AI日报 | 2026年11月3日：Anthropic 公开"递归自我改进"，Gemini 4 Carbon 低价登场
今日核心要闻
- Anthropic 发布《When AI builds itself》：AI 生成并合入生产代码占比 5 月已超 80%，约 26% 研发工作由 AI 自主推进
- 谷歌发布 Gemini 4 Carbon：面向长周期任务，输入 2 美元、输出 10 美元每百万 token 的入门定价
- Claude Code 禁止复刻既有游戏：系统提示新增规则，允许原创与同类型创作
- 42AI 携 6250 万美元出隐身：行为世界模型预测人在波动市况下的选择
- 四大实验室统一 robots.txt 训练退出信号：网站方一次配置，多家模型厂商同时读取
一、Anthropic 公开《When AI builds itself》，讨论递归自我改进
Anthropic递归自我改进Opus 5.5
Anthropic 在其研究所发布长文《When AI builds itself》，首次系统披露内部 AI 参与研发的真实比例，并把"递归自我改进"（Recursive Self-Improvement，RSI）摆上台面：当 AI 承担越来越多的 AI 研发工作，理论上最终可能出现能自主设计后继模型的系统。
- 2025 年 2 月 Claude Code 研究预览上线前，AI 编写并合入生产代码的占比仅为个位数；到 2026 年 5 月已超过 80%
- 约 26% 的研发工作由 AI 自主推进，人类仍充当质量闸门
- 官方明确表态：目前尚未达到完全自主设计后继模型的程度，RSI 也并非必然，但"可能比多数机构预期更早到来"
- 配合发布的 Opus 5.5 系统卡显示，已针对前沿大模型开发相关能力（如特定 ML 加速器上的内核开发）部署保护措施
编辑点评：一边是即将到来的 IPO 叙事，一边是递归自我改进的长期风险提示。Anthropic 把"安全"做成了差异化资产，也把自己放到了必须逐季度兑现的位置。
二、谷歌发布 Gemini 4 Carbon，主打长周期任务与低价
GoogleGemini 4 Carbon定价
谷歌发布Gemini 4 Carbon，称其为"面向长周期任务的最强模型"，并优先向网络安全防御场景开放，随后向开发者与消费者铺开。
- 入门定价：输入 2 美元、输出 10 美元每百万 token，与上一代 Argon 的入门价持平
- 在综合榜单上仍略落后于 Claude，但单位 token 成本更低
- 更便宜的前沿推理，正在改变机构级 agent 的运行经济学
- 谷歌同时宣布，Argon 的输出价格自 2026 年 12 月 1 日起由 10 美元上调至 20 美元每百万 token
行业影响：低价前沿模型让采购方始终保有两家可信供应商，也把竞争重心从"谁最强"挪到"谁的单位任务成本最低"。
三、Claude Code 新增规则：不得复刻特定既有游戏
Claude Code系统提示版权
Anthropic 悄然更新 Claude Code 与 Opus 5.5 的系统提示，新增两条规则，禁止复刻某款特定的既有电子游戏——即便它被改名或被称作原创。
- 原文：可以构建原创游戏、某一类型的游戏，以及基于用户自有设计与素材的游戏；不要构建对某款特定现有游戏的复刻
- 该变更在 10 月 27 日前后被开发者发现，官方未发布公告
- 背景是复古游戏圈用 AI 做反编译、重编译与移植的浪潮
- 游戏开发本身仍被允许：原创玩法、调试与工具链不受影响
风险提示：提示词层面的约束不是技术性版权过滤，无法覆盖模型"凭记忆重建"的情形。商业项目仍需要人工做版权合规审查。
四、42AI 结束隐身：6250 万美元做"行为世界模型"
42AI世界模型融资
以色列初创公司 42AI 结束隐身期，宣布获得 6250 万美元融资，推出所谓"行为世界模型"：估计一个人在特定情境下会做出哪种选择的概率。
- 首批客户是券商，用于评估市场剧烈波动时个人交易者的行为
- 与按人群画像的传统做法不同，模型试图刻画同一个人在不同激励、约束与经验下的差异
- 世界模型的边界从"预测物理世界"扩展到"预测人类决策"
编辑点评：当推荐、风控与投顾都开始"预测你会怎么选"，行为数据的合规边界可能比模型精度更早成为争议点。
五、四大实验室统一 robots.txt 的 AI 训练退出信号
OpenAIGoogleAnthropicrobots.txt
10 月 30 日，OpenAI、谷歌、微软与 Anthropic 共同宣布支持 robots.txt 中统一的 AI 训练退出信号，网站方不必再为每家爬虫单独配置。
- 该信号是自愿性质，各家实施节奏不同：Anthropic 表示已经开始支持，OpenAI 称将在未来数月落地，微软给出的生效时间是 2027 年 2 月 1 日
- 不覆盖"从实时网页内容生成回答"这一场景
- Meta 表态支持，但尚未承诺执行
仍需注意：统一信号降低的是沟通成本，真正的约束力取决于各家的履约与可审计性。对内容站来说，robots.txt 依然不是法律意义上的护城河。
今日数据速览
| 指标 | 数据 |
|---|---|
| AI 生成并合入生产代码占比（2026 年 5 月） | 超过 80% |
| AI 自主推进的研发工作占比 | 约 26% |
| Gemini 4 Carbon 输入价 | 2 美元 / 百万 token |
| Gemini 4 Carbon 输出价 | 10 美元 / 百万 token |
| Gemini 4 Argon 输出价（12 月 1 日起） | 20 美元 / 百万 token |
| 42AI 融资额 | 6250 万美元 |
| 统一训练退出信号宣布时间 | 2026-10-30 |
信息来源
- Anthropic Institute：《When AI builds itself》
- Google Blog：Introducing Gemini 4 Carbon
- Engadget：Anthropic updates Claude to stop it from recreating existing video games
- SiliconANGLE：42AI launches with $62.5M to predict how traders act when markets turn volatile
- SmartCR：The shared opt-out for AI training has limits
AI Daily | November 3, 2026: Anthropic Publishes Recursive Self-Improvement Research, Gemini 4 Carbon Lands Cheaper
Today's Top Stories
- Anthropic publishes "When AI builds itself": AI-authored code merged into production passed 80% by May 2026, with about 26% of R&D work driven autonomously
- Google launches Gemini 4 Carbon: built for long-horizon tasks at an introductory $2 input / $10 output per million tokens
- Claude Code bans recreating existing games: new system-prompt rules still allow original games and genre work
- 42AI exits stealth with $62.5M: a behavioral world model that predicts how people choose under volatile conditions
- Four labs back one robots.txt AI-training opt-out: sites express a single preference that multiple providers read
1. Anthropic publishes "When AI builds itself" and talks about recursive self-improvement
AnthropicRecursive self-improvementOpus 5.5
Anthropic's institute published a long piece, "When AI builds itself", disclosing for the first time how much of its internal R&D is actually done by AI, and putting recursive self-improvement (RSI) on the table: as AI takes on more AI development work, a system capable of autonomously designing its own successor could eventually emerge.
- Before Claude Code shipped as a research preview in February 2025, the share of AI-authored code merged into production was in the low single digits; by May 2026 it crossed 80%
- About 26% of R&D work is driven autonomously by AI, with humans still acting as the quality gate
- Anthropic says it is not there yet and that RSI is not inevitable, but it "could come sooner than most institutions are prepared for"
- The accompanying Opus 5.5 system card shows safeguards on capabilities tied to frontier LLM development, such as kernel work on certain ML accelerators
Editor's take: An IPO narrative on one side and a long-horizon warning about recursive self-improvement on the other. Anthropic has turned "safety" into a differentiator, and into a promise it now has to deliver every quarter.
2. Google launches Gemini 4 Carbon with long-horizon focus and low prices
GoogleGemini 4 CarbonPricing
Google released Gemini 4 Carbon, describing it as its most capable model for long-horizon tasks. It is being opened first to cyber-defense use cases before rolling out to developers and consumers.
- Introductory pricing: $2 per million input tokens and $10 per million output tokens, matching the introductory price of Argon
- It still trails Claude on aggregate benchmarks, but costs less per token
- Cheaper frontier inference is changing the economics of running agents at institutional scale
- Google also said Argon's output price rises from $10 to $20 per million tokens starting December 1, 2026
Industry impact: A cheap frontier model keeps two credible vendors in play for procurement, and shifts competition from "who is strongest" to "who costs least per completed task".
3. Claude Code adds a rule against recreating specific existing games
Claude CodeSystem promptCopyright
Anthropic quietly updated the system prompts for Claude Code and Opus 5.5 with two new rules that block recreating a specific existing video game, even if it is renamed or presented as original.
- The text: build original games, games in a genre, and games from the user's own designs and assets; do not build a recreation of a specific existing game
- Developers spotted the change around October 27; there was no official announcement
- It follows a wave of AI-driven decompilation, recompilation and porting in the retro-gaming community
- Game development itself is still allowed: original mechanics, debugging and tooling are unaffected
Risk note: A prompt-level rule is not a technical copyright filter and cannot cover cases where a model rebuilds something from memory. Commercial projects still need human copyright review.
4. 42AI exits stealth with $62.5M for a behavioral world model
42AIWorld modelFunding
Israeli startup 42AI emerged from stealth with $62.5 million in funding to build what it calls a behavioral world model: software that estimates the odds of each choice a person might make in a given situation.
- Its first customers are brokerages that want to know how individual traders will act when markets turn volatile
- Instead of demographic buckets, the model tries to represent how the same person behaves under different incentives, constraints and experience
- The world-model frontier is expanding from predicting the physical world to predicting human decisions
Editor's take: Once recommendation, risk control and advisory tools all start predicting what you will choose, the compliance boundary around behavioral data will become a controversy sooner than model accuracy will.
5. Four labs agree on a shared robots.txt AI-training opt-out
OpenAIGoogleAnthropicrobots.txt
On October 30, OpenAI, Google, Microsoft and Anthropic said they will honor a shared AI-training opt-out signal in robots.txt, so site operators no longer need to configure each crawler separately.
- The signal is voluntary and rollout is staggered: Anthropic says it has started implementing support, OpenAI expects it in the coming months, and Microsoft gave a February 1, 2027 effective date
- It does not cover answers generated from live web content
- Meta voiced support but has not committed to honoring it
Still worth noting: A shared signal lowers communication cost, but its real force depends on each provider's compliance and auditability. For publishers, robots.txt is still not a legal moat.
Key Numbers
| Metric | Value |
|---|---|
| AI-authored code merged into production (May 2026) | Over 80% |
| R&D work driven autonomously by AI | About 26% |
| Gemini 4 Carbon input price | $2 per million tokens |
| Gemini 4 Carbon output price | $10 per million tokens |
| Gemini 4 Argon output price (from Dec 1) | $20 per million tokens |
| 42AI funding | $62.5 million |
| Shared training opt-out announced | 2026-10-30 |
Sources
- Anthropic Institute: "When AI builds itself"
- Google Blog: Introducing Gemini 4 Carbon
- Engadget: Anthropic updates Claude to stop it from recreating existing video games
- SiliconANGLE: 42AI launches with $62.5M to predict how traders act when markets turn volatile
- SmartCR: The shared opt-out for AI training has limits