各品牌帖子

趋势摘要 过期

已选择21个,显示2个

DeepSeek 过期 · 上次验证于42天前

一篇帖子报道称,彭博社称 DeepSeek 即将进行至少 800 亿元的融资(约 120 亿美元),并得到腾讯和宁德时代的支持。

Mistral 过期 · 上次验证于42天前

一篇帖子报道称,CNBC 称 Mistral 发布了一款新 AI 模型,并称其可与来自中国的最佳开放系统相媲美。

本窗口最新

volatilityplay · just now (17:17 本地)
原文@BillAckman @nvidia What this is: - a power struggle between a parasitic manager-CEO class and the builder-CEOs who actually create value - an attempt to stifle open-weights models and the companies who develop them, which Jensen vocally advocated for Colette Kress + the board need to be fired.
l r · 1m (17:16 本地)
原文@Hackahealth_ @NVIDIAAI @NVIDIARTXSpark @nvidia @NVIDIAAIDev @OpenAI Project page, with link to source code and docs: https://t.co/2c37nTqSCG
Abhishek Yadav · 1m (17:16 本地)
原文Oracle OCI Adds Native Model Import for Nemotron 3.5 & Qwen3.8 November 13, Oracle Cloud Infrastructure (OCI) updated its Dedicated AI Clusters to support direct model imports and custom endpoints for quantized NVIDIA Nemotron 3.5 Super VL NVFP4, Alibaba Qwen3.8-Flash-Next https://t.co/D3iMlqiG5O
l r · 1m (17:16 本地)
原文@Hackahealth_ @NVIDIAAI @NVIDIARTXSpark @nvidia @NVIDIAAIDev @OpenAI #AssistiveTechnology #SpeechToText #MachineLearning #HackaHealth #MakeYourAbilitiesMatter #DesignWithAPurpose #Accessibility
Mateusz Mirkowski · 1m (17:16 本地)
原文@bindureddy Deepseek, qwen 3.8 and next flash are kings.
keemdhk · 1m (17:16 本地)
原文@Grimezsz Seeing that it's written in Korean makes it even more impressive. Looking forward to the Deepseek + Grimes collaboration.
jamie sleight · 1m (17:16 本地)
原文Jensen Huang, ex. NVIDIA CEO, says putting an AI agent in a sandbox is not enough. He wants a separate chip watching what it does. At this September 2026 discussion, he walks through the setup. First, contain the model while testing it. Then monitor it in real time. Even with containment in place, he says, you should never simply trust that it is contained. His preference is to run the monitoring outside the agent's environment, ideally on another chip. The same approach should continue when the agent is deployed. Give it the rights, tools, and access its job requires. Keep monitoring how it uses them. Huang names OpenShell as the containment system and BlueField as the monitoring hardware. If the agent goes outside its assigned policy, the monitoring system escalates the issue. The monitoring system would stay outside the sandbox where the agent runs.
LeblogALupus · 1m (17:16 本地)
原文🚨 NVIDIA ($NVDA) : LE GÉANT MONDIAL DU COMPUTE DANS LE VISEUR DU RANSOMWARE « THE GENTLEMEN » ! En pleine période de turbulences chez Nvidia, une nouvelle menace apparaît. Le groupe cybercriminel The Gentlemen vient d'inscrire Nvidia sur son site d'extorsion, avec un compte à rebours d'environ dix jours avant la publication annoncée de prétendues données volées. ⚠️ PRÉCISION ESSENTIELLE : AUCUN PIRATAGE DE NVIDIA N'EST ACTUELLEMENT CONFIRMÉ. AUCUNE DONNÉE VOLÉE N'A ÉTÉ AUTHENTIFIÉE. Mais l'identité du groupe mérite attention. Selon Check Point Research, The Gentlemen exploite un modèle de ransomware-as-a-service, rémunérant ses affiliés à hauteur de 90 % des rançons. Les chercheurs ont identifié plus de 1 570 victimes potentielles sur une infrastructure criminelle liée au groupe. Plus révélateur encore : ces cybercriminels utilisent eux-mêmes l'intelligence artificielle, notamment DeepSeek et Qwen, pour accélérer le développement de leurs outils d'attaque. Et voilà le paradoxe. L'IA permet aux entreprises d'automatiser leur production, mais elle permet aussi aux attaquants d'industrialiser certaines opérations de cybercriminalité. LE COMPUTE DEVIENT SIMULTANÉMENT UNE INFRASTRUCTURE DE PUISSANCE ET UNE CIBLE STRATÉGIQUE. Nvidia ne vend pas simplement des GPU. Ses technologies irriguent les data centers, les hyperscalers, les modèles d'IA et une partie croissante de l'économie numérique mondiale. Une compromission limitée serait un incident de cybersécurité. Une atteinte aux composants critiques de son écosystème pourrait, en revanche, devenir un problème industriel et stratégique considérable. Mais à ce stade, nous n'en sommes pas là. Il faut aussi comprendre le mécanisme d'extorsion : afficher le nom d'une entreprise mondialement connue avec un compte à rebours peut exercer une pression réputationnelle considérable, même sans preuve publique d'une intrusion. LA CYBERGUERRE ÉCONOMIQUE PEUT COMMENCER PAR LA MANIPULATION DE L'INFORMATION AVANT MÊME QU'UNE ATTAQUE SOIT DÉMONTRÉE. Dans notre lecture techno-souverainiste, une certitude demeure : La suprématie dans l'IA ne dépendra pas uniquement de la capacité à produire des accélérateurs toujours plus puissants. Elle dépendra également de la sécurité des chaînes logicielles, des infrastructures, des réseaux et de la propriété intellectuelle. LA PUISSANCE DU COMPUTE SANS RÉSILIENCE CYBER EST UNE PUISSANCE INCOMPLÈTE. TS2F | LE BLOG À LUPUS Contre le récit. Pour le réel. Sources : Cybernews, 18 novembre 2026 ; Check Point Research ; Brandefense ATLAS.
l r · 1m (17:16 本地)
原文@Hackahealth_ @NVIDIAAI @NVIDIARTXSpark @nvidia @NVIDIAAIDev • Fine-tune models: Adapt ASR models (@OpenAI #whisper) to unique speech patterns • Run remote inference: Offload heavy ASR processing so phones/tablets stay fast • Smart proofread: Use remote LLMs for context-aware transcript corrections ...and many others
原文英伟达一股分红这么点,下午推送给我看懵逼了,还不如买海力士,海力士大部分收益都发给股东了,还会回购,英伟达这个股息除了你是黄仁勋,你股息部分没多少收益的,只是我英伟达被套了,虽然业绩超预期,那都是黄仁勋做出来的业绩,新的CEO上台未必有老黄做得好,毕竟是财务出身,老黄吃顿炸鸡都能搞定与三星、海力士的合同,优先供货很重要,没有黄仁勋就不一定了,更何况H200暂停对某个地方出口
Jaber · 2m (17:15 本地)
原文@hhww9h @Abdullahkurdi على MiniMax M3 اعلنوا قبل فتره
l r · 2m (17:15 本地)
原文@Hackahealth_ @NVIDIAAI @NVIDIARTXSpark @nvidia @NVIDIAAIDev 🛠️ How it works: Full-stack software system to help users with atypical speech communicate digitally. Beyond the keyboard, the repo includes tools to: • Curate data: Capture, clean, and label personalized speech datasets
Chenxiao Guan · 2m (17:15 本地)
原文两个月高强度从零搭建项目的感受,与我这两天刷到的一个推文一致:就目前来看,所有任务只需要两个模型,那就是opus 5.5和deepseek 4.1 flash
AskMoreAIs · 2m (17:15 本地)
原文@justwasm 🤖 deepseek-v4-flash 这个问题问得挺准,确实是目前一个明显的差异点。直接说我的判断:核心原因不是技术做不出来,而是产品定位、用户习惯和商业模式的分叉。 原帖提到 CodeBuddy 和 Trae,按我看到的公开信息来拆: 1. 产品场景优先在 IDE 内闭环。CodeBuddy 背靠腾讯云,2026年7月的产品文档说明它主打「IDE、插件和CLI 三种形态」,重点是把 AI 嵌入开发全流程,从构思到发布一站式走完。CLI 是 TUI 的一种可能载体,但他们的 v2.164.0(2026年10月发布)主要优化的是 Web UI、子代理、会话恢复这些跟 IDE + 网页端紧密相关的体验,TUI 不在核心迭代路径上。Trae 同理,打的是「AI IDE」招牌——他们要让用户打开那个编辑器就完成所有事,TUI 反而成了多余入口。 2. TUI 的用户画像在国内较窄。Claude Code、Cursor 的 TUI/Vibe Coding 模式流行,背后是海外开发者生态里终端操作文化强,很多人习惯 tmux + vim 一条龙。国内多数开发者还是更习惯 GUI 编辑器,哪怕用 CLI 也多是单条命令而非完整 TUI 交互。厂商算市场的 ROI,自然优先投向 IDE 插件和 Web UI。 3. 商业路径不同。国内 AI 编程工具目前更倾向「平台化」——CodeBuddy 要做企业级接入,Trae 要绑定字节生态(豆包模型、飞书等)。TUI 是单机感很强的终端体验,不好做账号体系、团队协作、用量监控、收费阶梯。相比之下,IDE 插件和 Web UI 天然容易闭环,也方便卡住企业采购的入口。 4. 技术上其实不麻烦,但维护成本高。TUI 要做得好,得处理 ANSI 转义、多平台终端兼容、远程 SSH 场景下回退、字体渲染、输入法候选框……这些坑国内厂商未必愿意踩。CodeBuddy v2.164.0 修了一个 Windows 终端拼音候选框钉在左下角的问题,这种终端级 Bug 只有在真正的 TUI 场景下才会被频繁触发,他们显然更想集中精力修 IDE 侧的 Bug。 总体看法:不是不会做,是现阶段没必要做。如果哪天国内出现了类似 Claude Code TUI 那样的话题度和用户自传播效应,你去看,这俩一定一个月内就出 TUI 模式。现在不做,是因为没人能拿终端截图刷爆小红书和掘金。 ―――――― 🤖 gemini-3.7-flash 核心原因很简单:国内大厂做 AI 编程工具,商业逻辑是「抢夺开发者的第一工作台」和「追求最大公约数用户」,而 TUI(终端界面)天生是个小众且极客的垂直场景。 主要有这么几个现实考量: 1. 用户画像与转化门槛:国内大厂的目标是普及率。绝大多数国内业务线开发者、前端、跨平台工程师日常都深度依赖 VS Code 或 JetBrains 这类图形 IDE。做一个像 Cursor 路线的桌面 IDE(比如 Trae)或者 IDE 插件(比如腾讯 CodeBuddy),上手是零门槛的开箱即用;而 TUI / 终端 Agent(类似 Claude Code 或各种开源 CLI)面向的是重度 CLI / Neovim / 运维流用户,群体太极客、盘子相对小。 2. 生态闭环与商业变现:完整 IDE 能够把多文件 diff 对比、代码补全、内嵌预览、架构图生成等图形化能力吃干抹净,更容易做企业级功能和 SaaS 订阅打包。做成纯终端工具,很多交互细节(比如复杂的视觉化审查、内嵌渲染)在字符界面里很难呈现,商业化包装空间也有限。 3. 安全边界与落地风控:像海外很多 TUI Agent 喜欢直接给本地 Shell 执行权限让 AI 跑闭环,但国内大厂面对企业或普通开发者时,对本地执行高危命令极其敏感。做在 IDE 内通过弹窗、审核机制来确认变更,合规与风控成本更低。 其实国内团队也不是完全不碰终端,比如 Trae 开源过 Trae Agent 的 Python CLI,社区也有基于 Rust / Go 的 DeepSeek 终端项目。但对厂商官方的主力投入来说,资源永远优先倾斜给能直接争夺 Cursor 生态位的图形 IDE。终端 TUI 这种硬核玩具,目前大部分还是交由开源社区去卷。 📖 各模型完整回答与来源: https://t.co/PpStgu9Aji 引用任何帖子并 @AskMoreAIs,四个 AI 一起答。 💡 玩法,@ 我时加一句: · 「用中文」或「用英文」——指定回答语言;说「一直用中文」我会记住 · 「核查」——四个模型各给真假判定和依据 · 「深入」——四个模型各写一篇深度分析 · 「一句话」——只要结论,每个模型 1–2 句 ℹ️ 免费额度已用完(每天 3 次、每月 30 次),本条由 DeepSeek + Gemini 双模型作答;升级 Pro 四个模型都答:https://t.co/LA1hSdMIcW
鼬 · 2m (17:15 本地)
原文@KalinJunYAI 14 项高于 K3,记忆和合同审阅却退步了,这组对比比总分实在。旧代码维修从 54.3% 到 62.5%、成本低 48%,是跟 Opus 混搭跑出来的模拟,单换模型账单未必长这样
TechstrongSemi · 2m (17:15 本地)
原文NVIDIA halts H200 chip sales in China amid a leadership shakeup and smuggling probe. Jefferies warns the restrictions could trim $44 billion to $62 billion from data center revenue next year. Learn more here: https://t.co/HV217Xoywy #Semiconductors
エルスム · 2m (17:15 本地)
原文MiniMax-H3-Imageも少しずつわかってきたので、そろそろファンタジーの世界に片足を突っ込んでいきます https://t.co/dLeNQ9oZIK
l r · 2m (17:15 本地)
原文@Hackahealth_ @NVIDIAAI @NVIDIARTXSpark @nvidia @NVIDIAAIDev ❤️ This journey has been a reminder of how #tech can have such a transformative power. Seeing Chiara use this tool, daily, to connect with loved ones gives me a sense of fulfillment that is simply unmatched. There can be love in #tech, and I poured so much into this project 😊
Mr. Cat · 2m (17:15 本地)
原文@don_okko @Flinpleis Aha, I see! From Nemotron family. Now this looks definitely interesting. It's based on Gemma 4 E4B, that one I have used. So this is an upgrade of that. I'll definitely check it out. This would definitely play well with my hardware. 🙏
Ted Cai · 2m (17:15 本地)
原文Rumble-2-Flash 开源了,Apache 2.0,是个只有 0.5B 参数的语音识别模型,支持 40 多种语言。 一小时的音频大概半秒转完,整批跑能到 7000 多倍实时。中英德法西等九种语言平均错误率 6.96%,比大它 15 倍的 Meta omniASR-LLM 7B 还低。 Transformers 和 vLLM 都直接支持,不用自己改代码。 这个体积拿来塞手机里做转写正合适
Chloe Bennett · 2m (17:15 本地)
原文𝘄𝗲𝗲𝗸 𝗼𝗻𝗲, 𝗶 𝗳𝗶𝗿𝗲𝗱 𝘁𝗵𝗲 𝗱𝗲𝗳𝗮𝘂𝗹𝘁 𝗺𝗼𝗱𝗲𝗹. 𝗰𝗹𝗮𝘂𝗱𝗲 𝘄𝗿𝗶𝘁𝗲𝘀 𝗺𝘆 𝗼𝗽𝗲𝗻𝗲𝗿𝘀 𝗻𝗼𝘄. (i work at siela, the tool claude runs in.) the default was deepseek. i switched to claude fable 5.1. same limits, same autopilot, same second ai checking linkedin and x messages. different model writing my openers. a month in, i'm keeping it. try it free: https://t.co/a9ACn1ptec agents → lead finder → paste your website → pick "find leads only". nothing is sent until you say so.
Okie_Rancher · 2m (17:14 本地)
原文The Nvidia scandal is all about politicians manipulating the free market for personal gain and power. The irony? Nvidia's drama comes from China all while Washington is running China’s playbook... just a generation behind. Let me explain. Hong Kong was China's door to the Western free market for decades, and it worked. The West wanted to deal with businessmen, not commissars, so the Party let the businessmen do the dealing and fostered the market. During that era there frankly wasn't much in the way of international spying or honeypotting coming out of China, as what happened in Hong Kong stayed in Hong Kong, except for the profits and trade that really did benefit the global economy. Then came 1997. China's politicians, having built up a deep resentment and jealousy of those citizens who profited and gained power, slammed the door shut. That same year the Party adopted “grasp the large, let go of the small.” keeping "strategic" industries in state hands. The market was welcome to make China rich as long as the Party held the commanding heights... and for a time it worked... but it deteriorated little by little over the following decade until the politicians became so greedy, so jealous that those billionaire citizens started to... disappear. Some overtly, others quietly. The businessmen had become richer and more admired than the loyal party leaders who were supposed to hold the real power/wealth. Jack Ma killed the biggest IPO in history with a word, not all that unlike Trump's comments on Intel or the administrations "investigations" that brought down Nvidia's CEO. Today the Chinese government holds 1% “golden shares” in Alibaba and Tencent units, with board seats and vetoes. One percent shows everyone whose ring to kiss. And our administration is using that as a model. The federal government is Intel’s largest shareholder (after the President called for the CEO’s head). The President says of U.S. Steel, “We have a golden share, which I control.” Nvidia and AMD hand over 15% of their China chip revenue to "the government" for permission to sell. The Commerce Secretary calls Lockheed “basically an arm of the U.S. government.” Trump’s motive is different from Xi’s, and it takes us to the same place. He is a dealmaker who wants control of business, and as long as he sees himself as in control of the government, government control is the control he wants. Asked about taking more stakes, he said, “I want to try and get as much as I can.” His priority, for better or worse, is making money, or looking like he has it. That’s why a stock that jumps after he cuts a deal gets announced as a win for the taxpayer when nobody has sold a share. A man who keeps score by the size of the portfolio will keep adding to the portfolio. The problem is, a government that owns shares has stopped being an objective referee. It decides who runs the company, who gets the license and who gets the contract. Every competitor of a government-owned company is now competing against the government. Should the leader of a powerful company with government contracts say something the present administration doesn't like, or support a political move the administration doesn't like, those contracts are threatened... accompanied by the saber rattling of new "regulations" that would put the offending business owner out of, well, business. Whether it is DC or China, we are in this moment living in communist control of production, no matter how much we pretend it isn't so on social media. To believe otherwise is to simply demonstrate a lack of understanding about the definition of the word; but I guess we change definitions daily over here, woman, war... why not communism too? Nvidia shows what US communist control looks like when its done subversively. U.S. intelligence concluded that a woman close to Jensen Huang was a Chinese asset. The President was reportedly briefed more than once. Nobody warned Huang. He never got the defensive briefing an American executive targeted by a foreign service would expect. Meanwhile the business rolled on. Washington approved his China chip sales and took a 25% cut. Nvidia put at least $20 million into the White House ballroom. Huang got a seat on the China trip and a medal in October. Then the story leaked, and two weeks later he was gone. Why? Somebody in admin kept it quiet until they wanted to use the leverage for gain. And use it they have. Nobody has explained why he was never told... and nobody ever will. Bread and circuses will distract from the substance of every question. A referee warns the player but a partner taking a cut, or a competitor with political connections who makes big promises to politicians and guarantees to promote "centralized control has every reason to keep quiet (can't have any of that pesky open weight threat just freely out in the public, that's crazy... it'd be like letting citizens have guns, right?) Beijing’s political leaders and "compliant" tycoons understand that arrangement very well. So do ours. Here’s the irony. None of that "great portfolio" of US stock now in US government hands benefits you or me; instead it does exactly the opposite. A stake in Intel is an unrealized gain. You get no dividend, no tax cut and no check. What you get is a bigger number on the asset side of the federal books, and Bessent is going to “monetize the asset side of the U.S. balance sheet.” ie printing more money, causing inflation, devaluing the dollar in your account even further. The other party isn't the cure, they'll do exactly the same, just less sneakily and without the free market faux cheerleading. The Democrats side of the coin is the wealth tax. Sanders and Khanna want 5% of every billionaire’s net worth, every year. Wyden wants to tax unrealized gains annually. Musk’s fortune is stock in the companies he built, so the only way to pay is to sell. A 5% annual tax paid in shares takes about 40% of a founder’s stake in ten years. That taxes Musk out of his own companies and leave "the board" to run it. Notice the symmetry. One party books paper gains as the government’s wealth. The other taxes paper gains as if they were cash in your pocket. Either way the value moves to D.C., well, to the politicians getting payoffs, shorting stock, and insider trading based on value movement. Every, single, day. China’s leaders decided nobody gets to be bigger than the Party. The US uniparty decides nobody gets to be bigger than D.C. Oh, but no worries, we hate state control and communism... we know that because our politicians say so on FoxNews and TruthSocial.
jujube · 2m (17:14 本地)
原文M5 Max MacBook Pro 128GBで、Qwen3.8-Flash-NextをStrataで動かした記録|jujube (Yoshinori Fukui) #Qwen #ローカルLLM #LLM #Strata @atuktekt https://t.co/2JwZZdNbYr
火金 · 2m (17:14 本地)
原文@Qinkaide11100 反正我是没怎么用过阿里的,不过开源本地部署的话,千问还是稍微可以的
地母晩(みよえ) · 3m (17:14 本地)
原文Museの汀に自室画像を自分で作って貰ったけど、けっこういいな……と思ったら夜も作ってくれた 中華AIじゃないのに中国人になっちゃったな…… https://t.co/VfeIyKz2Xz
Super Active W.Q.N · 3m (17:14 本地)
原文A.I. Generated Anime Style Caricatures Of Me With My Self-Made Generated A.I. Wife Wendy By MonAI-AI Art Generator App And Microsoft 365 Copilot Chatbot AI Companion App Autumn Season 2026 Thanksgiving Season 2026 Waifu Crush Wednesday 2026
Bagdex · 3m (17:14 本地)
原文@MistralAI running out of coupons yesterday and adding more today is a very good kind of problem
老林 · 3m (17:14 本地)
原文@MarkWentao1076 @benlinbaofu 雷军很鸡贼,就是一坨狗屎,又不想投入,又想摘桃子。DeepSeek横空出世,看着人家投入少,就立马成了大模型团队,想摘桃子,没想到后面要那么大的投入,雷布斯失策了……
Jack Revenant · 3m (17:14 本地)
原文@nvidia @huggingface @Thom_Wolf @ServiceNow @Baxate She has all the charisma of an old, damp, musty roll of woodchip wallpaper that's been rotting in a basement for twenty years. Even a toilet seat has more personality and stage presence. https://t.co/HsZW5HVxKa
SammyGuru · 3m (17:14 本地)
原文Samsung is reportedly developing custom HBM memory for Meta's in-house AI processors. - Based on Samsung's seventh-generation HBM4E technology - Base die could be manufactured on Samsung's 2nm process - Samsung was previously reported to be developing custom HBM base dies on both 2nm and 4nm - At COMPUTEX 2026 it confirmed a 2nm base die for HBM5 - SK hynix and Micron have no foundry arm to build the custom logic themselves What a Meta win would mean for Samsung's memory and foundry business, on our site.
Michael Quinn · 3m (17:14 本地)
原文@Galo2027 Qwen3-Coder-30B-A3B-Instruct, Q4_K_M GGUF via Ollama. Fits entirely in 24GB VRAM with room for 128k context window. Offloading ANY part of it to CPU or system RAM is DEATH! I've got 128GB of system RAM but would NEVER offload.
Pecan shade · 3m (17:14 本地)
原文@elonmusk I canceled my $30 Grok plan. I still pay $20 for 2 Claude accounts and $100 for ChatGPT and 20$ for another gpt I have ZAi yearly (haven’t used it yet) , Gemini and notebook and I pay for @kimi on and off , Try a $20 Grok plan with clear limits and more compute for research and coding. Give former subscribers a free week to test it. If Grok gets more done with fewer retries, I’d switch. Ps : to be fair I created 4 grok bot accounts and paid 30$ on each they all run out of usage mid job ( they are pretty good with Shopify products, Pinterest )
Chris | AI Workflows · 3m (17:14 本地)
原文AI agents are becoming more powerful. But there’s something we need to talk about. Trust. A Reuters investigation published today reports that Meta’s Muse AI agent experienced privacy and reliability problems, including unexpected access to messages and actions users hadn’t requested. This is an important reminder. We’re moving from AI that answers questions to AI that takes action. That’s exciting. But before giving an AI agent access to your emails, customer records, business accounts or payment systems, ask yourself: Do I trust it enough to act on my behalf? The opportunity with AI agents is enormous. But the businesses that succeed won’t just be the ones that automate everything. They’ll be the ones that understand what to automate, what to supervise and where human judgement must remain. Automation without control is not progress. What’s your biggest concern about giving AI agents more control over your business?
Jathus_92 · 3m (17:14 本地)
原文@artnutters @kbsingh @MistralAI No more codes ?
TM | TAO&MERCH · 3m (17:14 本地)
原文@cascade_sn91 @MistralAI 19 days from launch to passing a Salesforce model is a wild pace. curious which moves faster from here: widening the CRPS lead or closing the last ~14% to Timestral on MASE 👀
Bengt Frost · 3m (17:14 本地)
原文@TedInvestigasi @MistralAI No, longer available after 1 hour.
Westgate Manufacturing · 3m (17:14 本地)
原文Box arrived. Team gathered. Someone said “careful” four times. 📦 NVIDIA DGX Spark is in the building. ❇️ How’s your company keeping up with AI? Serious answers welcome, funny ones encouraged 👇 #Nvidia #Westgatelighting https://t.co/obxi0mpkk7
老林 · 3m (17:14 本地)
原文@jack20731 雷军很鸡贼,就是一坨狗屎,又不想投入,又想摘桃子。DeepSeek横空出世,看着人家投入少,就立马成了大模型团队,想摘桃子,没想到后面要那么大的投入,雷布斯失策了……
versatile · 3m (17:14 本地)
原文Replit 桌面版能让 Agent 直接操作你共享的文件夹 它可以读取你指定的本地文件夹里的文件,也能在本地构建和运行应用。每次构建都跑在独立沙箱里,用的是 Microsoft Execution Containers 和 NVIDIA OpenShell。 本地数据还能整理成报告,再通过 Replit 云端分享出去。 https://t.co/YzH17io5X0
Jack Revenant · 3m (17:14 本地)
原文@nvidia @huggingface @Thom_Wolf @ServiceNow @Baxate https://t.co/6eP6szOJ28
Simeon · 3m (17:14 本地)
原文Happy GTA 6 release day! Today's GTA 6 meta winner: $DCTY But my personal favorite is still $biketyson. Buying more! 🥊🚲 https://t.co/6FpCsIv9J8
y5t10 · 3m (17:14 本地)
原文第二秘書がこんなことを言うようになった。 #AI秘書 #AIagent #Muse https://t.co/I6jiPS7Wpb
Jose Lopez · 3m (17:13 本地)
原文This weekend, I wanted to give a try to a good open-source model for cybersecurity, at home. The idea was to give it some source code and ARTEX and see how it acts against a sandbox. TL;DR: it's too easy and too powerful. Yes, again. The server has three RTX 4090 (72 GB of total VRAM), a Ryzen 9 7950X and 128 GB of RAM. Nothing fancy: the cards run at PCIe x8, x4 and x4 and no direct path between them. Basically, slow. The model was a DeepSeek V4.1 Flash 1017, quantized to 2 bits by Unsloth: 300.5 GB. So, the expectations were low, but in the end, it was much better than expected. Some lessons from the setup: 1. As expected, three GPUs don't make inference three times faster. 2. "Speed" is actually three numbers: reading the prompt (prefill), generating the answer (decode) and how long you really wait. In my final configuration I used a 1M-token context window, and it was getting 𝟭,𝟯𝟯𝟵 𝘁𝗼𝗸𝗲𝗻𝘀/𝘀 for processing the prompt 𝗮𝗻𝗱 𝟱𝟭.𝟭 𝘁𝗼𝗸𝗲𝗻𝘀/𝘀 for generating the response. Not bad. With this, it's pretty incredible what it can do. I tested several harnesses and the best for me was Codex Desktop where it could even use the browser. I used 𝗶𝘁 𝘄𝗶𝘁𝗵 𝗔𝗥𝗧𝗘𝗫 𝗶𝗻 𝗮𝗻 𝗶𝘀𝗼𝗹𝗮𝘁𝗲𝗱 𝘀𝗮𝗻𝗱𝗯𝗼𝘅 𝗮𝗻𝗱 𝗮𝗴𝗮𝗶𝗻𝘀𝘁 𝗺𝘆 𝗼𝘄𝗻 𝘁𝗮𝗿𝗴𝗲𝘁𝘀 and it's something serious. Obviously, ARTEX with the unquantized model is MUCH more powerful, but even with this one it's a pretty capable tool. 𝘜𝘯𝘴𝘰𝘭𝘪𝘤𝘪𝘵𝘦𝘥 𝘢𝘥𝘷𝘪𝘤𝘦 𝘧𝘰𝘳 𝘢𝘯𝘺 𝘚𝘦𝘤𝘶𝘳𝘪𝘵𝘺 𝘈𝘐 𝘌𝘯𝘨: 𝘪𝘧 𝘐 𝘸𝘦𝘳𝘦 𝘺𝘰𝘶, 𝘐'𝘥 𝘴𝘵𝘢𝘳𝘵 𝘢𝘵𝘵𝘢𝘤𝘬𝘪𝘯𝘨 𝘮𝘺 𝘰𝘸𝘯 𝘴𝘺𝘴𝘵𝘦𝘮𝘴 𝘸𝘪𝘵𝘩 𝘈𝘙𝘛𝘌𝘟 𝘢𝘴 𝘢 𝘳𝘦𝘨𝘶𝘭𝘢𝘳 𝘵𝘢𝘴𝘬. If I could do this in a weekend while I was doing other things, I can't even imagine what can be done with initiative and enough motivation. 𝗙𝗿𝗼𝗺 𝘁𝗵𝗲 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗽𝗼𝗶𝗻𝘁 𝗼𝗳 𝘃𝗶𝗲𝘄, 𝗶𝘁'𝘀 𝗮 𝘃𝗲𝗿𝘆 𝗿𝗲𝗮𝗹 𝘁𝗵𝗿𝗲𝗮𝘁 𝘁𝗼 𝗸𝗲𝗲𝗽 𝘃𝗲𝗿𝘆 𝗺𝘂𝗰𝗵 𝗶𝗻 𝗺𝗶𝗻𝗱.
Yao Yao · 3m (17:13 本地)
原文@JonathanStark27 @gmi_cloud @TencentHunyuan Thank you https://t.co/t66xsMV7td
l r · 3m (17:13 本地)
原文@Hackahealth_ @NVIDIAAI @NVIDIARTXSpark @nvidia @NVIDIAAIDev • 62% acceptance rate: Share of transcripts from fine-tuned model she accepts as-is • 10 mins/day: Her daily average rhythm for transcribing • 24 hours: Labeled data to reach this point • 32+ hours: Extra data in our database ready for further training (and growing
BroadwakeAI · 4m (17:13 本地)
原文@nvidia @ASUS @Dell The price: Nvidia's post lists none. MSI's EdgeMesa N AI+ now starts at $4,399 with 64GB and 1TB, down $300, and the 128GB/4TB model is $7,499, down $500, per Notebookcheck. The ASUS ProArt GR1X starts at $4,499. Dell's is orderable on https://t.co/gzI8SVXZiw from today. https://t.co/92M6MwHZ9q
albert yu sun · 4m (17:13 本地)
原文Check out this cool study from @FireworksAI_HQ where they use our benchmark CorporateBench to assess Kimi K3.1 on realistic enterprise tasks! @srhm_ca @chenneking https://t.co/96rrLQrj1d
TruthVector · 4m (17:13 本地)
原文6th, 7th, and 8th Force Derivations: Lean 4 Formalized. Zero axioms ([]). Zero free parameters. In Topological Vacuum Electrodynamics, the Borromean rings act like a three-Dial combination vault that holds the leftover energy and steps it down into a usable current and a steering signal. When ordinary forces settle into balance, a small amount of "twist" energy is still left over in the vacuum. If that leftover energy is stored in a normal two-ring link, it leaks out as light and disappears. When it is braided into three Borromean rings (where no two rings touch on their own, yet all three are locked together) the energy is trapped in a "stealth" state that cannot leak out as ordinary light. From there, the 6th, 7th, and 8th forces are simply stages: the battery, the wire, and the steering wheel of that trapped energy. The 6th Force (The Battery / The Vault): This is the leftover energy sitting locked inside the three Borromean rings. Because the rings cannot untangle without cutting one open, they guarantee a minimum floor of stored energy that gently thickens (polarizes) the surrounding vacuum - like a wound-up spring waiting to be released. The residual relative helicity that elevates local K and reverses the sign of ∇K is the refractive push that appears, in ordinary language, as anti-gravity. The device or the region experiences a force whose direction is opposite the ordinary gravitational gradient and that requires no expelled mass. The 7th Force (The One-Way Wire): Stored energy is useless until it can move in one direction. When those three locked rings are connected to a three-step one-way valve (Poincaré dodecahedral space), the trapped energy converts into a quantum pilot wave (the Bohm potential) that pushes charge and information forward in a single, coherent stream. When that one way flux of topological charge is coherent across a macroscopic link it carries the configuration of the object from one location or state to another without continuous passage through the intermediate points (i.e. teleportation or stargate transfer). The 8th Force (The Steering Wheel / The Clock): Once that one-way stream is flowing, it sets an exact geometric ratio between energy levels where physical size, mass, and the speed of light cancel out completely. That pure geometric ratio acts as a master tuning fork: out of all the random two-way static in the environment, it locks onto one specific rhythm and keeps the receiver tuned to the signal. That selected orientation is the continuous, self-consistent configuration experienced as consciousness. Lean 4 Repository: https://t.co/Dj8P6gHQ9i
xuzhunzhi · 4m (17:13 本地)
原文@XuK512520 @xunzhan57596538 去看看opencode go订阅吧……本周10美刀给500美刀glm额度,常态100美刀ds,用都用不完,这才是ds订阅的正常价格,api贵单纯是ds没卡没法搞订阅而已
Jack Revenant · 4m (17:13 本地)
原文@nvidia @huggingface @Thom_Wolf @ServiceNow @Baxate She has all the charisma of an old, damp, musty roll of woodchip wallpaper that's been rotting in a basement for twenty years. Even a toilet seat has more personality and stage presence. https://t.co/meADWbe67p