七月十三日

Thibault Sottiaux posted a detailed usage update for paid subscribers built around GPT 5.6 Sol, framing it as "no nerfing, only good stuff." He says inference optimizations are being passed straight through as roughly 10% more usage, and explains that quietly raising the product's context limit to 372k (up from 272k) had been draining subscriptions faster than intended, so they've reverted to 272k while they fix the rollout. He also rolled back some behind-the-scenes "juice" (reasoning-effort) changes, is trimming over-eager multi-agent usage at high and xhigh effort, and is temporarily leaving the 5-hour limit off. https://x.com/thsottiaux/status/2076495156757577895 He separately reassured users that GPT 5.6 Sol will stay in the Go, Plus, and Pro plans they pay for, "at least until we ship an even better model." https://x.com/thsottiaux/status/2076459871021736245

OpenAI 负责 Codex 与 ChatGPT 的 Thibault Sottiaux 发布了一条围绕 GPT 5.6 Sol 的付费用户额度更新，并强调"只有好消息，绝不缩水"。他说推理端的优化会原封不动地转化为大约 10% 的额度增加；同时坦承此前悄悄把产品的上下文上限从 272k 提到 372k，导致额度消耗比预期更快，因此先回退到 272k，等把发布流程修好再重新推出。他还撤销了一些幕后的"juice"（推理强度）调整，正在收敛 high 与 xhigh 强度下过于积极的 multi-agent 用量，并暂时保持 5 小时限制不生效。https://x.com/thsottiaux/status/2076495156757577895 他另外向用户保证，GPT 5.6 Sol 会继续留在你付费的 Go、Plus 和 Pro 套餐里，"至少在我们推出更强的模型之前是这样"。https://x.com/thsottiaux/status/2076459871021736245

Anthropic, via its official Claude account, announced it's extending access to Claude Fable 5 on all paid plans and keeping Claude Code's weekly rate limits 50% higher through July 19. As before, users can spend up to half their weekly usage limit on Fable 5; after that they can either continue with usage credits or switch to another model to keep working within their remaining limits. https://x.com/claudeai/status/2076351399999557669

Anthropic 通过其官方 Claude 账号宣布：延长所有付费套餐对 Claude Fable 5 的访问权限，并将 Claude Code 的每周速率上限维持在高出 50% 的水平，一直到 7 月 19 日。和此前一样，用户最多可以把每周额度的一半用在 Fable 5 上；用完之后，可以继续用 usage credits，或者切换到其他模型，在剩余额度内继续工作。https://x.com/claudeai/status/2076351399999557669

Aaron Levie argues that one of the defining business questions of the century is how you maximize your corporate IP — decisions, insights, workflow patterns, best practices — in a world where so much intelligence is packed into AI models. Counterintuitively, he says these questions don't get "bitter-lessoned" out of existence; they become more germane as intelligence gets more powerful, because when every firm has frontier intelligence, your edge is how you uniquely leverage it. That's why he sees so much value left to create in the applied-AI layer: evals for your workflows, routing across model tiers, capturing traces that improve those workflows, and making your information compound as AI gets better. "The companies that help figure this out for other enterprises will be in the best position to win the next enterprise workloads." https://x.com/levie/status/2076338364635287637

Box CEO Aaron Levie 认为，本世纪商业最核心的问题之一，是在如此多智能被封装进 AI 模型的世界里，你如何最大化自己的企业 IP——那些决策、洞见、工作流模式和最佳实践。反直觉的是，他说这些问题并不会被"苦涩的教训"（bitter lesson）抹平；相反，随着智能越来越强，它们只会变得更加关键，因为当每家公司都能用上前沿智能时，你的优势就在于如何独一无二地驾驭它。正因如此，他认为应用层（applied AI）还有巨大的价值有待创造：为你的工作流建立 evals、在不同智能档位的模型之间做路由、捕获能反过来改进工作流的 traces，并让你的信息价值随着 AI 变强而不断复利。"那些帮别的企业把这件事搞明白的公司，将最有机会赢得下一波企业级工作负载。"https://x.com/levie/status/2076338364635287637

Guillermo Rauch makes the case for owning your AI stack rather than renting it: "Make the model a cog in a machine you own." He maps it onto open layers — the AI SDK as an open model API, an open Agent API, and the AI Gateway for open zero-data-retention inference — and argues that startups and enterprises must own their data, evals, model choices, and software layer. His blunt framing: "Don't outsource your brain." https://x.com/rauchg/status/2076364176252191222

Vercel CEO Guillermo Rauch 主张要"拥有"而非"租用"你的 AI 技术栈："把模型变成一台你自己拥有的机器里的一个齿轮。"他把这套思路对应到几个开放层上——AI SDK 作为开放的模型 API、一个开放的 Agent API、以及用于开放式零数据留存（ZDR）推理的 AI Gateway——并强调初创公司和企业必须掌握自己的数据、evals、模型选择和软件层。他直白地总结道："别把你的大脑外包出去。"https://x.com/rauchg/status/2076364176252191222

Amjad Masad shared a "Vibe Research" experiment: fine-tuning a Qwen-8b model to play chess on Replit, running three parallel branches with different experiments and making real progress. His broader point is that models have gotten dramatically better at doing ML itself — work they used to be genuinely bad at — so someone with good intuition to guide the process can now do interesting ML work even if they've never done it before. https://x.com/amasad/status/2076227936202662357 He also showed Replit's own computer-use model playing against his new chess engine. https://x.com/amasad/status/2076356893736673507

Replit CEO Amjad Masad 分享了一个他称为"Vibe Research"的实验：在 Replit 上 fine-tune 一个 Qwen-8b 模型来下国际象棋，同时跑三个平行分支做不同的实验，并取得了实实在在的进展。他更大的观点是，模型在"做 ML 本身"这件事上已经进步惊人——这曾是它们真正很差的领域——所以如今哪怕一个从没做过 ML 的人，只要有好的直觉来引导过程，也能做出有意思的 ML 工作。https://x.com/amasad/status/2076227936202662357 他还展示了 Replit 自家的 computer-use 模型对战他新写的国际象棋引擎。https://x.com/amasad/status/2076356893736673507

Peter Yang argues that when community sentiment turns against you, the instinct to communicate less and in a more corporate way is exactly backwards — the right move is to get more human, be transparent about what's going on, and figure out solutions together with users in the open. https://x.com/petergyang/status/2076512796481880270 He points this specifically at Anthropic, saying they make great models but should engage the community as directly as OpenAI has been doing. https://x.com/petergyang/status/2076510899490480228 He also floats a rough read on model adoption: his wild guess is that 90%+ of people are using GPT 5.6 Sol and under 10% are on Terra or Luna. https://x.com/petergyang/status/2076519927843000448

产品负责人、newsletter 作者 Peter Yang 认为，当社区情绪转向对你不利时，那种"少说话、说得更官腔"的本能恰恰是错的——正确的做法是变得更有人味、坦诚地说明正在发生什么、和用户一起把问题解决掉。https://x.com/petergyang/status/2076512796481880270 他把这番话直接指向 Anthropic，说他们模型做得很好，但应该像 OpenAI 那样更直接地和社区互动。https://x.com/petergyang/status/2076510899490480228 他还大胆估了一下模型的采用分布：他猜 90% 以上的人在用 GPT 5.6 Sol，只有不到 10% 在用 Terra 或 Luna。https://x.com/petergyang/status/2076519927843000448

Nikunj Kothari pushes back on the SF "tokenmaxxing" culture — the people bragging about swarms of subagents looping things for them in the background. When he asks what they're building and for whom, very few can give a straight answer, which tells him that even in this insane AI era, simplicity and direction are still ridiculously important. His advice before you let your tokens "go brrr": stop and think about whether what you're building actually matters, because "time is literally the only thing you don't get back." https://x.com/nikunj/status/2076458876816540144

运营者、投资人 Nikunj Kothari 对旧金山那种"tokenmaxxing"文化提出了反驳——那些炫耀自己后台有一堆 subagent 在循环干活的人。当他问他们"在给谁、造什么"时，很少有人能给出清晰的答案；这让他意识到，哪怕在这个疯狂的 AI 时代，简单和方向依然重要得离谱。他给出的建议是：在你让 token"哐哐烧"之前，先停下来想想你正在做的东西到底重不重要，因为"时间才是唯一你拿不回来的东西"。https://x.com/nikunj/status/2076458876816540144

Zara Zhang shares a workflow she calls "meeting transcript as PRD": she talks through a feature's implementation with a colleague, sends the raw transcript to Codex, and it builds the prototype exactly as they discussed. Her line for it: "The meeting is the prompt." https://x.com/zarazhangrui/status/2076300222884626754

投资人、builder Zara Zhang 分享了一个她称之为"把会议记录当 PRD"的工作流：她和同事把某个功能的实现方案聊一遍，把原始的会议记录直接发给 Codex，它就会照着他们讨论的样子把原型搭出来。她的一句话总结是："会议本身就是 prompt。"https://x.com/zarazhangrui/status/2076300222884626754

The Takeaway: Nuclear's real bottleneck was never physics or reactor design — it's the will to actually build, iterate on hardware, and treat reactors as manufactured products rather than one-off megaprojects.

Isaiah Taylor is the founder and CEO of Valar Atomics, and he's an outsider to the credentialed nuclear establishment — which is exactly the point. His company just became the only private startup founded since the discovery of fission to make nuclear power, going from a Delaware filing to its first split atom in two years and four months. The metric he obsesses over is "tick rate," borrowed from video games: the time between one reactor turning on and the next. It's already down from 28 months to seven, and he intends to drive it into the minutes.

His central heresy is that nuclear is a hardware-execution problem, not a design problem: "Nuclear has never had its Ford moment or its Tesla moment." Most of the industry, he argues, is really a "modeling and simulation industry" producing beautiful paper reactors — so Valar refused to call itself a nuclear startup until it had physically split an atom, because "companies are what they do."

Two ideas stand out. First, safety by consequence, not odds: rather than engineering the probability of failure toward zero, Valar builds a TRISO-fueled, helium-cooled reactor that stays safe even if literally every system fails at once — they scram it, cut all power, and let passive physics carry away the decay heat. Second, ruthless verticalization to kill fake costs. Quoted $5 million and two and a half years for a reactor protection system, five engineers built their own in six weeks for $400,000. "Everywhere you look, it's a fake industry that has not been building anything for forty years." The prize: energy 10x cheaper, and a market that's effectively infinite because cheaper energy induces its own demand. https://www.youtube.com/watch?v=5Xvbq_zvOQ4

要点：核能真正的瓶颈从来不是物理或反应堆设计，而是"真的去建造"的意志——去迭代硬件，把反应堆当成可批量制造的产品，而不是一次性的超级工程。

Isaiah Taylor 是 Valar Atomics 的创始人兼 CEO，他并非科班出身的核能圈内人——而这恰恰是关键。他的公司刚刚成为自核裂变被发现以来、唯一一家真正发出核电的私营初创公司：从在特拉华州注册到第一次裂变，只花了两年零四个月。他最执着的一个指标叫"tick rate"，借自电子游戏，指的是一台反应堆点火到下一台点火之间的时间。这个数字已经从 28 个月降到 7 个月，而他打算把它压进"分钟级"。

他最核心的"异端"观点是：核能是一个硬件执行问题，而不是设计问题——"核能从未迎来它的 Ford 时刻或 Tesla 时刻。"他认为，行业里大多数公司其实是"建模与仿真公司"，产出的是漂亮的"纸上反应堆"；所以在真正裂变出第一个原子之前，Valar 拒绝自称核能初创公司，因为"公司就等于它做过的事"。

有两点尤其突出。第一，靠"降低后果"而非"降低概率"来做安全：与其把故障概率工程到无限接近零，Valar 造的是一座 TRISO 燃料、氦冷的反应堆，哪怕所有系统同时失效也依然安全——他们急停反应堆、切断全部电源，让被动物理把衰变热自然带走。第二，为了消灭虚高的成本而不惜彻底垂直整合。一套反应堆保护系统被报价 500 万美元、耗时两年半，五名工程师却在六周内、用 40 万美元造出了自己的版本。"你放眼望去，这是一个四十年没建造过任何东西的虚假行业。"而回报是：让能源便宜 10 倍，以及一个近乎无限的市场——因为更便宜的能源会自己创造需求。https://www.youtube.com/watch?v=5Xvbq_zvOQ4