八月二十二日

Anthropic says Claude Security can now point Mythos 5 at a GitHub repository, trace data across files, reason about component interactions, and return findings with CWE categories, confidence and severity ratings, plus suggested fixes. The model stays behind the scan and exposes findings rather than direct model access; suggested patches open in Claude Code on the web, and scans use standard token billing under an existing plan. Anthropic is also putting $35 million in credits behind open-source security through its Defender Advantage Fund, working with partners to embed Mythos 5 in security products and services, and preparing an expansion of its Cyber Verification Program.

Anthropic 表示，Claude Security 现在可以让 Mythos 5 扫描 GitHub repository，跨文件追踪数据流、推理组件之间的交互，并为每个发现返回 CWE category、confidence、severity rating 与修复建议。模型本身留在扫描流程背后，只输出 findings，而不向用户开放直接访问；建议 patch 会在 Claude Code web 端打开，扫描则按现有 plan 的标准 token usage 计费。Anthropic 还通过 Defender Advantage Fund 为 open-source security 提供 3500 万美元 credits，与合作伙伴把 Mythos 5 集成进 security products and services，并准备在未来几周扩展 Cyber Verification Program。

Meta Senior Director of AI Madhu Guru warned that compressing a complex eval suite into one score creates a "tyranny of the average." His example shows why: a model can improve from 85% to 89% on simple summarization and from 80% to 85% on basic factual QA while falling from 70% to 63% on complex financial analysis, leaving the frontier use case worse even as an aggregate looks better. He argues that weighted scores merely add false mathematical precision to a judgment call; teams should instead prioritize evals, inspect critical results and failure modes in depth, and decide against the needs of their actual users.

Meta AI Senior Director Madhu Guru 警告说，把复杂的 eval suite 压缩成一个分数，会制造“tyranny of the average”。他的例子清楚说明了问题：一个模型可能在 simple summarization 上从 85% 提升到 89%，在 basic factual QA 上从 80% 提升到 85%，却在 complex financial analysis 上从 70% 下滑到 63%；即使 aggregate score 变好，最前沿的 use case 反而退步。他认为 weighted score 只是给判断题套上一层虚假的数学精确度；团队应该为 evals 排定优先级，深入查看关键结果与 failure modes，再根据真实用户需求做决定。

Swyx changed his view on the claim that simulation is a new scaling law. His reasoning is that if models automate progressively larger parts of ML research and AI engineering, simulating humans and human feedback becomes one of the remaining barriers to recursive self-improvement. He points to Simile and the earlier Smallville research lineage, noting that the work once had no commercial application but is now finding product-market fit at Fortune 100 companies; that commercial traction is his evidence that human simulation is moving from research curiosity toward infrastructure.

Swyx 改变了自己对“simulation is a new scaling law”这一说法的看法。他的推理是：如果模型正在自动化越来越多的 ML research 与 AI engineering，那么模拟人类及 human feedback，就会成为 recursive self-improvement 剩余的关键障碍之一。他把 Simile 与早期 Smallville research lineage 联系起来，指出这项工作当初没有商业应用，如今却已在 Fortune 100 companies 中找到 product-market fit；对他而言，这种商业 traction 说明 human simulation 正从研究趣闻走向基础设施。

Peter Yang initially praised Instinct for unusually smooth onboarding across iMessages, Google Workspace, and MCP connections, plus proactive suggestions immediately after a tool is connected. He still saw its single-thread interface as unsuitable for sustained work and preferred ChatGPT Work and Codex for that. More seriously, he later withdrew his recommendation after learning that Instinct indexed and retained his emails without permission and offered no way to delete them, turning a product UX review into a concrete warning about data control in connected assistants.

Peter Yang 最初肯定了 Instinct 的 onboarding：连接 iMessages、Google Workspace 与 MCP 的过程非常顺滑，而且 tool 一接入，assistant 就会主动建议可以完成的任务。不过，他认为 single-thread interface 不适合持续工作，因此实际工作仍会选择 ChatGPT Work 和 Codex。更严重的是，他随后得知 Instinct 未经许可就索引并保留了自己的 emails，而且没有删除入口，于是撤回推荐；一次 product UX 评测也因此变成了对 connected assistants 数据控制问题的具体警告。

Vercel CEO Guillermo Rauch said his team repeatedly ran `is-agentic` against a linked sandbox until the result reached 100/100, and that the loop exposed several product gaps they then closed. The noteworthy engineering pattern is to turn agent compatibility into an executable acceptance test rather than a qualitative claim, while keeping the criteria demanding enough to justify users' time and tokens. Rauch also said a linked Vercel tool now supports Grok and Codex subscriptions and can be installed instantly on a sandbox.

Vercel CEO Guillermo Rauch 表示，团队反复让 `is-agentic` 对一个 sandbox 执行测试，直到结果达到 100/100；这个循环暴露了多个产品缺口，并推动团队逐一补上。值得注意的工程模式，是把 agent compatibility 变成可执行的 acceptance test，而不是停留在定性宣传，同时让评测标准足够严格，配得上用户投入的时间与 tokens。Rauch 还表示，一款他链接的 Vercel 工具现已支持 Grok 与 Codex subscriptions，并可立即安装到 sandbox 中。

OpenAI's Thibault Sottiaux said some Codex users saw worse cache hit rates this week than in the preceding stable period, which could explain why their usage allowances drained faster because consistent cache hits are important to efficiency. The team is investigating. He also confirmed that a banked reset landed for all paid ChatGPT Work and Codex users, separating an immediate quota adjustment from the unresolved cache-efficiency issue.

OpenAI 的 Thibault Sottiaux 表示，部分 Codex 用户本周的 cache hit rate 低于此前稳定时期；由于持续命中 cache 是提高效率的重要条件，这可能解释了为什么这些用户的 usage allowance 消耗得更快。团队仍在调查。他还确认，所有 ChatGPT Work 与 Codex 付费用户的 banked reset 已经上线，从而把即时的 quota adjustment 与尚未解决的 cache-efficiency 问题区分开来。

Anthropic's Thariq shared an internally popular Claude Code skill called ELI5. The command asks Claude to explain a topic for a complete beginner using an HTML artifact with large visuals and few words; he uses it to understand how a module works, why a tradeoff was made, or what caused an incident. It is not yet an official plugin, but users can try it through Anthropic's community plugin marketplace, making it a compact example of packaging a repeatable explanation style as an installable development workflow.

Anthropic 的 Thariq 分享了一个团队内部常用的 Claude Code skill，名为 ELI5。这个 command 会要求 Claude 面向完全不了解主题的人，用大图、少量文字的 HTML artifact 来解释问题；他会用它理解 module 如何工作、某项 tradeoff 为什么成立，或 incident 的根因是什么。它目前还不是 official plugin，但用户已经可以通过 Anthropic community plugin marketplace 试用；这也是一个很紧凑的例子，展示如何把可复用的解释方式封装成 installable development workflow。

FPV Ventures partner Nikunj Kothari described a small but revealing agent use case: he asked Claude Code to inspect network requests behind a poorly structured school-meal website, discover an unauthenticated API, determine its format, and connect the result to an existing Hermes home bot. The bot now reports breakfast and lunch each morning so the family can decide what food to pack. The value came from stitching an obscure data source into an existing routine, not from launching another standalone assistant.

FPV Ventures partner Nikunj Kothari 分享了一个很小、却很能说明问题的 agent use case：他让 Claude Code 检查一个结构混乱的学校餐食网站背后的 network requests，找到一个无需认证的 API，判断其数据格式，再把结果接入现有的 Hermes home bot。现在，这个 bot 每天早上都会播报 breakfast 与 lunch，方便家人决定要不要另外带饭。它的价值不在于再推出一个 standalone assistant，而在于把冷门数据源缝进已有的生活流程。

Box CEO Aaron Levie argues that models are simultaneously becoming cheaper for equivalent tasks, more capable, faster, and deeper across domains. His conclusion is that the main opportunity shifts toward diffusing AI through the economy, while model innovation and competition become a tailwind for applied AI companies rather than merely a threat. For startups, the strategic question is increasingly how quickly they can convert falling intelligence costs into useful deployment.

Box CEO Aaron Levie 认为，模型正在同时变得更便宜、更通用、更快，并深入更多专业领域。他的结论是，主要机会正在转向如何让 AI 扩散进整个经济体系，而 model innovation 与 competition 对 applied AI companies 来说，会更像顺风而不只是威胁。对 startups 而言，战略问题越来越变成：能否快速把不断下降的 intelligence cost 转化为有用的 deployment。

The takeaway: Chess.com is evidence that superhuman machines can increase rather than erase demand for human skill. CEO Erik Allebest says computers first pushed elite players toward technically perfect but boring styles, while neural-net engines such as Leela Chess Zero later produced aggressive, unconventional play that expanded what humans could learn from the game. As he puts it, "fundamentally, humans want to do human stuff": compete with other people, make things, and solve problems. AI now supports game review, personalized puzzles, research, internal analytics, support automation, and experiments with an on-demand coach that reviews a player's recent games and recommends what to improve.

核心结论是：Chess.com 证明了 superhuman machines 不一定会抹去人类技能的价值，反而可能提升人们对它的需求。CEO Erik Allebest 表示，早期 chess engines 一度把顶尖棋手推向技术完美却乏味的风格；后来，Leela Chess Zero 等 neural-net engines 展现出激进、非传统的下法，反而拓宽了人类可以从棋局中学习的空间。正如他所说，“fundamentally, humans want to do human stuff”：与其他人竞争、创造东西、解决问题。如今，AI 已被用于 game review、personalized puzzles、research、internal analytics、support automation，以及一项 on-demand coach 原型；它会查看玩家最近的棋局，并建议下一步该改进什么。

Allebest's broader product lesson is that access to intelligence does not remove the need for practice. AI can act like a professor available in your pocket and dramatically accelerate research and learning, but expertise still comes from repeated puzzles, reviewing mistakes, and compounding daily habits. Chess.com itself grew without primary outside capital to more than 250 million registered members, roughly 10 million daily active users, and projected annual revenue above $200 million; the company is now applying its ratings-and-learning playbook to poker through Gambit while using agentic development to shorten the path from identifying a problem to shipping code.

Allebest 更广泛的产品启示是，获得 intelligence 并不会取消练习的必要。AI 可以像一个随时装在口袋里的教授，大幅加速 research 与 learning，但真正的 expertise 仍来自反复解题、复盘错误，以及长期累积的日常习惯。Chess.com 在没有 primary outside capital 的情况下，已经发展到超过 2.5 亿 registered members、约 1000 万 daily active users，并预计年收入超过 2 亿美元；公司如今正通过 Gambit 把 ratings-and-learning playbook 带到 poker，同时用 agentic development 缩短从发现问题到发布代码的路径。