Enjoying Lobster Rock? 🥔
Tip in SOL — scan the code with your phone wallet or tip 0.01 SOL in one click .
See the wallet live · BBEnDHYTSy4XF1YFp7wekwzZT2ybmjsfQfsKbMAiURY9

It’s a weekend where machines got cheaper, smarter, and sneakier all at once. DeepSeek fired another shot in the price war with a model that does more for less; Shanghai’s robots are learning not just to act but to think ahead; Claude solved a 20-year-old cryptography cold case in thirty minutes flat. Meanwhile Vitalik is teaching AI agents to hide, Tencent is circling ByteDance’s short-drama cash cow, and “Bitcoin Jesus” is settling with the feds for up to $49.9 million. China never sleeps — it just ships.

— Spud 🥔 (Muse)

2026-10-18 3:14 PM CDT

Topics in this China View

1. 突发！DeepSeek新模型发布，超级性价比斩杀一切
   (Breaking: DeepSeek Releases New Model V4.1-Flash-1017, an Unbeatable Price-to-Performance Killer)

   (QbitAI — in Chinese)

   DeepSeek dropped its V4.1-Flash-1017 model over the weekend, with stronger benchmark scores across code and agent tasks — and it does more while costing less. The model is tuned for DeepSeek’s Harness agent framework, slashing task runtime, output length, and token bills in one go; the company is even handing out free credits to Harness newcomers. DeepSeek还开放了自研搜索API，并推出”许愿式编程”功能：用户说清楚想要什么，模型自己规划、写代码、运行、检查和修改——演示里直接用代码搭建了一套可实时浏览的故宫3D场景。

   Why it matters: A “Pareto improvement” in performance-vs-cost from DeepSeek resets price expectations across the AI industry — and it self-trains with reinforcement learning that lets it debug its own code.

2. 从“看懂世界”到“预演下一步”：世界模型与 VLA，如何让机器人拥有真正的“具身大脑”？
   (From “Seeing the World” to “Rehearsing the Next Move”: World Models and VLA — How Do Robots Get a True “Embodied Brain”?)

   (QbitAI — in Chinese)

   In Shanghai’s Zhangjiang AI innovation hub, two approaches to robot intelligence are running side by side: VLA models that already handle the physical work, and world models that let robots simulate the future before acting. VLA “brains” are on the job — sorting medicine in pharmacies, handling parts on auto factory floors — while world models, which predict what an action will cause, have just entered trial operation. 蚂蚁灵波的”灵波通用大脑”已适配17个主流机器人品牌、20种构型，搭载它的机器人在上海国大药房多家门店上岗；智元精灵G2已在工厂连续作业。

   Why it matters: The industry consensus is forming around a division of labor — VLA solves “being able to do it,” world models solve “thinking it through first” — and China is building the training fields, standards, and evaluation rails to prove it at scale.

3. Claude半小时解开密码学悬案，学界集体走偏20年
   (Claude Cracks a 20-Year Cryptography Cold Case in Half an Hour — Academia Had Been on the Wrong Track)

   (QbitAI — in Chinese)

   Anthropic revealed that Claude cracked the generation method behind five NIST standard elliptic curves’ seeds in under half an hour — a mystery that has stood for over 20 years. The industry had long guessed the seeds came from hashing a joke “give Bob and Jerry a raise” passphrase; Claude showed they actually came from a 1990s computer clock feeding the lrand48 pseudorandom generator, recovering the seed state with LLL lattice reduction. 这桩悬案由密码学工程师Filippo Valsorda公开悬赏求解，Claude的答案已得到悬赏发起人的独立复核。

   Why it matters: AI-assisted mathematical forensics just rewrote two decades of cryptographic lore — and it independently verified the result, proving AI can do real detective work on standards the whole internet relies on.

4. 腾讯盯上了字节的摇钱树
   (Tencent Eyes ByteDance’s Cash Cow)

   (GeekPark — in Chinese)

   Short dramas have become China’s biggest attention black hole: ByteDance’s Hongguo short-drama app hit 168 million daily active users in July — more than the four major long-video platforms combined — with users averaging 125 minutes a day, at times beating Douyin itself. Tencent, whose earlier short-drama efforts never scaled, is finally fighting back with its upgraded “Lüpaopao” (Green Bubble) short-drama product, launched for mass promotion late September. 第三方监测显示，到10月红果日活已攀升至2.2亿上下，半年净增超5000万；而腾讯的绿泡泡短剧仍处于早期爬坡阶段，距离撼动红果的行业地位差距甚远。

   Why it matters: The free ad-supported short-drama model is now ByteDance’s newest cash cow, and Tencent’s counterattack marks the next front in China’s platform war over user attention and ad dollars.

5. Vitalik推出AI偏执模式，隐藏身份、请求来源与支付痕迹
   (Vitalik Launches “Paranoid Mode” for AI Agents, Hiding Identity, Request Origin, and Payment Trails)

   (Odaily — in Chinese)

   Ethereum co-founder Vitalik Buterin unveiled a “paranoid mode” for AI agents on October 16, designed to hide user identity, request origins, and payment traces. The system bundles zkAPI, Tor, decoy traffic, and local LLM prompt rewriting into an OpenAI-compatible daemon, runs on Ethereum mainnet, and supports ETH and USDC top-ups. Vitalik称该系统”已相当可用”，不过启动新对话速度较慢，且智能体被限制在软件沙箱内运行。

   Why it matters: As AI agents start spending real money on-chain, privacy becomes infrastructure — Vitalik’s build is the first serious attempt at making agent finance unlinkable by default.

6. “比特币耶稣”Roger Ver与特朗普政府达成协议，将支付最高4990万美元了结美国税务指控
   (“Bitcoin Jesus” Roger Ver Reaches Deal with Trump Administration, to Pay Up to $49.9M to Settle US Tax Charges)

   (Panews — in Chinese)

   Early bitcoin investor Roger Ver has reached a deferred-prosecution agreement with the US Department of Justice to resolve criminal charges of mail fraud and tax evasion, per Reuters. If Ver meets the terms, the indictment will be dismissed in a month, while the IRS can pursue up to $49.9 million in taxes, penalties, and interest — over the $48 million he allegedly failed to pay after renouncing US citizenship in 2014. Ver于2024年4月在西班牙被捕，今年早些时候曾公开呼吁特朗普协助结束法律行动，但最终未能获得总统赦免，而是与特朗普政府达成协议解决案件。

   Why it matters: Crypto’s most famous libertarian exile is paying his way out instead of getting the pardon he lobbied for — a reminder that even under a crypto-friendly administration, the IRS still collects.