六月二十八日

Latent Space and AI Engineer cofounder Shawn "Swyx" Wang picked up Noam Brown's point about holding inference budget constant when reporting evals and gave it a twist: open models get far more dollar-per-token mileage than closed APIs, so anyone shipping an open model should report reasoning effort measured by dollars of inference on popular providers rather than by raw token count on the x-axis. It's a fairer way to compare "thinking levels" across open and closed systems.

Latent Space 与 AI Engineer 的联合创始人 Shawn "Swyx" Wang 接着 Noam Brown 那个观点展开：在报告 eval 结果时应保持 inference 预算恒定。他的延伸是，open model 每美元 token 的性价比远高于闭源 API，所以任何发布 open model 的人都应该用主流 inference 服务商上每美元的推理成本来衡量 thinking 等级，而不是在横轴上用原始 token 数。这样去比较开源与闭源系统的"思考程度"才更公平。

OpenAI's Codex and ChatGPT lead Thibault Sottiaux shipped a batch of Codex improvements: smoother handling of very long threads, a hoverable navigation rail to preview and jump between turns, expanded settings search, fixed tooltip and dialog misalignment at non-default zoom levels, and Markdown-preserving copy into Slack. "And most importantly: a dedicated Pets panel."

OpenAI 负责 Codex 与 ChatGPT 的 Thibault Sottiaux 一口气上线了一批 Codex 改进：更顺滑地处理超长对话线程、可悬停的导航栏方便预览和在各轮之间跳转、扩展了设置项搜索、修复了非默认缩放比例下 tooltip 和弹窗错位的问题、复制到 Slack 时保留 Markdown 格式。"而最重要的是：一个专门的宠物面板（Pets panel）。"

AI tutorials creator Peter Yang pushed back on a viral "problem-severity levels" framework, arguing that escalating only at level 5, after a fire has burned for days and your "solution" never actually fixed anything, leaves you far worse off than someone at level 1 who flagged the issue immediately and found a fix together: "everything has nuance." Separately, he showed his Hermes setup, which emails him a weekly health check by pulling from his Withings smart-scale API, Fitbit and Google Health, plus an MCP server and fitness app he vibe-coded.

AI 教程创作者 Peter Yang 反驳了一个很火的"问题严重程度分级"框架，他认为等到 level 5、火都已经烧了好几天、而你的"解决方案"根本没修好任何问题时才升级上报，处境其实比一上来就在 level 1 直接说出问题、然后大家一起找到解法的人糟得多，"凡事都有 nuance"。另外，他展示了自己的 Hermes 配置：每周给他发一封健康检查邮件，数据来自 Withings 智能秤的 API、Fitbit 和 Google Health，以及一个他自己 vibe coding 出来的 MCP server 加健身 App。

Linear head of product Nan Yu offered a contrarian one-liner on the same framework debate: "If 90% of problems you bump into aren't worth solving, then level 1 and level 6 are 90% the same."

Linear 产品负责人 Nan Yu 针对同一个框架争论抛出一句很反直觉的话："如果你碰到的问题里有 90% 根本不值得去解决，那么 level 1 和 level 6 其实有 90% 是一样的。"

Vercel CEO Guillermo Rauch warned that the Mythos / Sol class of AI cybersecurity tools is dual-use, as useful offensively as defensively, and that if adversaries get an equivalent offensive capability, US companies still unaware of their latent vulnerabilities are seriously exposed. His advice: proactively run deepsec or a similar harness with frontier models to find your own holes first.

Vercel CEO Guillermo Rauch 提醒，Mythos / Sol 这类 AI 网络安全工具是双刃剑，进攻能力和防守能力一样强；一旦对手拿到同等的进攻能力，那些还没意识到自身潜在漏洞的美国公司将面临严重威胁。他的建议是：主动用 deepsec 或类似的 harness 配合 frontier model，先把自己的漏洞找出来。

Box CEO Aaron Levie argued that AI token-cost optimization goes nowhere without a layer that deeply understands your actual workflows, context, and business processes, and that individual companies can't build that well at scale. That, he says, is "effectively the playbook for any applied AI company right now": eval models for the specific use case, master the domain, tune the UX, and support adoption through FDEs, so enterprises actually get more intelligence per dollar.

Box CEO Aaron Levie 认为，AI 的 token 成本优化离不开一个真正深入理解你实际工作流、上下文和业务流程的中间层，而单个公司很难自己把这件事做到规模化的好。他说这"基本就是当下所有 applied AI 公司的打法"：针对具体用例去 eval 模型、吃透领域、打磨 UX，再通过 FDE 支持落地，这样企业才能真正做到每一美元买到更多智能。

FirstMark VC Matt Turck posted a wry history of smart-glasses hype, walking from Google Glass in 2013 through Microsoft HoloLens, Meta, Apple Vision Pro, and Snap in 2026, with every pitch met by a polite "no" from everyone. A reminder of how stubbornly the category has resisted adoption no matter who tries.

FirstMark 的 VC Matt Turck 调侃式地梳理了一遍智能眼镜的炒作史：从 Google Glass（2013）一路到 Microsoft HoloLens、Meta、Apple Vision Pro，再到 Snap（2026），每一次推销都被所有人礼貌地回了句"不要"。这再次说明无论谁来尝试，这个品类都顽固地不被市场接受。

Builder Zara Zhang reflected on going from barely knowing how GitHub worked a year ago to 10k GitHub followers, all while still not writing code by hand, a marker of how far AI has lowered the barrier to shipping real side projects.

Builder Zara Zhang 回顾说，一年前她连 GitHub 怎么用都还搞不太清楚，如今已经有 1 万个 GitHub 粉丝，而且至今还不会手写代码，这恰恰说明 AI 把"做出真正能用的 side project"的门槛降到了多低。

Claude Code now supports artifacts — Claude Code can now capture an in-progress session as a live, shareable artifact: a web page (PR walkthrough, dashboard, incident timeline, release checklist) built from the full session context, including your codebase, your connectors, and the conversation itself. Pages update in place at the same link as the session works, with version history and a gallery to browse everything you've made. Anthropic says debugging is the most common internal use case: an engineer kicks off an incident investigation before standup, and Claude Code publishes a timeline, the suspect commits, and an error-rate chart, then republishes as the investigation progresses, so teammates "don't have to 'walk us through what the agent found' because they're all looking at the same view." Every artifact is private to its author by default and shareable only with authenticated members of your org (it cannot be made public), with admin access controls, retention policies, and a compliance API. It's available in beta to Claude Team and Enterprise orgs from the CLI and desktop app, viewable in any browser.

Claude 博客 —《Claude Code now supports artifacts》 — Claude Code 现在可以把进行中的会话沉淀成一个可分享的 live artifact：一个网页（PR 讲解、dashboard、事故时间线、发布清单），由完整的会话上下文生成，包括你的代码库、连接的各类工具，以及对话本身。随着会话推进，页面会在同一个链接上原地更新，并带有版本历史和一个可以浏览所有作品的 gallery。Anthropic 表示内部最常见的用例是 debug：工程师在 standup 之前启动一次事故排查，Claude Code 会发布一份时间线、可疑的提交和错误率图表，并随排查进展不断重新发布，这样队友"就不必再听我们复述 agent 发现了什么，因为大家看的是同一个视图"。每个 artifact 默认仅对作者私有，只能分享给经过身份验证的组织成员（无法公开），并配有管理员访问控制、留存策略和合规 API。目前以 beta 形式面向 Claude Team 和 Enterprise 组织开放，可在 CLI 和桌面端使用，任意浏览器都能查看。

The Takeaway: The next leap in useful AI won't come from a bigger frontier model, it'll come from models that keep training on your private context, so the more you work with one, the better it gets at your specific job.

Dan Biderman and Jessy Lin are cofounders of Engram, a research lab betting exclusively on memory and continual learning; Biderman came up through neuroscience. Their core claim cuts against frontier-lab orthodoxy: raw intelligence is no longer the bottleneck, understanding new, evolving, often private context is. Today people solve that with ever-larger prompts and context engineering, but Engram wants to bake context directly into model weights via lightweight adapter fine-tuning (LoRA, prefixes, sparse architectures), training a per-team model that knows a company "the way that an employee that's worked at your company for years has."

The payoff they pitch is dramatic: internalizing context can cut inference tokens by orders of magnitude. As Biderman puts it, the model can answer "within 100 tokens" what "the best frontier models would consume 100,000 tokens doing," because knowledge about people, teams, and priorities rarely lives in a single retrievable document. They frame this as a "rag killer," while admitting that what to internalize versus retrieve is still unsolved: you probably don't need last year's hotel room number in your weights, but your home password might be worth imprinting.

The most vivid technical aside is on memory efficiency: the KV cache for one Wikipedia article can balloon to roughly 80GB of GPU memory, yet a 70B model packs a lossy version of the whole internet into about 100GB of weights, proof that gradient descent compresses information far more tightly. Engram's bet is to spend compute offline compressing that cache by around 1000x. Their endgame: everyone carries a personalized model, and the skills you build at work travel with you to the next job.

要点：AI 真正变得好用的下一次飞跃，不会来自更大的 frontier model，而是来自能持续在你私有上下文上训练的模型，你用得越多，它就越懂你具体的工作。

Dan Biderman 和 Jessy Lin 是 Engram 的联合创始人，这家研究实验室把全部赌注押在 memory 和 continual learning 上；Biderman 出身神经科学。他们的核心观点与 frontier lab 的主流叙事相左：raw intelligence 已经不再是瓶颈，理解全新的、不断演变、而且往往是私有的上下文才是。如今大家靠越来越长的 prompt 和 context engineering 来解决这个问题，但 Engram 想通过轻量的 adapter fine-tuning（LoRA、prefix、sparse 架构）把上下文直接烤进模型权重里，训练出一个"像在你公司干了好几年的员工那样"了解这家公司的 per-team 模型。

他们描绘的回报相当惊人：把上下文内化可以把 inference token 砍掉好几个数量级。正如 Biderman 所说，模型能"在 100 个 token 内"回答出"最好的 frontier model 要消耗 10 万个 token 才能完成"的问题，因为关于人、团队和优先级的知识很少会写在某一份可检索的文档里。他们把这称为"rag killer"，同时也承认到底什么该内化、什么该检索仍是未解之题：去年住过的酒店房间号大概不必塞进权重，但你家门的密码也许值得印进去。

最生动的技术细节是关于 memory 效率：一篇 Wikipedia 文章的 KV cache 可能膨胀到约 80GB 的 GPU 显存，而一个 70B 模型却把整个互联网的有损版本压进了约 100GB 的权重里，这证明 gradient descent 能把信息压缩得紧凑得多。Engram 的赌注就是用离线算力把这份 cache 压缩约 1000 倍。他们的终局设想是：每个人都随身带着一个个性化模型，而你在工作中练就的技能也能跟着你走到下一份工作。