AI News | Oct 18, 2026

AI News | Oct 18, 2026

Lede: Anthropic lobbies Congress with a safety bill it wrote itself, Meta quietly deletes "train" from its commercial terms, Gemini 4 Argon's ARC score halves under a different harness, DeepSeek's V4.1 Flash open update halves agent task cost, 90% of Claude Code is now written by Claude, Elorian debuts visual-reasoning model Glimpse, and hedge funds dump AI stocks five days straight — today is a triple act of legislation, benchmarks, and capital.

1. Anthropic Quietly Pitches Its Own Frontier AI Safety Bill to Lawmakers

Politico reported on Oct 16 — confirmed by Anthropic the same day — that the company is pitching lawmakers a frontier AI safety bill it drafted itself: only the Commerce Department, with a federal judge finding an "imminent" risk of over $1B in damages or 50+ deaths/serious injuries (limited to bioweapons, loss of control, or cyberattacks), could halt a model; developers must report incidents within 72 hours.

  • Summary: A rare sight: an AI lab writing the law it wants to be governed by. The trigger requires a Commerce Department request plus a federal judge's finding of "imminent" risk — over $1 billion in losses or 50+ deaths/serious injuries, confined to three categories: bioweapons, loss of control, and cyberattacks. Incident reporting to Commerce within 72 hours. Anthropic confirmed the draft's existence to Politico the same day.
  • Why it matters: The bar is so high it "may never trigger," critics say — a shield dressed as a safeguard. More subtle: the bill would bar other federal agencies from halting models independently. Whether that's locking in regulation or buying insurance is for Congress to decide.
  • Sources: The Hindu BusinessLine (via Reuters on Politico's reporting)

2. Meta Quietly Deletes the Word "Train" From Its Commercial Terms

On Oct 15, Nieman Lab exposed Meta's new Commercial Terms authorizing scraping of business accounts' websites for AI training with no opt-out; two days of backlash later, late on Oct 16, Meta quietly deleted the word "train" from the terms. A spokesperson said the merged crawling clause now applies only to ads and commercial products, with the robots.txt policy unchanged.

  • Summary: The Oct 15 Nieman Lab exposé: starting Oct 30, Meta's terms would let it crawl "any content" on business accounts' websites for AI training, no opt-out. Late Oct 16, Meta removed "train" from the clause; the spokesperson says the merged scraping provision covers only ads and commercial products, and the robots.txt policy stands.
  • Why it matters: A clean reversal of yesterday's story — AI companies keep probing the "license vs. scrape" boundary: write it in, watch the backlash, quietly delete it. Publishers won this round, but the scraping authorization itself remains in the terms. The next round won't be far off.
  • Sources: The Diary of AI (Oct 17 coverage)

3. Gemini 4 Argon Scores 97.5% on ARC-AGI-3 — Then 49.8% Under a Different Harness

ARC Prize's Oct 15 verified results: Gemini 4 Argon scored 97.5% (at ~$4.8K cost) under the "Provider Adapter Harness," becoming the second lab after OpenAI to cross the 95% bar — but dropped to 49.8% under the standard test framework. GPT-6 Astra fell from 99.9% to 62.7% as well.

  • Summary: ARC Prize's Oct 15 verified scores put Gemini 4 Argon at 97.5% under the Provider Adapter Harness (~$4,800 inference cost), making Google the second lab past 95% after OpenAI. Under the standard harness: 49.8%. GPT-6 Astra showed the same pattern, 99.9% to 62.7%. The Epoch index additionally shows Argon still trailing rivals in math and software engineering.
  • Why it matters: The same model can score twice as high depending on how much "scaffolding" the harness provides — the perfect case study for 2026's benchmark hype. The number doesn't speak; the harness does.
  • Sources: The Decoder

4. DeepSeek Releases V4.1 Flash 1017 Open Update: Agent Task Cost Halved Again

On Oct 17, DeepSeek released the V4.1 Flash 1017 snapshot with open weights (MIT license), just 5 weeks after the September release; official charts show DeepSWE at 77.0%, with per-task cost down from ~$0.50 to under $0.20.

  • Summary: DeepSeek's Oct 17 V4.1 Flash 1017 snapshot ships open weights under MIT, five weeks after the September release. Official charts claim 77.0% on DeepSWE, with per-task inference cost cut from about $0.50 to under $0.20 — nearly halving agent task cost.
  • Why it matters: China's open models keep doubling down on "cheap and capable" — a halved agent-task cost directly benefits every developer building agent applications. One caveat: the scores are self-reported and benchmarked against Opus 5, not the newer Opus 5.5; third-party verification still pending.
  • Sources: The Diary of AI (Oct 17 coverage)

5. Anthropic: ~90% of Claude Code Is Now Written by Claude

In an Oct 15 claude.dev post, Anthropic disclosed that roughly 90% of Claude Code's code is written by Claude itself; merged PRs doubled from June to September while rollback rates fell over 40%. Code review is now a seven-layer system (pre-PR validation, parallel AI reviewers, dedicated security review, automated checks, risk-tiered human approval, merge queue, fast rollback) — no longer "did a human look at it," but "which layer catches the risk."

  • Summary: Anthropic's Oct 15 claude.dev post: ~90% of Claude Code written by Claude; PR merges doubled Jun–Sep; rollback rate down over 40%. The review stack: pre-PR validation, parallel AI reviewers, security review, automated checks, risk-tiered human approval, merge queue, fast rollback. The companion Code Review for Claude Code is now GA for Team/Enterprise.
  • Why it matters: "AI writes code" is shifting from gimmick to process standard — Anthropic is productizing its internal methodology. When 90% of code is AI-written and rollbacks still fall 40%, the definition of human review in software engineering is being rewritten.
  • Sources: WebProNews (Oct 17 coverage)

6. Ex-DeepMind Researcher Andrew Dai's Elorian Debuts Glimpse: A "Visual Draft Space" for Mid-Reasoning Image Checks

On Oct 16, Elorian released its first model, Glimpse (research preview, API available): instead of "glance once, convert to text, then reason," it can crop regions, pull video frames, and write code to measure what it sees — mid-reasoning. A small team trained it on a few hundred GPUs; pricing at $0.60/$3.00 per million tokens, with self-reported CharXiv-R chart-reasoning scores above GPT-6 Astra.

  • Summary: Elorian — founded by ex-DeepMind researcher Andrew Dai — launched Glimpse on Oct 16 (research preview, API live). Unlike glance-then-reason pipelines, Glimpse revisits images mid-reasoning: cropping regions, extracting video frames, writing code to measure what it observes. Trained by a small team on hundreds of GPUs; priced $0.60/$3.00 per 1M tokens; self-reports beating GPT-6 Astra on CharXiv-R. Backed by Nvidia, with Jeff Dean involved.
  • Why it matters: Multimodal's pain point is "hallucinating what it can't see clearly" — counting, spatial relations. Glimpse makes "looking again" a core mechanism and attacks that pain point head-on. Star-team debut worth watching; scores are self-reported, so stay in spectator mode.
  • Sources: TechCrunch (Oct 16)

7. Hedge Funds Dump AI Stocks 5 Days Straight: Top 20 AI/Semiconductor Stocks Lose $1T+

Goldman Sachs Prime data shows hedge funds net-sold global AI stocks for five straight trading days (Oct 8–14), erasing all cumulative net buying since early 2025; the 20 largest AI and semiconductor stocks shed over $1 trillion in combined market value; the US semiconductor sector fell 8 days straight — the longest streak since 2008.

  • Summary: Goldman Prime data: five consecutive sessions of net selling in global AI stocks (Oct 8–14), wiping out all net buying since early 2025. The 20 largest AI/semiconductor names lost over $1 trillion combined; US semis fell 8 days in a row, the longest since 2008. Broadcom CEO Hock Tan: "capital constraints, not capacity, will be AI growth's biggest bottleneck."
  • Why it matters: Breakthroughs keep coming, but markets are questioning capex returns — the AI narrative is decoupling from the market. Builders should watch the money too.
  • Sources: Sina Finance (Oct 18 morning, citing Goldman data)

One-line takeaway: Today's AI news is a triple act of legislation, benchmarks, and capital — Anthropic tries to write its own rules, Meta is forced to rewrite its terms, Gemini's scores expose the rules of scoring itself, and markets vote with their feet: the tech is sexy, but the bill is real.

中文版：https://cutting-edgetechnology-ai.blogspot.com/2026/10/ai-20261018.html

评论

发表评论

此博客中的热门博文

每日 eSIM | Orange Travel：法国官方出品，带真实号码、能收短信的全业务 eSIM

图片
h1>每日 eSIM | Orange Travel：法国官方出品，带真实号码、能收短信的全业务 eSIM 「每日 eSIM」每天介绍一款 pay as you go / 预付费 / 有免费试用的 eSIM。只讲干货：多少钱、能不能收验证码、适合谁。今天第八期，主角是法国电信巨头 Orange 官方的旅游 eSIM——Orange Travel，200+ 目的地一张卡搞定，关键是它带真实电话号码。 Orange Travel 官网：覆盖 200+ 目的地，流量+通话+短信三合一（travel.orange.com 官网素材） 一句话定位 Orange Travel 是法国 Orange 官方直营的旅游 eSIM（不是第三方转售），覆盖 200+ 国家和地区。跟 Airalo、Nomad 这类纯流量卡最大的区别：它是 全业务卡 ——流量、通话、短信三合一，还带一个 法国 +33 真实手机号码 ，套餐可续费充值，用完再买，没有月租。 套餐类型：预付费固定套餐，按有效期分档 按有效期分档 ：7 天 / 14 天 / 30 天 / 45 天 / 90 天多档可选，固定流量包，用完流量或到期即止； 全业务捆绑 ：各档套餐一般含欧洲境内无限通话和短信，另加拨往世界各地的分钟数和短信条数（档位越高越多）； 多国/全球包 ：除单国包外还有 World 系列多国包，一次买断跨多国使用； 可续费 ：套餐到期或流量用完可直接在账户里续费/充值延长，不用重新装卡。 Orange Travel App：选目的地、买套餐、续费充值都在一个 App 里搞定（travel.orange.com 官网素材） 价格参考（截至 2026年10月，esimdb 数据） 入门：5GB/30天 约 $14.99； 常用：20GB/14天 约 $29.99，20GB/30天 约 $26.99（欧洲区）； 大包：100GB/30天 约 $82.99；500GB/90天 约 $119.99。 价格因目的地差异很大（单国包便宜、多国包贵），且官网常有折扣，下单前以 travel.orange.com 实时报价为准。 关键问题：能收短信验证码吗？ 能。 这是 Orange Travel 相对纯流量 eSIM 最大的优势。esimdb 2026 年 1...
阅读全文

AI 大事件 | 2026年10月6日

今天的 AI 世界很不平静：美国新锐 Reflection AI 发布了首款开放权重模型 Beam，叫板中国开源模型；彭博报告显示美中 AI 性能差距已缩小到历史最低的 3%；纽约市议会把 OpenAI、Anthropic、谷歌、Meta 拉上听证席，追问 AI 的法律责任。以下是过去 24 小时最值得关注的 AI 大事件。 1. Reflection AI 发布开放权重模型 Beam，对标中国开源模型 Nvidia 投资的美国新创公司 Reflection AI 于 10 月 5 日正式发布首款开放权重模型 Beam，这是一款拥有 5010 亿总参数、每次推理仅激活 230 亿参数的混合专家（MoE）文本模型。该公司声称 Beam 在高级推理基准上的表现可媲美智谱的 GLM-5.2，推理算力需求仅为同类西方开放模型的三分之一到四分之一。 为什么重要： 这是美国方面对 DeepSeek、Qwen、Kimi 等中国低成本开源模型最有力的回应之一。如果效率主张成立，Beam 将大幅拉低企业和开发者部署前沿 AI 的算力成本，重塑开源模型的竞争格局。不过目前所有基准成绩均为公司自报，开放权重需到本月晚些时候才会发布，尚待独立验证。 来源： TechCrunch 2. 彭博：美中 AI 性能差距缩小至 3%，创历史最低 据彭博社 10 月 5 日报道，彭博行业研究（Bloomberg Intelligence）最新报告指出，自 DeepSeek 于 9 月发布 V4.1 Flash 后，中国顶尖模型与美国竞争对手在基准测试得分上的差距已缩小至仅 3%，而 5 月时约为 9%，今年早些时候为 15%。 为什么重要： 这意味着美国在模型能力上的传统领先优势几乎被抹平，AI 竞争正进入"毫厘之争"阶段。对企业而言，选型时地缘、成本、生态的权重将超过单纯的性能比较；对政策制定者而言，出口管制和算力封锁的效果正受到严峻考验。 来源： 新浪科技转载观察者网 3. 纽约市议会举行 AI 企业宣誓作证听证会，审议 10 项 AI 法案 10 月 5 日（当地时间），纽约市议会举行全体委员会听证会，Anthropic、OpenAI、谷歌和 Meta 的代表宣誓作证。议长朱莉·梅宁追问：如果 AI 造成重大经济损失或人员死亡，企业是否承担法律责任...
阅读全文

出海贸易 | eu.org 免费域名申请全攻略（2026 版）

做海外业务，域名是门面：品牌官网、域名邮箱、支付回调（webhook 必须挂在域名上）都离不开它。平台送的二级域名处处受限—— 换平台带不走，https 权限也不在你手里。 eu.org 是 1996 年起运营的公益免费域名，至今仍可申请，但它是“审核制 + 全人工”，懂得避坑才能拿到。 一、先说结论 仍免费 ：nic.eu.org 正常运营，申请不收费（截至 2026 年 10 月 10 日）。 审核无官方时限 ：社区反馈从几小时到数年不等，急着用的别拿它赶 deadline。 最大坑 ：官网明确警告——Gmail 会拒收它的验证邮件，注册邮箱别用 Gmail。 申请前必须先配好 NS ：eu.org 只处理域名委托记录，没有 NS 的申请等于白交。 配 Cloudflare 完全可行：改 NS 接入，免费 SSL 自动签发。 二、申请前先搞懂规则 按 nic.eu.org 官网政策整理： eu.org 面向“付不起域名费的用户和非营利组织”； 只处理 NS 委托记录 ——填表前，你的 NS 服务器必须已经配好并能解析； 官方建议 别直接挂 eu.org ，优先选开放三级域，如 ASSO / EDU / INT / NET.eu.org（直接挂限制更多，先读 direct subdomain 政策）； 没有外部注册商体系，域名只能在官网 arf 界面管理， 基本无法转出（待核实） 。 三、手把手：六步申请流程 按官网流程（nic.eu.org/register.html）整理。精确点击路径未实操， 以官网当前页面指引为准 。 第 1 步：读政策，选对后缀 打开 nic.eu.org/register.html，读通用政策（general policy）； 按政策 选开放三级域 （如 example.NET.eu.org）——官方明确建议；实在想直接挂 eu.org，先读 direct subdomain 政策再决定。 第 2 步：先配好 NS（申请前必须） eu.org 只处理域名委托记录，没有 NS 的申请等于白交； 注意 ：Cloudflare 不能提前添加未注册域名。社区经验是先在 DNS.he.net、DNSPod 这类允许“先加 NS 记录”的解析商处配好 NS，注册成功后再迁到 Cloud...
阅读全文