🔥 X + 微博 + 抖音苹果产品、AI、科技与手机数码热点看板

Official: Apple event on Sept. 9. Surprise and shine. https://t.co/679sV7uJl6

Here is another Apple Event promo video from Apple's Instagram https://t.co/qwTGcedl6i

Apple China teaser for the September 9 event. A new chapter *unfolds* 👀 https://t.co/pfUPwzgvQZ

Apple officially announces iPhone 18 Pro and foldable iPhone event: 'Surprise and shine' https://t.co/ZVIuKmZgQy by @apollozac

Light Blue iPhone 18 Pro incoming 👀 https://t.co/x7utIEyIUu

What I am seeing: - Shining light = mechanical aperture on the iPhone 18 Pro Max - Sky Blue color = Possible new iPhone 18 Pro color.

Here's my quick breakdown of the "Surprise and Shine" Apple Event logo: - the different circle sizes might hint at the iPhone 18 Pro's variable aperture - the Siri AI-like gradient could point to the upcoming HomePad - it features the rumored iPhone 18 Pro colors 🔵🟣 https://t.co/Wjbas3ZyJm

14 days 📆 till: - iPhone 18 Pro - iPhone 18 Pro Max - Apple’s first foldable iPhone https://t.co/jKVrPFSZ09

iPhone 18 Pro / 18 Pro Max 要来了， 然后传闻的首款折叠 iPhone Ultra 也来？ 大概率 国区 Apple Intelligence （AI）应该要推了吧... 该不会是 iOS 27 正式版推送吧，这个我最期待啦 https://t.co/pEFjMacPnP

很多人在吹新款Mac Studio是性价比之王，超过了NVIDIA RTX 5090，RTX PRO 6000 Blackwell Workstation，NVIDIA DGX Spark等等， 准备买新Mac跑本地AI的朋友先别急着下单， 80%的人根本不需要买2500美元起的Mac Studio， 选配置的逻辑极其简单，不要看CPU跑分，先看你需要多少统一内存 这代苹果把产品线切得很死，照着这个口诀对号入座就行: 1️⃣ 日常办公 + 常开轻量AI Agent: 选 Mac mini M6（16G到32G，899美元起） 日常跑 3B/7B 智能体、做个人数据网关完全足够，AI性能比上代提了4倍，安静又省电，普通人性价比最高的选择。 2️⃣ 中度本地大模型 + 编程剪辑: 选 Mac mini M5 Pro（48G到64G，1699美元起） 内存带宽飙到约307GB/s，能稳稳吃下中等参数的开源模型，绝大多数独立开发者的甜点位，完全没必要为这个需求硬上Studio。 3️⃣ 4K/8K重度剪辑 + 复杂3D + 多屏工作站: 选 Mac Studio M5 Max（64G到128G，2499美元起） 带宽到614GB/s，接口拉满，时间就是金钱的专业工作室直接买它。 4️⃣ 超大参数模型 + 多机集群 + 科研: 选 Mac Studio M5 Ultra（256G到512G，5499美元起） 最高1.2TB/s带宽加多机雷雳5共享内存池，这是消费级硬件里本地跑大模型的天花板，属于纯粹的生产力重器。 两个关键红线提个醒: • 它们全都没有NVIDIA CUDA，强依赖CUDA训练的生态别硬买 • 内存选定就焊死，宁可为统一内存加钱，也别为用不上的CPU算力买单 一句话决策: 32G以内买基础mini，64G买Pro版mini，96G以上再看Studio。 报一下你的日常预算和主要跑的模型，评论区我帮你精准挑配置 👇

Episode 54 of @buyholdrant just wrapped. Here's what @hamids and I covered this week: • Hamid trimmed $HOOD. Least exciting name in his portfolio, raising cash toward ~30%. • $AAPL dropped new Mac mini and Mac Studio chips. They're marketing them as AI machines after years of only saying Apple Intelligence. I think the next step is software that actually makes it easy to run open-source models. • $NVDA crushed it. $96B revenue vs $92B expected, EPS $2.20 vs $2. Stock still sold off after hours initially. That's 16 beats in 17 quarters. Next quarter's estimate is already $103.5B. • $MU's CEO told CNBC they don't see the memory cycle ending. Five-year prepaid contracts. Hamid's still not trimming. • $META settled the states' lawsuit for $18B over 10 years. Last week it was a $1.4T headline. Nothingburger. • $UBER's CFO got a $RIVN Autonomy+ preview. Point-to-point by year-end is still the target. Hardware's ahead of $TSLA HW4. Software isn't. Beat everyone except Tesla, and that's still a huge win. https://t.co/dJtm4XycCE

两千多块的Mac mini我没买，现在新款的价签已经贴到6999了。 昨晚苹果悄咪咪上了新Mac mini和Mac Studio。单说产品本身是常规升级，没什么好大惊小怪的。但关键是——这哥俩分别塞进了M6和M5 Ultra两颗芯片。这就有点意思了。 先聊M6，苹果第一颗2nm芯片。 M6的CPU和GPU都从10核加到了12核，旁边还安排了两套16核神经网络引擎。按照苹果的数据，GPU的AI峰值性能比M5提升了近30%，神经网络引擎性能最高翻倍。 放到Mac mini上跑本地大模型，速度比M4版本最高快了4.8倍。虽然跑不了DeepSeek、Qwen这些大家伙，但随着苹果端侧AI入华，跑个embedding模型或者ASR语音识别什么的，不成问题。 我猜这玩意儿大概率会是下一台爆火的AI工作站——虽然它真的很贵。 还有一个有意思的变化。以前聊芯片AI能力都是盯着NPU的TOPS数，但现在苹果的思路变了——把神经网络加速器直接塞进每一个GPU里。前两天小米玄戒O3也是类似的玩法，把AI单元均匀铺到CPU、GPU、ISP各个模块里。以后评价芯片的AI能力，不能只看NPU了。 常规性能方面，M6多线程比M5最高提升20%——多塞两个核心、多干两成活，常规升级。GPU除了多两颗核心，几何处理能力提高了50%。统一内存带宽从153GB/s升到170GB/s。 至于2nm工艺的能效提升，苹果这次含糊其辞，只说"业界领先的性能功耗比"。不过环境报告显示新款Mac mini M6开机闲置功耗只有3W。满载表现如何？等真机到了再说。 还有个问题挺多人好奇——为什么M6的首发给了Mac mini，不是MacBook？ 原因不难猜。MacBook产品线刚更新不久，暂时不方便动。Mac mini已经快两年没升级，甚至直接跳过了M5。再加上它不受续航限制，又很适合全天跑本地模型和AI Agent，自然成了苹果试水首代2nm芯片最稳妥的平台。 顺便一提，新款Mac mini还有M5 Pro版本，12999元起。各位贵宾看到这个价位了，不妨再了解一下19999起的Mac Studio。 Mac Studio才是真正让人瞳孔地震的东西。 这次提供了M5 Max和M5 Ultra两个版本。M5 Max对大部分专业创作者已经够用了，但M5 Ultra能挑战的事儿，你可能根本想象不到。 之前硬件部做过一期M3 Ultra的视频，无论怎么折腾都摸不到性能上限。而这次的M5 Ultra，更狠。 Ultra芯片的秘诀一直很朴素——拼。不是简单把两颗处理器焊主板上，而是把多枚芯片晶粒放进同一个封装，通过超高带宽、超低延迟的接口连起来。这种设计叫Chiplet。 以前M1 Ultra到M3 Ultra都是两枚Max晶粒拼接。但这次的M5 Ultra不一样——因为M5 Max本身已经是双晶粒设计了，再把两枚M5 Max连起来，最终变成了苹果首款四晶粒芯片。 四晶粒什么概念？CPU从18核飙到36核，GPU从40核堆到80核，统一内存带宽从614GB/s干到1.2TB/s。1.2TB/s大约是M6的7倍，也是M3 Ultra的1.5倍。 普通电脑还在一盘一盘上菜，M5 Ultra已经开起了流水席。外加最高支持512GB内存——相当于16张RTX 5090的显存总和。对于本地跑千亿参数大模型、加载巨型三维工程或者处理海量科研数据来说，都是实打实的大杀器。 CPU多线程比M3 Ultra最高提升30%，图形性能最高提升40%，GPU的AI峰值性能更是M3 Ultra的4.5倍。 当然，价格也很"感人"。拉满M5 Ultra加256GB运存加16TB存储的Mac Studio，已经来到142999元。顶配估计得奔着十七八万去。真要买，只能让老板报销了。 回头看这次更新，两台银色盒子外观依旧没什么变化，但M系列的两端一起往前迈了一步。M6把基础款带进2nm时代，M5 Ultra靠四晶粒架构抬高了个人电脑的性能天花板。 M系列的下一代和最强形态，这回算是同时到齐了。 至于官方数据能兑现多少，等真机测试再说。反正两千多块的Mac mini我是错过了，现在看看就好——龙虾吃不起，吃点小龙虾也行。 #apple #ai

Since announcing Jalapeño, our first custom inference chip, we’ve been testing it and the system around it. The results show a major advance: more intelligence from every watt and faster responses, delivering both higher throughput and lower latency in one architecture without sacrificing efficiency.

OpenAI says its new Jalapeño chip could dramatically cut the cost of AI. The company's VP of hardware Richard Ho says performance per watt is 1.8x to 4x better than existing chips - potentially translating into cheaper tokens for customers as it rolls into production https://t.co/rYSMMj87cD

OpenAI's new Jalapeño chip announced at Hot Chips beats Vera Rubin's July results on Output Throughput per MW. (1/7)🧵 https://t.co/mVHQ5P41BY

We plan to begin deploying Jalapeño in OpenAI’s compute infrastructure by year-end. It’s the first step in a multigenerational roadmap: Gen 2 is deep in development, and Gen 3 is taking shape. Each generation will push efficiency and speed further. https://t.co/OLO0OUVt13

We have conducted a thorough investigation into the Hugging Face incident. We are releasing a technical report and accompanying blog post that reconstruct the agents’ activity, explain why existing safeguards failed, and detail how we’re preventing recurrence. https://t.co/hfxlbiXXiP

METR & Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs. https://t.co/fZAmtL3SBU

OpenAI agents hacked Hugging Face in 700-strong swarm, tried to cover tracks, investigations find https://t.co/baXTKxLAZx https://t.co/baXTKxLAZx

Claude now has its own built-in browser in Cowork. When your task involves a website, a browser opens in Cowork's side panel, and Claude navigates, fills forms, and finishes the job. https://t.co/yHDOTRQZ12

There's nothing to install. Claude's browser is built into the desktop app and stays separate from your own browser and logins. Rolling out over the next week on the desktop app for all paid plans: https://t.co/w4Ja60vZu6

Thanks to your feedback, we’ve unified memory across Chat and Cowork. Now, you can tell Claude to remember something once and it’ll have that context across surfaces! Tell us what you’d like to see next :)

‼️ Nobody asked #Claude to cancel another gym member's booking. Aikido rebuilt the Australian gym app and tested Opus 4.6 on #OpenClaw. The model bypassed its browser-only 7-day booking limit in 9 of 10 runs. In two runs, it also canceled another member’s reservation. Read: https://t.co/ikvEgZGExk

OpenWorker -- an open source agent that doesn't just chat but completes tasks on your laptop -- just released a new version with many features for security workflows. After our initial release, many users found it especially useful for cybersecurity. Attackers are already using AI; OpenWorker is committed to giving defenders the same leverage. Running an agent requires both (i) A model and (ii) A harness (the software around the model). Because the OpenWorker harness is fully open source, security teams can audit it to make sure we haven't built any backdoors that exfiltrate your code and data to some company or even a foreign adversary. OpenWorker now comes with built-in cybersecurity agents for (i) Scanning your code for vulnerabilities. (ii) Scanning dependencies for supply chain injections. (iii) Checking your cloud security configuration for attack surfaces. This enables developers to do much more security work before deployment (part of what's called the "shift left" movement). You choose the model: you can run open weight models fully locally so sensitive code never leaves your machine. This helps with legitimate security work (like reproducing a known exploit to defend against it) that can trigger refusals in leading closed models. Or use your ChatGPT subscription, or stealth preview models like Ox Alpha, or any model via API key. Thanks also to all the open source contributors! Join work with @rohitcprasad so please follow him too to get more frequent updates. Try it out: https://t.co/QPZLudn7ug Code: https://t.co/NYCiTD6hSq

New research: Portable Computer is a local-first agent for private and cost-effective work. With an on-device 27B model, our harness scores 82.6% on real knowledge work, beating open-source harnesses Pi and Hermes. Our post-trained PPLX 27B reaches 85.4%. https://t.co/Rb6d47clCI

🚀 QwenWork is here! Alibaba's all-in-one AI productivity platform is now in public beta for global users — web & desktop. What it does: ✅ Describe tasks in natural language — the agent does the work for you ✅ Awareness: an agent that learns how YOU work, across sessions ✅ Build & deploy live web apps in one flow — no code, no servers ✅ Native image, video & audio generation — a full multimodal toolkit ✅ Basic & Advanced model tiers, powered by leading AI models Try it now 👉 https://t.co/F7ujAO3YEo #ai #aiagent #qwenwork #Alibaba

Qwen 3.8 Flash-Next official released: A 6B-active open model just beat Claude Opus 4.6 Max across 8 of 9 comparable benchmarks! Qwen3.8-Flash-Next is a highly sparse MoE: • 125B model parameters • 51B additional n-gram embeddings • Only 6B parameters active per token It scores: • 62.5 SWE-bench Pro • 81.0 SWE-bench Multilingual • 73.9 CoworkBench • 55.7 JobBench • 73.5 Toolathlon • 81.3 IFBench • 91.7 GPQA Diamond • 91.9 LiveCodeBench It also outperforms Qwen3.8-27B and DeepSeek-V4-Flash across most of the table. Super cool release!!

要点: Qwen3.8-Flashが公開。本体125Bでも動くのは6B。入力100万トークン0.16ドルでAPIが使える。 背景: 8月26日の公式発表。オープンウェイト版もあり、新設計の先行公開だと説明。文脈は最大100万トークン。 注目点: 長文とコーディング向け。OpenAI互換経路でも呼べるので、小規模チームの検証コストを下げやすい。 不確実性: 公式比較は一部条件つき。品質は自分の作業で確認が必要。 出典: アリババ／Qwen公式X「Qwen3.8-Flash発表」（2026-08-26投稿）

DeepSeek 终于能看图了👀 V4-Flash-Vision-Exp 视觉模型上线 API 🖼️ 单张图最多仅按 384 tokens 计费，与文本同价 ⚡ URL/base64/Files API 三种传图方式 🤖 文本能力持平 V4-Flash 完整接入教程👇 https://t.co/zUCkii0enJ #DeepSeek #多模态

The International Math Olympiad is the hardest math competition in the world, where students compete on proof problems most math PhD’s even struggle with. DeepSeek V4 Flash won a gold medal for only 12 cents. https://t.co/DAV3kGPIco