🏆 Headline
A 10-year-old network beats two flagship LLMs: dermatologist's test exposes the soft spot of "general" AI
Dr. Magnus Lynch, a consultant dermatologist at Guy's Hospital in London, ran a test so plain it borders on offensive: 200 biopsy-confirmed skin lesions, each with a clinical photo and a dermoscopic image, and a six-way diagnosis multiple choice for three "examinees." The first two were today's strongest flagship models — GPT-6 Astra and Claude Fable 5.1. The third was a filler contestant: Inception v3, a convolutional network Google published in 2015, fine-tuned on 68,011 public dermatology images for about 20 minutes on a single H100. The result was a blowout. The old CNN named the right diagnosis for 125 of 200 lesions (62%); the two flagships managed 84 and 85 (42%). The melanoma numbers sting the most: of 30 cases, Astra's first choice caught 8, Fable 11, the CNN 13. The article's lead figure is an invasive melanoma on an ear, under 0.8 mm thick — both LLMs called it a harmless seborrhoeic keratosis and said the patient could be discharged. Notably, the LLMs' default posture was nearly all-refer (they kept 96-97 of 100 risky lesions), yet still let 3-4 melanomas go; with thresholds matched, the CNN's missed-melanoma count drops to 1-2 — still fewer. There's an interesting patient-perspective twist: asked "do I need to see a doctor?", both LLMs advised seeing one in 99 and 98 of a further 100 biopsy-proven melanomas — in chat mode they are conservative enough, but as diagnostic aids their read of the image fails the grade. Lynch draws his own honest boundaries: small sample, Inception v3 is only a baseline, and regulated clinical AI products use far stronger ensemble models. And MILK10k has been public since 2025, so the LLMs may well have seen it in pretraining — given every advantage, they still lost.
Source: Lynch Lab (Dr. Magnus Lynch, Guy's Hospital) | 2026-10-25
Flagship models cut prices up to 80% in months: a timeline exposes a strange pre-IPO script
Developer Max Woolf's weekend analysis hit 50 points on Hacker News by lining up OpenAI's and Anthropic's pricing moves since June: on July 30, OpenAI cut small-model Luna's price by 80%; on August 10, Anthropic made Sonnet 5's "temporary launch discount" permanent; on September 1, Fable 5.1 cut cache-read pricing by 75% — and Woolf does the math: in agentic coding, cache reads are 96%-99% of total token spend, so that "footnote price" cut equals an overall bill at roughly a quarter; on September 22, GPT-6 Sol launched at half the already-discounted predecessor price. Within months, same-tier models now cost a fraction of what they did in early summer. The most suspicious explanation is pre-IPO optics — trading inference margin for a growth story; but there is a non-conspiracy explanation too: Anthropic itself has said its new models are cheaper to serve and priced accordingly. Woolf also caught a reverse move: OpenAI simultaneously halved the usage allowance for Codex's $200/month tier, "to reflect our increasingly efficient models." The same day, The Information reported Replit moved the bulk of its $100M+-per-year AI spend from Anthropic to OpenAI — triggered by an OpenAI discount north of 30% while Anthropic discontinued its own. The price war has reached customers' procurement contracts.
Source: minimaxir (Max Woolf analysis) / The Information | 2026-10-25
Cursor wires hands onto its coding agents: any desktop app is now clickable
Cursor shipped Computer Use on October 23: its coding agents can no longer only write code — they can click, type, and operate any application on your desktop. The showcase is a full loop: the agent writes an iOS feature, launches the Simulator itself, taps through the flow, inspects the result, and goes back to revise the code — no human hands required. Invoke it with @Computer Use in a prompt, or let it trigger automatically when a task needs UI interaction; the prerequisite is macOS Accessibility and Screen Recording permissions. No pricing change was announced. The coding-agent race is shifting from "write the code" to "verify the code" — whoever automates the write-test-fix loop inside the IDE claims the most expensive stretch of a developer's attention.
Source: Cursor official / press reports | 2026-10-23
Grok's Odyssey releases its first seven minutes: three months of sparring with Nolan, seven minutes delivered
Three months ago, Musk declared that Grok would make a full-length, "historically accurate" Odyssey, taking aim at Nolan's theatrical version and its casting. This weekend, xAI released "Episode 1" of the series — all of seven minutes, which Musk promptly amplified. The tech stack is nothing to sneeze at: generated by Grok Imagine Video 1.5, 15-second clips with synchronized native audio in a single pass, and xAI claims first place on blind text-to-video arena benchmarks ahead of Sora 2, Veo 3.1 and Kling. A $175,000 creator contest ($100K/$50K/$25K) rides along. Musk promises a full-length feature by end of 2026. 💬 Seven minutes is barely a trailer for an epic, but it was released deliberately for comparison: on one side a practical production spanning six countries with thousands of cast and crew, on the other generated video stitched from 15-second segments. What Musk wants isn't box office for this one — it's whether "generated film" becomes a category. The full feature at year's end is the real exam.
Source: TechCrunch / xAI | 2026-10-25
Kimi K3.1 Pro launches: beats GPT-6 Astra on terminal benchmarks, open weights next Wednesday
Moonshot AI released Kimi K3.1 Pro on October 23: built on the nearly 3-trillion-parameter K3 line, 1M-token context, API available now, full weights pledged open by November 4. Benchmarks: 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra's 58.2% and closing on Claude Opus 5.5's 66.4%; its own stress test cuts hallucination from 23.4% to 8.3% (self-reported, third-party replication pending). More notable than scores is the subtraction before release: cyber-attack data was removed from both training stages, which Moonshot frames as clearing misuse concerns before open-sourcing — cut the capability first, open the weights second. 💬 "Subtract first, then open" is itself the news: the loudest objection to open weights is that anyone can take them and do harm. By stripping dangerous skills before publication, Moonshot half-answers the industry's openness-versus-misuse question in advance. After November 4, whether third parties can replicate these benchmarks is the real touchstone.
Source: Moonshot AI / VentureBeat, TNW reports | 2026-10-23
Sweden's Government Offices block ChatGPT and Claude: one approved tool remains
Sweden's Government Offices announced on October 23 that civil servants' work computers will lose access to ChatGPT and Claude starting Monday, citing failed security review; Microsoft Copilot is the only AI tool that passed. The Swedish Tax Agency and the National Board of Institutional Care had already blocked them, with the Enforcement Authority considering following. A security-review veto is becoming a new battleground beyond the enterprise market for model vendors — an industry that sells "enterprise-grade security" being locked out wholesale by one country's government offices hurts more than losing a few benchmark points.
Source: Swedish Government Offices announcement / press reports | 2026-10-23
Three-time founder Naveen Rao bets: inference hits $10T in a decade, robotics 100x bigger than coding agents
Unconventional AI founder Naveen Rao delivered a three-part thesis on the 20VC podcast last Friday: the inference market reaches $10 trillion within ten years; within three years, 80% of AI tokens run on open models, leaving only 20% for frontier labs; and robotics is a market 100 times larger than coding agents. The supporting math is about power: America's spare grid capacity runs out within a few years, and a new gigawatt of power costs roughly $50 billion — more than a semiconductor fab — so "hardware is the place to be over the next 10 years." His track record lends weight: Nervana sold to Intel, MosaicML to Databricks for $1.3 billion, and his third company builds AI hardware designed around energy efficiency. 💬 When someone who has sold two companies at high marks reads the industry, the valuable part isn't the conclusions but the coordinate system: when power, not algorithms, is the scarce resource, value flows to whoever turns energy into intelligence. If that holds, today's model-application companies are all working for the people who collect the electricity bill.
Source: 20VC podcast / Scuttlebutt summary | 2026-10-24
Elvis Presley's music rights holders sue Anthropic: 274 songs, itemized
The owners of Elvis Presley's music catalog (under Authentic Brands Group) filed suit on Friday alleging Anthropic trained on 274 Presley songs without a license — the list includes "Can't Help Falling in Love," "Don't Be Cruel" and "All Shook Up." Lead attorney Richard Busch wrote that the defendant "built its entire business model based on copyright infringement at a scale never seen before." For the generative-AI industry, what matters here isn't the legal open question but the method: rights holders are no longer gesturing at "training infringement" in the abstract — they itemized 274 specific works. The more specific the case, the easier the settlement number is to negotiate.
Source: Billboard / The Hollywood Reporter | 2026-10-25
🏆 今日头条
十年前的老网络打赢两个旗舰大模型：皮肤科医生实测揭出「通用」的软肋
伦敦 Guy's Hospital 皮肤科顾问医生 Magnus Lynch 做了一场朴素到近乎「冒犯」的测试：200 例经活检确诊的皮肤病变，每例配一张临床照片和一张皮肤镜图像，让三个「考生」做六选一的诊断。前两位考生是当下最强的旗舰模型——GPT-6 Astra 和 Claude Fable 5.1；第三位是陪跑选手：Google 2015 年发表的卷积网络 Inception v3，只用 68,011 张公共皮肤病图像、在一张 H100 上微调了约 20 分钟。 结果是压倒性的。老 CNN 答对 125 例（62%），两个旗舰模型分别只有 84 和 85 例（42%）。最扎心的是黑色素瘤：30 例里 Astra 的首选诊断只认出 8 例，Fable 11 例，CNN 13 例。文章首页的例图是一处长在耳部、厚度不足 0.8 毫米的侵袭性黑色素瘤——两个 LLM 都诊断为无害的脂溢性角化病，建议患者「可以出院」。值得一提的是两个 LLM 的默认姿态近乎全员留检（100 例风险病变留了 96-97 例），即便如此仍放走了 3-4 例黑色素瘤；把判定阈值对齐后，CNN 放走的黑色素瘤降到 1-2 例，还是更少。有意思的还有「患者视角」：当被问「我需要看医生吗」时，两个 LLM 在 100 例黑色素瘤里分别有 99 例和 98 例建议就医——聊天场景里它们足够保守，但作为辅助诊断工具，判读力是不及格的。Lynch 自己也把边界划得很诚实：样本量小、Inception v3 只是基线，临床级的医疗 AI 产品用的是更强的集成模型；而且 MILK10k 数据集 2025 年就公开了，两个 LLM 很可能在预训练里见过它——占尽便宜的还是输了。 > 💬 这份实测最值钱的是对照设计：CNN 没见过测试集，两个 LLM 却可能见过这批公开数据，结果专业的老网络还是赢了。通用模型「什么都会」的卖点，在专科医生眼里恰恰是问题——什么都会等于什么都不精，而诊断错误的账单是要用命来付的。AI 进医疗的路子可能不是把大模型塞进诊室，而是给每个专科配一个便宜、可验证、错了能定位的老网络。
来源：Lynch Lab（Guy's Hospital 皮肤科医生 Magnus Lynch 实测）| 2026-10-25
旗舰模型几个月降价最多八成：一份时间线拆出了上市前的奇怪剧本
开发者 Max Woolf 周末发表的分析在 Hacker News 拿下 50 分，他把 OpenAI 和 Anthropic 六月以来的定价动作排成一条时间线：7 月 30 日，OpenAI 把小模型 Luna 的价格砍了 80%；8 月 10 日，Anthropic 把 Sonnet 5 的「限时优惠价」转正；9 月 1 日，Fable 5.1 把缓存读取价砍了 75%——Woolf 算了一笔账，智能体编码场景里缓存读取占总开销的 96%-99%，这处「脚注价」的降幅等效于整体账单打两五折；9 月 22 日，GPT-6 Sol 直接以比上代打折价再低一半的价格出场。几个月下来，同档模型的价格普遍降到原来的两三成。最可疑的解释是上市前的报表需求——推理毛利让位于增长故事；但也有不阴谋论的解释：Anthropic 自己说过新模型服务成本更低、定价反映了这一点。Woolf 还逮到一个反向动作：OpenAI 同期把 Codex 200 美元档的用量砍了 50%，理由是「模型越来越高效」。同一天 The Information 报道，Replit 把每年上亿美元规模的 AI 支出大头从 Anthropic 挪到 OpenAI——起因是 OpenAI 给了 30% 以上的折扣，而 Anthropic 收回了自己的折扣。价格战已经打到客户的采购合同里。 > 💬 一边把 API 价格砍到骨折、一边把订阅用量砍半，这两个动作放在一起读很有信息量：「便宜」正在从卖点变成武器——预算少的让位，预算多的用折扣锁长约。对开发者是实打实的红利，但每张折扣单背后都是同一个问题：上市之前，到底是毛利重要还是故事重要。
来源：minimaxir（Max Woolf 分析文）/ The Information | 2026-10-25
Cursor 给编码智能体接上了手：桌面上的任意应用都能点
Cursor 10 月 23 日发布 Computer Use：编码智能体不再只能改代码，可以直接点击、输入、操作桌面上的任意应用。招牌场景是一个完整闭环——智能体写完一个 iOS 功能，自己启动模拟器、把界面点一遍、检查结果、回去改代码，全程不用人碰。调用方式是在提示词里加 @Computer Use，或者任务需要操作界面时自动触发；前提是授予 macOS 的辅助功能与屏幕录制权限。官方没有同步调整定价。编码智能体的竞争正从「写代码」卷向「验证代码」，谁能让智能体自己跑完「写-测-改」的循环，谁就拿走了开发者最贵的那段注意力。 > 💬 从「帮你写」到「替你试」，这一步比任何跑分都重要——软件工程的瓶颈正在从产出代码转移到验证代码，把这个环节自动化进 IDE，才是智能体真正替你上班的开始。
来源：Cursor 官方 / 多家媒体报道 | 2026-10-23
Grok 的《奥德赛》放出前七分钟：对线诺兰三个月，交出七分钟
三个月前，Musk 宣布 Grok 要拍一部「历史考据级」的完整版《奥德赛》，矛头直指诺兰的同名院线巨制及其卡司选择；这个周末，xAI 放出了「系列第 1 集」——全长 7 分钟，Musk 随后转发造势。配置倒不含糊：由 Grok Imagine Video 1.5 生成，15 秒片段带同步原生音频一次成型，官方称其在盲测文生视频榜单上排名第一，压过 Sora 2、Veo 3.1 和 Kling；配套还挂出 17.5 万美元的创作者大赛（10 万/5 万/2.5 万三档）。Musk 承诺 2026 年底前交出完整长片。 💬 七分钟对一部史诗来说连预告片都嫌短，但这七分钟是故意放给你比的：一边是六国实景、数千剧组成员的实拍巨制，一边是 15 秒一段拼出来的生成视频。Musk 要的不是这一部的票房，是「生成电影」这个类目能不能站住——年底的完整长片才是真正的交卷时刻。
来源：TechCrunch / xAI | 2026-10-25
Kimi K3.1 Pro 发布：终端基准超 GPT-6 Astra，权重下周三开源
月之暗面 10 月 23 日发布 Kimi K3.1 Pro，基于近 3 万亿参数的 K3 系列，支持 100 万 token 上下文，API 即日起可用，完整权重承诺 11 月 4 日开源。基准成绩：Terminal-Bench 4.0 拿下 65.7%，超过 GPT-6 Astra 的 58.2%，紧追 Claude Opus 5.5 的 66.4%；自家压力测试把幻觉率从 23.4% 压到 8.3%（自家口径，待第三方复现）。比分数更值得注意的是发布前的主动做减法：两个训练阶段都剔除了网络攻击类数据，官方称之为替开放权重扫清滥用顾虑——先砍能力，再开源。 💬 「先做减法、再开放」这个顺序本身就是新闻：开源权重绕不开的争议是「谁都能拿去干坏事」，月之暗面选择在权重公开前先把危险技能拆掉，等于替全行业把「开放与滥用」这道题先答了一半。11 月 4 日开源之后，第三方能不能复现这些基准，才是真正的验金石。
来源：Moonshot AI / VentureBeat、TNW 等报道 | 2026-10-23
瑞典政府办公厅封杀 ChatGPT 和 Claude：办公电脑只留一家过审的
瑞典政府办公厅 10 月 23 日宣布，从周一起公务员办公电脑将无法使用 ChatGPT 和 Claude，理由是两者未通过安全评估；微软 Copilot 是唯一过审的 AI 工具。瑞典税务署和机构关怀管理局此前已各自封禁，监察局也在考虑跟进。安全审查的一票否决，正在成为模型厂商在企业级市场之外的新战场——拿「企业级安全」当卖点的行业，被一个国家的办公系统整体拒之门外，比掉几个点的基准分疼得多。 > 💬 跑分决定发布会，审查决定采购单。对大模型厂商来说，能不能留在公务员的桌面上，正在变成一个比榜单名次更现实的商业问题。
来源：瑞典政府办公厅公告 / 北欧科技媒体等报道 | 2026-10-23
三次创业者 Naveen Rao 押注：推理十年万亿美元，机器人比编码智能体大 100 倍
Unconventional AI 创始人 Naveen Rao 上周五在 20VC 播客里给出三连判断：推理市场十年内会做到 10 万亿美元；三年内 80% 的 AI token 将跑在开源模型上，只有两成留给前沿实验室；机器人市场是编码智能体的 100 倍。支撑论据是一笔电力账：美国电网的备用容量几年内就会耗尽，而新增一吉瓦电力成本约 500 亿美元，比一座晶圆厂还贵——所以「未来十年，硬件才是好生意」。这位创业者履历堪称豪华：Nervana 卖给了 Intel，MosaicML 以 13 亿美元卖给 Databricks，第三次创业做的是围绕能效设计的 AI 硬件。 💬 连续把两家公司卖出高价的人看行业，最值得听的不是结论而是坐标系：当电力而不是算法成为稀缺资源，价值就流向把能源变成智能的那一层。这个判断如果成立，现在卷模型应用的公司，十年后都是在替「收电费的人」打工。
来源：20VC 播客 / Scuttlebutt 整理 | 2026-10-24
猫王音乐版权方起诉 Anthropic：274 首歌，清单开到了庭上
猫王 Elvis Presley 音乐版权的所有方（Authentic Brands Group 旗下）周五向法院递状，指控 Anthropic 未经授权使用 274 首猫王歌曲训练模型，清单包括《Can't Help Falling in Love》《Don't Be Cruel》《All Shook Up》等代表作。原告首席律师 Richard Busch 在诉状中写道，被告「以前所未见的规模把侵权建成了商业模式」。对生成式 AI 行业来说，这场案子的看点不在法理悬案，而在打法：版权方不再空谈「训练侵权」，而是拿出 274 首具体作品逐项索赔——案子越具体，和解的价码就越容易谈。 > 💬 内容授权的价目表正在被一场一场诉讼标出来。对模型厂商来说，与其等判决，不如把「买歌」当成和买算力一样的常规采购——只是这张清单会越来越长。
来源：Billboard / The Hollywood Reporter | 2026-10-25