Eric Xu (e/Mettā)

@nicdunz Yep, it's the red queen race. The "patent" period of frontier models is now 3-6 months or less. They *have* to reduce the price in this overly competitive market.

GPT-5.4 full at xhigh scored 51, exactly where Luna max sits today. GPT-5.4 costs $2.50/$15; Luna now costs $0.20/$1.20. In other words, roughly four months later, OpenAI is selling March’s full flagship intelligence at about one-thirteenth the token price.

I wrote about the "patent" period of frontier models a few days ago. New data point is clear: for those "little brother" models, frontier labs are increasingly losing the pricing power, thanks to fierce competition. This time around, GPT 5.6 Terra ($2/$12) is right below the Kimi K3 price ($3/$15). The pricing dynamics will again be similar to big Pharma -- the premium drugs are premium, but the last gen and generics are cheap and competitive. I won't be surprised that a natural next step for frontier labs is to open-weight those models one generation behind, because it would be easier to lock the ecosystem onto the harness + tax the inference providers than keeping up all the inference stack internally. ---- GPT 5.6 的中档和低档模型都降价了，而且和我前文所写的“专利”药品一样：越是低档的价格降低越多。这次上线才三周时间就下调价格 无论是被迫还是主动选择：前沿实验室一定会持续使用这个策略 -- 即除最强模型之外（有定价权），在中档和低档模型上，不寻求很高的 margin. 目前开放权重模型也就是比最好的模型差半档的样子，因此为保障用户锁定在生态系统，就需要中档和低档模型的定价低于对应的开放权重模型。 Terra 的 $2/$12 就正好切在 Kimi K3 的 $3/$15 下方一点。 这样外推下去的话，一年内 OpenAI /Anthropic 把上一代的中档，低档模型开放权重我都不感到惊奇：为了继续让用户为最高档的模型付费，除了模型要保持领先外，生态也很重要。为了培养生态，把已经不能够赚取高额利润的模型开放出来给社区（继续有商业协议保护）是一个理性的选择。

Hey Haiping, a few months ago I took a stab at this and built a “vibefs”. I shifted my focus to other projects but happy to see others taking a fresh look. What I learned: git workflow is very sticky. When it is coupled with bash the design space is narrower. There might be a way to redo the whole tool chain, but to make today’s agent work you might end up with a fairly complicated overlay fs. Google’s internal srcfs etc were very successful and the cost is to not use git workflow at all. Mac new FS support is getting to be good but the API so far is swift only. Managing node_modules/ etc via an overlay is fairly costly. It’s doable, but the hypotheses that you can do it in SQLite might be false. Happy to share more. I know a few folks had kernel level proposals.

I’ve been thinking a lot about git worktree lately. It’s one of Git’s most powerful features. Being able to work on multiple branches without cloning an entire repository is fantastic. But I don’t think it’s the endgame. For modern projects—Rust, Node.js, Next.js, Go, Java, Python, and beyond—the biggest cost isn’t cloning the source. It’s rebuilding the world. Every worktree eventually accumulates its own build artifacts: Rust: target/ Next.js: .next/ Node.js: node_modules/, package caches Go: build cache and module cache Java/Kotlin: build/, .gradle/ Python: virtual environments, __pycache__ C/C++: build/, object files Once you start working across multiple branches—or let multiple AI agents collaborate—you end up with massive duplication. The source code is mostly the same. The dependencies are mostly the same. Most build outputs are identical. Yet every workspace stores its own copies. I think we need a new abstraction: a snapshot-native development workspace. Imagine creating a new workspace in milliseconds: - Instant filesystem snapshots (APFS/Btrfs/XFS reflinks or CoW) - Immutable source shared across workspaces - Shared dependency stores - Shared compilation/build caches - Branch-local writable state - Automatic garbage collection - First-class support for parallel AI agents Instead of treating every workspace as a complete copy of a project, we treat it as a lightweight overlay on top of shared immutable data. This isn’t just about saving disk space. It’s about making developer workflows—and AI agent workflows—effectively free. Spawn 50 workspaces. Run 50 experiments. Benchmark multiple implementations. Open 20 pull requests. Launch 100 coding agents. Each gets an isolated environment, while the underlying source, dependencies, and caches are shared wherever possible. git worktree solved the “multiple branches” problem. The next step is solving the “multiple execution environments” problem. As AI agents become a core part of software development, I think snapshot-native workspaces will become as fundamental as Git itself. I’m seriously considering building this.

@jiayq Congrats! I also like the //. logo -- human writes the code comment and AI completes the rest.

We are starting Intent Lab, building an autonomous team we call "fleet" that turns intent into production software. Today we are sharing some early results: the fastest GLM5.2 inference engine, one shot database creation, and a fully verified agent filesystem. intentlab.ai/blog/turn-your…

@HiTw93 恭喜！ Mole 对于非常不健康的 Mac 清理生态真的是一阵凉风！现在 Noah 有些 Mac 用户告诉我说他们用不到高级修理功能的时候，我都直接说去用 Mole 😂

🎉恭喜 Mole Cli 达到了 60K 的 Star，从去年国庆当时我在三亚泳池写下几百行代码开源至今，居然发布了累计 55 个版本，累计 112 位真实的来自全球各地的开发者贡献了非常多的代码，累计解决了 812 个功能建议和Bug，合并了 483 条有效的来自全球各种 Mac 电脑的环境下的清理优化功能迭代，非常感谢大家各种帮助。 开源应该是友好的、快乐的、可以相互学习的、1+1>>2的那种纯粹，也是认识很多志同道合朋友的一个很好的渠道，也让我更加相信长期主义，以及在开发过程中对于各种功能的把控，深刻体会到至繁归于至简，如无必要、勿增实体的产品关键的实践，接下来 Mole 会继续强大，期待有一天全球每 50 台 Mac 就有一台安装了 Mole Mac 或者 Cli ，这一天我相信会来的，那是会多么有趣的一天。 github.com/tw93/mole

You really don't need "software factory" or "loop engineering" to work effectively, or in this case burn $10K per month on a subscription of $200/month. A few very simple tips: * make every ticket ready to start -- pay attention to the grooming and design phase of any task -- ruthlessly delete tickets; * set up a separate repo for the OS that's different than the code repo -- this repo stores working agreements, workflows, learnings, docs. Keep the code repo(s) clean and dry -- pure code and nothing else. * always start working from the OS repo: OS + linear tickets == durable work context. Code is just a short-lived artifact * if you actually need to use goals, write them down in the OS first and then issue them to agents. * don't ever read code, but design a few "gates" your agent have to go through. For me: pixel gate for me to review UX, code review gate(automatic), release gate (automatic). * Design your "squad" -- a group of agents together doing one task. One squad for one complete epic. Tickets should be filed at squad level, such as "finish onboarding v2", not at task level, like "add a button to onboarding". -- 过去的 30 天其实没有用任何所谓的“软件工厂”/loop/graph，就是遵照一些基本的软件工程的原则，默默地就烧掉了 13B 的 tokens，折合 API 价格 $10110. 就是几个很普通的原则如上

几周前的推文 x.com/xleaps/status/…

Just like I predicted: Kimi K3's license requires inference providers to enter a separate commercial agreement (ARR > $20M). Kimi might be the first company doing a "fabless" business model. -- 验证了我几周前的预言: Kimi 的开放权重模型 K3 ，有单独的商业授权协议，在这样的协议下，模型训练和推理可以在商业上分开，就和半导体设计与制造一样。 围绕着开放权重模型会有一整套的生态和公司，比如专门从事为企业应用微调的，专门做 prefill-only 适配的等等。这种生态系统层面的力量，是封闭厂商完全没法匹配的。

Thanks for sharing. The intuition that MLP stores facts is known since the ROME paper and externalizing the layer via vector store, but having a closed form recipe is very nice. I am curious to know in an actual transformer, how does it work for facts to go to each layers, and why.

MLPs store facts in language models. Can we write them into Transformers without training? New work w/ amazing team @garctrob @ronnygjunkins @EyubogluSabri, Atri Rudra & @HazyResearch gives a ✨closed-form✨ recipe for fact-storing, Transformer-ready MLPs. Accepted at COLM 2026!

@larryhaonlp 愿闻其详，比如 hawk-chicken 模式的均衡是什么样子的。 （最根本的问题是这个棋局上有许多玩家，谁也没有假设游戏里只有“双方”的自由）

非常精彩的文章！但两个纳什均衡的论证细节似乎不太够？@xleaps 根据我浅显的博弈论知识似乎存在一个反例，在同样“溢价窗口很短”的叙事下，合理收益矩阵既可产生两个不对称纯均衡加一个混合均衡，也可产生唯一的“双方都等待”均衡；所以“自动滑向同时出发”不是纳什均衡定义推出的结论吧？期待补充更多细节

大模型是一个专利期十个月的制药生意 x.com/i/article/2080…

@Mononofu @JensenHuang Anthropic is quickly burning all its credits with the community, and unfortunately there is no weekly credit reset in this case.

I’m so excited that @JensenHuang is a believer in open source now, looking forward to the CUDA and GPU driver open source release!

@dotey advisor 和 orchestrator 两个范式最终都可以归结到一个问题：系统最大的不确定性，也就是最需要高级智能的场景在哪里 如果把这两个范式贸然糅合到一起，有可能是问题的边界和上下文策略还没想清楚

我有个疑惑，如果让弱一些的模型设计，聪明的模型当顾问，那么有几种可能： 1. 它设计就是很糟糕的，问的问题也问不到点子上，那再聪明的顾问也给不了好的建议 2. 它很自负，不觉得自己的有问题，明明该去问聪明模型而不去问 3. 它很不自信，一点小问题就去问聪明模型 这些可能都让所谓的 /advisor 模式没有什么用。 为什么不让聪明的模型去设计，让弱一点的模型去执行，最后聪明的模型去验收？这样应该才是性价比最高的组合方式。

@hello_mirr89197 同意 参见我前面的一个简评：开放权重就是为了解决模型训练出来了没有算力部署的瓶颈。比如 Kimi 3 ，已经是在商业授权这条路上了。这时候新云厂商，包括云服务商，就会问：我手里的 GPU 到底是和谁签协议去服务推理呢？

@xleaps 就目前国模推理算力短缺的现状，就算国模能力足够接近，而且 API 标价只有美国模型的十分之一。也仍然推不出：“美国模型的定价权窗口会从十个月缩短到三个月。”因为缩短到三个月我认为默认了一个前提就是推理算力足够。

@tony9527167 好几年前就开始买了，谢谢推友 @hzhu_ NET 股票我觉得是有多少可以买多少 （是我个人意见，不构成投资建议）

@xleaps 你这么说我就要吹 Cloudflare 了 ，它们都用，流量越大付费越多，好消息 Cloudflare 是上市公司股票随便买😃

大家或许不知道的是，那些 OpenAI 和 Anthropic 的账单，包括 API 和月订阅 都是 Stripe 处理的 不是说要卖铲子么 Stripe 就是卖铲子的其中一家 从产业链上每 Watt AI 拿走大约 $0.05 但支付本身消耗瓦特极低，所以是一个近似于 Meta 的极好的生意 可惜是未上市标的，否则上涨应该是超过内存股票的

Two papers at #SIGIR2026: ROO and SilverTorch. Both were bets from 2021: fix the logging schema and better models will follow; a multi-funnel retrieval system reduces to one trainable layer. The full story, five years later, and how to build a scaling law: blog.jiaqizhai.com/2026/two-bets-…

这个模式我在 LinkedIn 上前几天写过 其实是一个真正追求产品质量的团队最终必然要实现的一种模式 （我原创的名词：叫做 shadow steering ） 它是这样的： 平时一个普通的模型工作。另一个更聪明的模型在后台并不需要做什么 甚至无需输出 只要输入 （输入成本极其便宜）来维护一个 prefill kv. 到了某个时刻　harness 侦测到小模型质量下降干不下去的时候，大模型介入 输出一百个 token 的指导 小模型被点悟 一点即通 继续干活 这个 “影子点拨”可以藏在小模型的思考轨迹里 反正用户看不见 这种直接在轨迹上的点拨比 off policy 不知道高到哪去了，因为可以在长程任务中期给一个 on-policy 的指点 后面拿来蒸馏非常舒服 这个模式可以用来节省大量中等模型成本，也可以用来蒸馏，用来测试，用来收集长程任务质量等等 回到这则新闻本身。我看到的是它只能说明一件事情，DeepSeek 下一代模型已经训练完毕，在影子测试而已 智能程度让用户感知到了 Fable 5 而已。

And to your point about "tasks in progress" -- yes the Wall I have is exactly that. I call it "Floor view" -- the hustling and busting of the agents. It's actually pretty easy to one shot build that view -- just ask each Agent to post to the Wall at critical moments. And the backend of that is dirty cheap.

I made a few "gates" -- basically similar idea to yours but in my own workflow -- pixel gate (any work that changes desktop UI need to render a pixel for me to review and approve); release gate (only one agent called Shipper can cut a new app release and I approve that), among others. When Squad meets the gate, they stop and ask me for approval. There is also some Key Decision gate for Growth but TBH I am trying to shift them to an aligned Chief of Staff once I have all business North Star (and PostHog) exposed to it and get it aligned well.