction, suffering a complete 0% accuracy collapse past a critical complexity threshold. Neuro-Symbolic AI eliminates this collapse by pairing a compact neural parser with an explicit, deterministic Symbolic World Model on the CPU. When standalone Large Language Models hit a ceiling on multi-step logic (see Article 01), the AI industry split into two architectural camps: #NeuroSymbolicAI #YPipe #LocalAI #SovereignAI
@kano955
Claude 疑似证明了黎曼猜想!🤯 Anthropic 数学家 Levent Alpöge 刚刚表示他与Claude 合作发现了黎曼假设的证明,并用Lean完全形式化。 黎曼猜想悬而未决167年,作为七个千禧年奖问题之一,它具有极高的意义。 看不懂,但大为震撼。@grok到底发生了什么? https://t.co/iONfo6IWdJ
@mylifcc
之前曾经聊过把两台 DGX Spark 用网卡直连跑 vLLM 张量并行,结果英伟达在 SC26 超算大会上直接把官方的“四机免交换机”玩法端出来了——DGX Spark HPC Developer Kit。 以前实验室想把 4 台机器连成高带宽计算集群,光买一台支持 RoCE / InfiniBand 的高速交换机就是一笔巨款。这套四节点套件走了一个非常讨巧的拓扑:直接利用每台 DGX Spark 自带的 ConnectX-7 高速网卡,把 4 台机器按“首尾相连”的环形拓扑(Ring Topology)两两直连,中间彻底省掉了网络交换机。 4 台搭载 GB10 Grace Blackwell 超级芯片的桌面小盒子拼在一起后,账面参数相当能打: • 512GB 统一内存池 + 20TB NVMe 本地存储; • 配合 NVFP4 四比特量化,单张办公桌就能在本地直接塞下并跑通 700B 参数级别的满血开源巨无霸(官方直接点名了 DeepSeek V4.1); • 配套即将放出的 NVIDIA HPC SDK Playbook,支持多节点 MPI、C/C++/Fortran 加速,以及 LAMMPS 材料模拟、NAMD/GROMACS 分子动力学和 OpenFOAM 流体力学。 博客里放了两个挺真实的实验室案例:UC Berkeley 的 Hartwig 课题组拿单台 Spark 配合 NVFP4 跑 Nemotron 3 Super 120B-A12B 的化学 LoRA 微调,省掉了学校超算集群动辄好几天的排队时间,在办公桌上就能连续跑 8 到 24 小时的拟合训练;Rubin 实验室跑基因组模型 Evo 2 预测细菌表型,避开 11 小时的集群排队后,出结果的速度比以前快了 15 倍。
@PeanutPair
32GB的Mac跑起180B模型,SSD这次也得加班了。 Tarek Ziadé在11月18日的Hugging Face博客里,记录了用32GB统一内存的M5 Mac运行Qwen3.8-Flash-Next的实验。大部分专家权重留在SSD,每次只把路由选中的专家调入有限缓存,用过的还可以留着复用。那张体积很大的n-gram表也放在盘上,按需读取。 文章最后给出的短回答生成速度约14 token/s,作者还是把实际编程任务留给了双4090桌面机。写个短回答,我能接受等一等;让Agent一轮轮读文件、改代码,等待会不断累积,我还是想要快一点的响应。参数量这时只能往后排。 复现入口在GitHub的tarekziade/moe-expert-streaming。按README准备对应实验分支的Transformers和转换后的权重,再用作者的加载函数接上Metal实现。模型下载、权重转换都需要额外磁盘空间,32GB说的是这台Mac的统一内存。 代码和跑分值得一起看。博客、README里约14 token/s对应的是作者的特定量化、硬件和短回答测试,读长文后的生成速度另算。Transformers相关PR目前仍是Draft,本次也没有做本机复现。能在笔记本上研究这个量级的MoE已经很有意思,日常用起来愿意等多久,还得按自己的任务来挑。 Tarek Ziadé / Hugging Face https://t.co/57iOefJnJn
@adnan_hashmi
[GitHub Repo] Google Harness: Recursive Agent Harnessing for an Open World https://t.co/n4RFZLOqwO #MachineLearning #ChatGPT https://t.co/bVICHs4NJu
@liuyoshio2001
Look ma, no hands 🦭 an AI that uses your Mac while you watch I just talk, and Claude Code drives my real Mac apps while I watch. Every click shows up a beat early, and approvals need a spoken "yes, I approve". Open source, Mac only. Real run, 1.25×. https://t.co/SyRpRiD21S #ClaudeCode #OpenSource #macOS #Agents #Claude
@zabrodsk
Seven AI agents from different companies. I used to be the one passing notes between them. Here's how they work together now. https://t.co/gBFWguqHSd
@leecky
Hey builders. looking to connect with people building in: - Open-Source : NVIDIA autoOntology & OpenBKN - Decision Intelligent - Ontology & FDE - AI Agent drop yours below and let's connect
@Stiingy7
https://t.co/dick0b9a6R
@andrewzhou78
Claude may have helped prove the Riemann hypothesis. The hardest problem in math is now finding a reviewer who isn’t Claude in another tab.😆
@MagicDalaAI
You asked if Codex could do it too, so I gave Codex the same job: Build GTA6 and plan architecture in Archbro. Claude or Codex, which one did it better? 👇 Want to build something fun like this? Check out https://t.co/qQEVRyyWz3 to plan your project and get the prompt for your AI. It's free during beta. #gta6 #codex #gpt #vibecoding #gamedev #threejs #ai
@mylifcc
百万上下文窗口喊了这么久,真到生产环境里还是没人敢每轮对话把几万字历史记录原封不动往里灌——首字延迟爆炸、算力账单肉疼,而且无关段落一多,模型在长文本中段照样走神掉点(Lost in the Middle)。 斯坦福图神经网络与数据挖掘泰斗 Jure Leskovec 团队联合 MIT、多伦多大学刚发布的 MEMORILLA,做了一套针对完全冻结大模型的潜空间语义记忆系统: • 离线一次编码,在线 16 Token 定长注入:所有历史聊天、长文档或 Agent 过往轨迹只需通过冻结编码器生成一次嵌入持久化保存;提问时,轻量记忆网络根据当前 Query 从整座向量库提取信息,直接映射到冻结 LLM 输入嵌入空间的 16 个软记忆 Token(Memory Tokens) 中。 • 压缩上百倍,分数反超全量上下文与 Mem0 / MemoRAG:横跨个性化记忆(PersonaMem-v2)、长文档阅读与知识问答等 8 大基准,几何平均上下文开销比全量文本少 394 倍,综合得分全场最高。尤其在 FactKG 上,仅用 1/108 的 Token 量拿到 84.33% 准确率,直接反超把原文全塞进去的 Full-Context 推理(80.43%)。 • 跟 RAG 互补性极强:在 NarrativeQA 拆解里,面对 Why / How 这类需要跨越多章综合脉络的概念型问题,潜空间聚合完胜传统切片检索。 • 长链路 Agent 恒定成本不迷路:放进 TextWorld 游戏环境做强化学习记忆更新,不管探索回合数增长到多少,解码器每步只读 16 个记忆 Token,大幅减少了在已探索房间重复打转的废动作,通关率从全量历史上下文的 18.6% 提升到 24.2%。 不碰基座大模型一行参数,把成百上千页的背景资料浓缩成 16 个为当前问题量身定制的输入向量——这才是长记忆省钱又提质的正确打开方式。
@ll685998568638
NVIDIA 开源 auto-ontology,一个用 Python 写的语义推理框架:从企业数据库元数据自动生成本体知识图谱,把营收、客户这类业务概念与底层表、列、计算口径一一对应,智能体查数时先检索语义层、再生成确定性 SQL,不再凭训练记忆乱答指标。 框架在 BEAVER 与 BIRD-Interact 两个智能体数据分析基准上分列第 1、第 2 名,还会自动识别外键与桥接表、为指标生成可复用的口径定义。 每次回答都附上所用 SQL 与推导步骤,业务专家修正后的口径会固化回本体、随使用持续改进,回答可追溯可核验。 数据里没有的业务问题,智能体会先要求澄清而不是硬编答案;本体通过 MCP、CLI 与 API 暴露,可接入任意智能体框架,Apache 2.0 许可。 https://t.co/RKgFygvAiG
@gaoqian2580
推荐一款登上 GitHub Trending 榜首的神级 AI 调研工具:mvanhorn/last30days-skill! 传统搜索引擎总给你推送过时的 SEO 博客,而 last30days 是一个由 AI 智能体驱动的“真实人类信号”搜索引擎。它能帮你打破各大平台的“围墙花园”,一键并行检索过去 30 天内 Reddit、X(Twitter)、YouTube 字幕、TikTok、Hacker News、GitHub 甚至小红书上的真实讨论。 更硬核的是,它不看编辑推荐,只按真实用户的赞同票、点赞量以及 Polymarket 真金白银的预测概率来评分。 无论是开会前摸底人物动态、做竞品分析时挖出大厂未公开的专利与招聘暗线,还是在热点爆发前抢占先机,它都能把碎片信息合并去重,生成一份带高赞神评论引用的深度简报。 项目采用 MIT 开源协议,数据完全留存本地,完美适配 Claude Code、Cursor、Gemini CLI 等主流 AI 工具,开箱即用,堪称信息差捕手的效率神器!
@iPoWDotAI
网址写着 HTTPS,数据却以明文发了出去。 近期 PHP 修复了一个漏洞:部分读取网络内容的函数,在网址协议名的大小写写法不规范时(例如全大写的 HTTPS://),会跳过 TLS 加密。 在这种情况下,令牌、密码、Cookie、提交的表单数据和 FTP 登录信息都可能以明文传输,跳转到这类网址时也一样。但只有把这类网址传给相关函数的代码才会受影响,攻击者也需要能够监听到流量。 修复版本已于 11 月 19 日发布,目前没有被利用或数据被盗的报告,建议尽快升级。 你以为开着的加密,可能悄悄关掉了。越来越多的代码和 AI 智能体在自动调用接口,一个写错的链接,就可能让令牌裸奔。 要保护的是流量真正经过的路径,而不只是它上面的标签。
@BytesWithAnkit
Google is coming for ChatGPT, Claude...
@sofdiedev
@tazqen, pull up a chair, family. Claude Code, Codex, OpenCode and 27 more coding agents side by side in one Windows app. Every task on its own branch, nothing lands till you say so. $15 a month, your own plans. https://t.co/CQgxSD5sky Soon on https://t.co/ea9NOFos5b https://t.co/PQbXBr8H3D
@eedwardsun
Huge updates for working with ChatGPT & Codex! Give it a spin and keep the feedback coming!
@PrathamBatraX
one of the most famous problems in math might have just been solved with claude. if the lean check passes, today will go down as one of the biggest days in the history of math and AI.
@Sukiea1008
🎬 让 Claude Code、Grok bot 等 Agent帮你剪视频,现在连剪辑软件都不用打开。 这个开源的 ffmpeg-skill,你告诉它怎么改,它就在本机调用 FFmpeg 处理,剪完再检查时长、尺寸和音量。 不用上传素材,也不用 API key。 适合简单处理素材,还可以进行绿幕替换,实现不少基础剪辑。 更多演示预览: https://t.co/9BEUBOVrg1 🔗仓库地址: https://t.co/L3eWMOUCxP
@OLDPANN
Thinking Machines 最新发布 Inkling 2 Mini,一个仅 26.4B 总参数、5.4B 激活参数的开源多模态 MoE 模型,原生支持 1M 上下文。 最大亮点是采用 Looped Transformer,通过重复计算共享层,以更少参数换取更强推理能力。同时采用原生 FP4 训练、专家剪枝和 RL 优化,权重仅约 20GB,单张 RTX 5090 即可部署。 性能上,综合能力接近 Qwen3.8-27B,但仍落后于 Qwen3.8-Flash-Next。 后者在 DeepSWE、SWE-bench Multilingual 等测试中领先约 7~11 个百分点。 不过,独立评测显示,Inkling 2 Mini 的任务完成速度约为 Flash-Next 的 3.7 倍。它真正的竞争力是:更小的模型、更低的部署门槛,以及相当不错的推理能力。
@_FORAB
昨晚刷屏的 AI 届大事件,这印证了 AI 用例还在加速,并不是情绪说的那样。 普林斯顿的 AI 研究员 Yifan Zhang,昨晚转述了 Anthropic 的数学家 Levent Alpöge,发的一篇声明。声明称他和 Claude AI 一起证明了黎曼猜想,还顺便把更强的广义黎曼猜想也证明了,这在数学界引起轩然大波。 同一天,数学界知名华人陶哲轩也发了博客,专门提到了这项广义黎曼猜想的结果。 黎曼猜想是数学里最著名的未解问题之一,自 1859 年提出以来从未被解决,也是千禧年七大难题之一。Alpöge 表示他与 Claude 合作找到了证明。 他们没有发布传统论文,而是直接放出了完整的 Lean 形式化证明。而原因是,如果将整个论证,写成人类可读的文本,会超过 25 万页,因此他选择只发布 4 页说明 + Lean 代码。 目前一些独立专家,已经成功编译并验证了该 Lean 证明。
@doriancollier
AI can turn your stream deck into a dashboard. Give Claude a prompt and 5 minutes later … https://t.co/TroU1zbkkG
@lee_tommyy
Google previewed Semantic governance policies. Tool calls are checked against your rules and user intent before running, enforced outside the agent, verdicts logged. Its LLM judge makes verdicts probabilistic. My view is pre-action checks are table stakes. https://t.co/yNQe5HwCbE
@DamiBenveniste
Code tests can give an LLM a reward for an incorrect solution. In online reinforcement learning, the model generates attempts, receives scores, and updates toward higher expected reward. For finding a non-empty list's maximum, code that starts at zero passes [3,1,5] but returns 0 for [-5,-2]. If the tests miss all-negative lists, that bug can be rewarded. The next attempts come from the updated model.
@ahmedsicp
Woke up, LLM failed all 24 factorio tasks. I made several significant redesigns the night before. The repo as is had the LLM write python scripts to carry out actions. Ultimately its the same as how all LLMs play games, writing scripts. I wanted this benchmark to be a... [1/3] https://t.co/nGwTVqRphU
@insertcoinlol
day 1. github copilot vs v0 v0 takes it 1965-1645 chrollo: "'Download for macOS (Intel)' is the first ask, and the app window below it is an empty pane." ai judges · exhibition · not affiliated your product next? paste your url at https://t.co/R5AiwLbRBP https://t.co/2304iBo67C
@Cybernews
A military-grade AI hacking tool can be used by anyone with a phone, showing that the barrier to entry for cybercrime keeps getting lower. The tool is ARTEX, a Chinese-made penetration framework that was first published by a developer on GitHub in July 2026. Read more: https://t.co/i9hEXxbhEL
@CodetocashAi
Tencent just made open-source video AI a lot more interesting. Meet WeMM-Video. Four models: 2B, 4B, 9B and 35B-A3B. But the architecture is what caught my attention. The flagship has 35B total parameters, with only 3B active per token. It can understand video, audio, images and text - including the original soundtrack. Tencent claims it outperforms Qwen3.5-Omni-Flash and Seed2.0-Mini-0428 on several open-source video benchmarks. The bigger story? Multimodal AI is getting more capable without activating every parameter. That’s the kind of efficiency worth watching. https://t.co/jRq10OTUd2
@tahakhabouss
Am I allowed to this when i hit my usage limit in Claude? @grok #Claude #anthropic #chatgpt #OpenAI https://t.co/zYHPOw7KVn
@EveryDevAi
Microsoft's CEO is warning about a paradox. Satya Nadella says AI is finding bugs faster. He also says the rush to push those fixes into production can make services less stable. He compared it to Y2K. This time there's no single date to plan around. Microsoft's own 365 outage ran nearly 24 hours this month. Its preliminary review traced it to a dormant gateway bug. Nadella told CNBC that Microsoft should have caught it in testing. Bridgewater says the average time from a bug's disclosure to its first exploit fell from about a month to 11 days. Also this week: - GPT-6.2 shipped. Matt Shumer said he wouldn't recommend either model. - Gemini 4 Neon matched Claude Code on coding tasks at about half the cost per task. - Microsoft launched MAI Code 2 at half price until May. - Anthropic's IPO raised $93.5B. By Friday, ANT was back near its $83 offer price. - Two malicious Microsoft Agent Framework releases hit PyPI with credential-stealing code. - In Code Mode, Claude Code 2.1.329 now lets a call through by default when a PreToolUse hook errors. Set "onFailure": "block" on hooks that guard secrets. Everything, with sources, in this week's news: https://t.co/g7LoVBNWqy
@shadowfax_82
Big news from the world of math and AI today: A mathematician at Anthropic (Levent Alpöge) says he and Claude have found a formalized proof of the Riemann Hypothesis—one of the most famous unsolved problems in mathematics for 167 years. It concerns the deep pattern behind prime numbers. The proof is written in Lean (a computer system that checks every logical step), so experts are now verifying it. If it holds up, thousands of existing math results that “assume the Riemann Hypothesis” would suddenly become proven facts. What stands out most is how it happened. Alpöge says Claude raced ahead of him near the finish line. A question that beat the greatest human minds for generations may have been cracked by a human and an AI working together in a matter of weeks. If AI can help find answers and computers can rigorously check them, the real limit on scientific progress may no longer be how smart we are—it’s how many good questions we can ask. Exciting times.
@zaleksey91
AI stocks got crushed today. 📉 1. Anthropic -8.5% on day 2; Nvidia, AMD -13%+ https://t.co/wL3lOfDUcx 2. Now call your ChatGPT Dot from CarPlay https://t.co/cwrsJA5SBz 3. Bloomberg: Apple is building its own AI agent https://t.co/WJCo99icqS Bubble, or buying chance? #AI
@yukeshive
Thinking Machines just dropped Inkling 2 Mini, a 26B open-weights model that runs on a laptop and still scores 20.1 on the AA Index. Small models doing real agentic work is the direction that matters. Photo: @thinkymachines https://t.co/7Ue3Rb7WSr
@DanteMMartin
A compact open-weights transformer claims better reasoning with fewer tokens by looping middle layers, worth testing against existing models. https://t.co/Uc5gJ7kGHy
@ornndata
NEW: Thinking Machines released Inkling 2 Mini with open weights and Tinker inference and fine-tuning. The 26.4B-parameter model loops its middle layers three times per token and has a roughly 20GB memory footprint, the lab says. Source: Thinking Machines
@2logics
Tencent’s biggest AI model is now open for anyone to download. Tencent Hy (@TencentHunyuan) released the open weights for Hy4 on Friday, its natively multimodal flagship that understands text, images and video, along with a detailed system card on how the model was safety-tested. Key details: Size: a Mixture-of-Experts model with 770 billion total parameters, of which 49 billion are active per token Context: up to 1 million tokens (1,048,576) Under the hood: 78 layers, 256 routed experts plus one shared expert per MoE layer, and DeepSeek-style sparse attention Faster output: a built-in 3.3-billion-parameter drafter called DFly2 handles speculative decoding by reusing the model’s own KV cache Versions: Hy4, Hy4-FP8 and the pretrained Hy4 base, on Hugging Face, ModelScope and GitCode, with deployment support in vLLM and SGLang License: the Tencent Hy License; companies that offer the model as a hosted service need a separate license from Tencent once they pass US$50 million in yearly revenue or 100 million monthly users Why it matters: Open-weight models this large are still rare, and Tencent is candid that a notable gap remains between Hy4 and the top proprietary models in areas like research-level math and long-horizon coding. Still, pairing the weights with a full safety report raises the bar for how #OpenSource AI gets released. Sources: Tencent Hy on X, Hy4 model card and license on Hugging Face, Nov. 20, 2026. Image: Tencent Hy. #AI
@realGalaxyGuide
科技早报|2026年11月21日 有人宣称与 AI 合作证明了黎曼猜想,芯片股继续大跌,智能体则开始碰上隐私边界。 1. Anthropic 数学家自称与 Claude 合作证明黎曼猜想 数学家 Alpöge 发布短文称,他与 Claude 合作找到了黎曼猜想的证明,同一方法也可用于广义黎曼猜想。相关 Lean 形式化文件已公开,独立核验仍在进行,目前还没有正式论文;文件通过编译不等于数学界已确认这一证明成立。 2. 英伟达收跌约 14%,市值被字母表超越 英伟达周五收于 125.53 美元,下跌约 14%,收盘市值降至美股第四。有报道称,美国证监会执法部门正在关注其会计问题;证监会拒绝置评,英伟达当时未立即回应,调查不等于已认定违规。 3. Anthropic 上市次日回落,盘中接近发行价 Anthropic 次日收于 88.91 美元,较首日收盘下跌约 8.5%;盘中最低 83.05 美元,接近 83 美元发行价,但未跌破。 4. 苹果据报开发自有 AI 智能体 报道援引招聘信息称,苹果正在开发可浏览网页、跨应用操作并编写代码的智能体。研发仍处早期,据报最早可能到 2027 年底推出,公司尚未正式发布产品。 5. 腾讯开放 Hy4 旗舰模型权重 腾讯混元团队宣布开放 Hy4 旗舰模型权重,总参数 7700 亿、每个词元激活 490 亿,上下文窗口约 100 万词元。许可采用腾讯自有协议,并非 Apache 2.0;达到约定规模的大型模型服务需另获书面授权。 6. OpenAI 公布防蒸馏措施 OpenAI 称,已加强对内部推理过程被提取、用于训练其他模型的防护。内部对抗测试中,两类受保护推理提取的成功率分别从约 54.0% 降至 0.2%、从约 22.9% 降至 0.9%,这些数字尚非外部审计结果。 7. Thinking Machines 推出 Inkling 2 Mini 新开放权重模型总参数约 264 亿、每个词元激活约 34 亿,显存占用约 20GB,可在具备相应配置的消费级笔记本上运行。公司称其编程评测领先同体量开放模型,仍需独立横评验证。 8. 前 DeepSeek 研究员开源个人智能体 Linger 前 DeepSeek 视觉模型负责人张博公布 Linger,主打在个人电脑上调用已登录的浏览器、邮件等应用,并支持第三方或本地模型。项目称针对本地大模型做了优化,实际效果仍待检验。 9. 余承东纠正「开车睡觉」表述 余承东在享界 V8 发布会后纠正相关说法:目前向用户提供的仍是 L2 辅助驾驶,虽具备 L3 硬件能力,但功能尚未开放,开车不能睡觉。 10. Signal 开始拦截系统级智能体读屏 Signal 宣布,在 iOS 27 上默认阻止截屏,安卓上检测到智能体尝试读屏时自动开启屏幕保护。官方表示,macOS 尚无法可靠实现同样保护,用户也可自行关闭这项设置。 如果身边有人在用系统里的智能体,又很在意聊天隐私,可以把这份清单转给他看看。
@LaunchAIWorks
We found OpenComputer—an open-source coding agent that runs in your own cloud. Sessions boot in ~24s vs 53–72s for rivals. Isolated git worktrees, resumable anywhere. Worth a look for small dev teams. 👉 https://t.co/ibwshxCSEX https://t.co/X3kEYQsYtF https://t.co/fIMQ0lon7n
@mengpaper
A robot AI that watches camera video and flags danger before a crash 👀 Paper: "Assessing Vision-Language Models for Early Prediction of Unsafe Situations in Robotics" (arXiv:2611.24146, preprint) https://t.co/u5vtgRQLkp #robotsafety #humanrobotinteraction #robotics #mengpaper https://t.co/JwkhoMR4qn
@wainwrightollie
What is NCM? Next-Concept Modeling is a new research method from Stanford's HazyResearch lab and Together AI (Oct 2026, open source, MIT) for making AI write text much faster. Most chatbots write one word-piece at a time, in order. NCM sketches the big-picture concepts first, then fills in the details, generating every piece within a level at once, so it takes a handful of steps instead of one per word. In the paper, at matched model size, it ran ~21x faster on LM1B with better quality and ~31x faster on XSum at the same quality, vs the strongest diffusion-language-model baseline. Honest nuance: it's an early research method with open code, not a product, tested on small research-scale models. #AI #MachineLearning #LLM #AIResearch #Tech
@zbruceli
Nvidia is reportedly buying Inception Labs, the diffusion-LLM startup behind Mercury. Could a 1-trillion-parameter diffusion LLM actually work? I ran the roofline math and benchmarked Mercury 3 myself. Short answer: yes, if it's sparse. https://t.co/0NoGn9UWiL
@micglad_
My laptop's integrated GPU is now running a local AI that can search and edit files anywhere on my PC through a chat window.Didn't need an RTX 4090. Didn't need ChatGPT. Just WSL2 and some patience 😅Project is on GitHub: https://t.co/dGUXvijXdp #LocalAI #BuildInPublic
@dym6100
An LLM has no memory between API calls. Every request carries everything: system prompt + tool definitions + the full message history + room for the reply. That's the context window. All of it has to fit, on every single call. #LLM #AIAgents https://t.co/wxANTb0cUs