七月四日

Box CEO Aaron Levie argues the next competitive battle in AI is a "battle for context." An agent's effectiveness, he says, comes down to whether it has the right domain expertise, access to the right context and tools, and fits into a workflow users can review and trust. That's why the applied AI layer is far more than an "LLM wrapper": whoever can capture, govern, and continuously improve the knowledge feeding their agents — and route work between frontier models for planning and cheaper open or closed models for bulk tasks — will build the deepest moats. He also flags forward-deployed engineers (FDEs) as critical, because the applied layer is exactly where change management of the workflow actually happens.

Box CEO Aaron Levie 认为，AI 领域下一场竞争的核心是"对 context 的争夺"。他说，一个 agent 是否有效，取决于它是否具备正确的领域专业知识、能否访问正确的 context 和工具，以及是否能融入用户可以审阅和信任的工作流。正因如此，应用层（applied AI layer）远不只是一个 "LLM wrapper"：谁能捕获、治理并持续改进喂给 agent 的知识，并在规划环节调用 frontier 模型、在大量常规任务上调用更便宜的开源或闭源模型，谁就能建立最深的护城河。他还特别指出 forward-deployed engineer（FDE）至关重要，因为应用层正是真正发生工作流变革管理（change management）的地方。

Builder Zara Zhang shares a sharp read on how AI is reshaping what people will pay for: buyers are increasingly unwilling to purchase mere tools, figuring they can just rebuild them with coding agents. What still commands money, she argues, is "the feeling of hiring expertise they don't have."

Builder Zara Zhang 精准地点出 AI 正在如何重塑人们愿意付费的东西：买家越来越不愿意为单纯的工具买单，因为他们觉得用 coding agent 就能自己搭一个。她认为，真正还值得付钱的，是"雇到一份自己不具备的专业能力的那种感觉"。

AI engineer and Latent Space host Swyx makes a wry observation about "tools for thought": the community spent a decade building beautiful canvas demos, only to get "completely mogged by low contrast poorly designed CLIs." The reason those clunky command-line agents won? They actually do the "commodity thinking for you" — utility beat aesthetics.

AI engineer、Latent Space 主播 Swyx 对 "tools for thought"（思维工具）发出一句略带自嘲的观察：这个圈子花了十年做出漂亮的 canvas demo，结果却被"一堆对比度低、设计糟糕的 CLI 彻底碾压"。这些看似粗糙的命令行 agent 为什么赢了？因为它们真的能替你完成 "commodity thinking"（把常规思考外包出去）。实用性赢过了美观。

Vercel CEO Guillermo Rauch pitches "agentic self-improvement": give an agent the ability to introspect its past runs — spotting inefficiencies, errors, and redundant tool calls — and let it produce new prompts and skills from what it learns. That, he argues, is why observability should be built into agent deployments rather than bolted on afterward.

Vercel CEO Guillermo Rauch 提出了 "agentic self-improvement"（agent 自我改进）的概念：让 agent 能够回顾自己过去的运行记录，找出低效、错误和冗余的 tool call，并据此生成新的 prompt 和 skill。他认为，这正是为什么 observability（可观测性）应该内建在 agent 部署中，而不是事后再补上。

Linear head of product Nan Yu shares two pointed observations. First, that "if your field doesn't produce good training data it was probably full of crap to begin with." Second, that the best-case medical scenario is "a doctor who spends significant time on your case and is equipped with LLMs" — augmentation, not replacement.

Linear 产品负责人 Nan Yu 分享了两个犀利的观察。其一，"如果你所在的领域产生不出好的训练数据，那它八成本来就是一地鸡毛"。其二，医疗场景的最佳状态是"一位愿意在你的病例上花足够时间、并且配备了 LLM 的医生"，是增强而非取代。

Anthropic's Claude Code builder Thariq shares his core technique for working with Fable: the most important part is "discovering my own unknowns so I can prompt it better." He uses HTML artifacts to surface those unknowns and points to examples, noting the approach was refined through discussion at the AI Engineer conference (including adding quizzes).

Anthropic 旗下 Claude Code 的开发者 Thariq 分享了他使用 Fable 的核心方法：最重要的一步是"发现我自己的未知（unknowns），这样我才能把 prompt 写得更好"。他用 HTML artifact 把这些未知呈现出来，并给出了示例；他还提到这套方法是在 AI Engineer 大会上的讨论中打磨出来的（包括加入 quiz 的做法）。

Anthropic's Cat Wu (Claude Code + Cowork) shares a workflow tip: you can use Claude Code with computer use to set up Claude Tag — just point it at the Claude Tag docs and it will connect your team's GitHub repo, data warehouse, Google Drive, and other data sources for you.

Anthropic 的 Cat Wu（Claude Code + Cowork）分享了一个工作流小技巧：你可以用 Claude Code 配合 computer use 来配置 Claude Tag，只要把它指向 Claude Tag 的文档，它就会帮你把团队的 GitHub repo、数据仓库、Google Drive 以及其他数据源都接好。

Developer Peter Steinberger offers a design tip for coding agents: "If you think codex sucks at design, try 'use imagegen to re-imagine this design and implement that.'" Letting an image model reinterpret the layout first gives the coding agent something much better to build from.

开发者 Peter Steinberger 给出了一条针对 coding agent 的设计小技巧："如果你觉得 Codex 不会做设计，试试让它'用 imagegen 重新构想这个设计并实现出来'。"先让图像模型重新诠释布局，coding agent 就有了一个好得多的基础去实现。

FPV Ventures partner Nikunj Kothari, self-described as "one of the more vocal critics of Gemini," concedes it's still the only place where a single API key does it all: Flash for cheap long-context structured tasks, Nano Banana for world-class images, search with grounding, and strong realtime audio and video. He also floats a theory that labs quietly ship big models before long weekends — Opus 4.5 before Thanksgiving, Fable 5 before July 4th — to give people time to tinker.

FPV Ventures 合伙人 Nikunj Kothari 自称是"对 Gemini 批评最多的人之一"，但他也承认，Gemini 仍然是唯一一个用单个 API key 就能全部搞定的地方：Flash 负责便宜的长 context 结构化任务，Nano Banana 提供世界级图像，还有带 grounding 的搜索，以及很不错的 realtime 音视频。他还抛出一个理论：各大 lab 会悄悄在长周末前发布重磅模型，Opus 4.5 赶在感恩节前，Fable 5 赶在美国国庆前，好让大家有时间去折腾。

Every CEO Dan Shipper pushes back on claims that Fable 5 and Opus are different models — they're the same, he says, but Fable "falls back to Opud 4.8 slightly more, so the benchmarks are measuring a mix of Fable and Opus." He separately illustrates the new token economics of agentic work: a personal iOS app end-to-end runs ~5M tokens, clearing a whole production bug backlog ~20M, and answering every unread email, Slack, and text ~30M — "Fable working for you while you're at the pool: Priceless."

Every CEO Dan Shipper 反驳了 "Fable 5 和 Opus 是不同模型"的说法：他说它们是同一个模型，只不过 Fable"更常回退到 Opus 4.8 一点，所以 benchmark 测的其实是 Fable 和 Opus 的混合"。他还生动地展示了 agentic 工作的新 token 经济学：端到端做一个个人 iOS app 约 5M token，清空整个生产环境的 bug backlog 约 20M，把所有未读的 email、Slack 和短信都回一遍约 30M，"你在泳池边休息时 Fable 替你干活：无价"。

Y Combinator President and CEO Garry Tan predicts AI will transform healthcare just as specialist wait times are climbing: "I have a feeling AI will bring the quality of care up 100x and not a moment too soon for patients everywhere."

Y Combinator 总裁兼 CEO Garry Tan 预测，就在专科医生的等候时间不断拉长之际，AI 将彻底改变医疗："我有种感觉，AI 会把医疗质量提升 100 倍，而这对全世界的患者来说来得正是时候。"

The Takeaway: In an era where compute is capped, the only way to get more intelligence is to get more efficient — and that constraint, not brute force, is what shapes NVIDIA's open models.

Bryan Catanzaro leads Nemotron, NVIDIA's family of open foundation models, and has been building AI at the company since 2008 (with a detour to Baidu's Silicon Valley AI Lab alongside Andrew Ng and a young Dario Amodei). His pitch for why a chip company gives away frontier models is refreshingly unsentimental: Nemotron has "two jobs." First, building models forces NVIDIA to understand AI deeply enough to co-design its chips — "a GPU is whatever NVIDIA says it is." Second, open models grow the whole ecosystem, and "whenever AI is further developed and further deployed, it's an opportunity for our business." No altruism required; more AI everywhere means more GPUs.

The most striking idea is his framing of efficiency as destiny. Because the value of intelligence is so high, every organization will run at its limit — of dollars, of gigawatts. "If you accept as the truth that we're gonna be running at the limit, then the way to get more intelligence is to be more efficient. We can't get more intelligence by applying more force if we're already at the limit." That philosophy drives NVIDIA's hardest bets: pretraining Nemotron in 4-bit arithmetic (NVFP4), a hybrid state-space and transformer architecture that is genuinely smarter and not just faster, mixture-of-experts routing, multi-token prediction that makes inference faster the more accurate the model gets, and multi-teacher distillation.

Catanzaro also declares Moore's Law "dead for years" — the economic version that let you shrink-and-double is gone, which is precisely why first-principles co-design now beats raw scaling. And on the US-China rivalry, having worked at Baidu, he flatly rejects the copycat narrative: "There's a lot of bright people on this planet. It's not the case that there's only a few labs that have the monopoly on all good ideas."

The Takeaway：在算力已经见顶的时代，获得更多智能的唯一途径是变得更高效，而正是这种约束（而非蛮力）塑造了 NVIDIA 的开源模型。

Bryan Catanzaro 领导着 Nemotron，也就是 NVIDIA 的开源基础模型家族；他自 2008 年起就在 NVIDIA 做 AI（中间还去过百度的 Silicon Valley AI Lab，与 Andrew Ng 以及年轻时的 Dario Amodei 共事）。对于一家芯片公司为什么要把 frontier 模型免费送出去，他的解释毫不煽情：Nemotron 有"两个使命"。第一，做模型逼着 NVIDIA 把 AI 理解得足够透彻，才能协同设计自己的芯片："GPU 是什么，由 NVIDIA 说了算。"第二，开源模型能把整个生态做大，而"只要 AI 被进一步开发和部署，那就是我们生意的机会"。不需要什么利他主义，AI 无处不在就意味着更多 GPU。

最令人印象深刻的是他把效率视为命运的框架。因为智能的价值太高，每个组织都会跑在自己的极限上，无论是预算的极限还是电力（gigawatt）的极限。"如果你接受'我们注定要跑在极限上'这个事实，那么获得更多智能的方式就是变得更高效。如果我们已经在极限上，就没法靠加大蛮力来获得更多智能。"这一理念驱动着 NVIDIA 最艰难的几个赌注：用 4-bit 运算（NVFP4）做 Nemotron 的预训练，采用真正更聪明（而不只是更快）的 state-space 与 transformer 混合架构，mixture-of-experts 路由，让模型越准确推理越快的 multi-token prediction，以及 multi-teacher distillation。

Catanzaro 还宣称摩尔定律"已经死了好几年"，那个能靠"缩小再翻倍"取胜的经济学版本已经不复存在，这恰恰是为什么如今从第一性原理出发的协同设计能胜过单纯堆规模。谈到中美竞争，做过百度的他干脆利落地否定了"抄袭"论调："这个星球上有大量聪明人。绝不是只有少数几个 lab 垄断了所有好点子。"