今日AI领域呈现训练范式转变与安全工程并重态势：参数数量不再是唯一标尺，后训练缩放定律受关注，同时企业落地与隐私保护加速。
模型训练范式迎来关键转折。Latent Space刊文报道Z.ai CEO Jie Tang关于GLM 5.3及后训练缩放定律的观点，指出参数数量需结合数据规模、计算分配和运行条件综合考量，单纯参数计数已不足以衡量模型能力。同时开放权重模型进展迅速，如Qwen 3.8 Max和Kimi K3等，逼近Fable类模型水平，显示行业对模型规模的认知正在深化。
AI安全与隐私议题持续升温。OpenAI发布AI Futures博客，探讨AI如何重塑权力、治理、经济和个人自由，同时重申零数据保留政策并预览隐私安全处理机制，体现对数据隐私的重视。这些举措与日益紧迫的AI治理讨论形成呼应。
AI在软件开发中的实际应用取得新进展。Simon Willison记录使用smolmachines作为沙盒执行不可信代码的探索，过程中遇到环境限制但通过GitHub Actions解决，展现AI代理的主动性。他同时引用Jeremy Morrell观点，认为LLMs降低扩展作者成本，推动可扩展软件发展。此外，Stampli案例显示使用ChatGPT Work和Codex将发布准备时间压缩68%，验证了AI提升开发效率的实际价值。
基础设施与生产力讨论并行。vLLM发布v0.28.0rc1，尽管细节未披露，但表明推理基础设施持续演进。Simon Willison还讨论代码行数作为生产力指标的新意义，指出编码代理可大幅提升调试后代码产出量，挑战传统度量观念，为AI驱动的开发效率提供了新视角。
本汇总由 AI（deepseek-v4-pro）自动生成，朗读由阿里云 CosyVoice 合成；可能存在疏漏，请以原文为准。
vLLM 发布 v0.28.0rc1 候选版本，但页面加载错误，未能显示发布说明或更新内容。
There was an error while loading. Please reload this page.
There was an error while loading. Please reload this page.
There was an error while loading. Please reload this page.
There was an error while loading. Please reload this page.
There was an error while loading. Please reload this page.
Z.ai CEO 唐杰表示仅看参数数量不够，还需考虑数据、算力和运行条件；文中谈及 GLM 5.3 及新的后训练缩放定律，并提到年底或出现开源 Fable 级模型。
We’ve covered GLM 5.2 very excitedly before, and Prof Jie Tang’s belief that there will be an open weights Fable-class model by end of the year (spot check - with 134 days left, there are now two 2-3T models (Qwen 3.8 Max and Kimi K3) with estimates that Fable is 3-7T, and only 2 points higher on the AA index.)
Prof Jie Tang is back on X to tell us that our shorthand for model sizes is no longer enough: “Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions.”
We have covered Chinchilla (and post-Chinchilla) scaling laws in past LS years, but, so we will skip the history lesson, but it is good to level-set on why Chinchilla’s assumptions were wrong in the Inference Inflection world (no fixed number, between 200-900 toks/param, citing Roberts et al on task dependence).…
在截止日期固定、设计资源紧张的情况下，Stampli 借助 Codex 和 ChatGPT Work，将原本数周的发布制作压缩至数天，实现发布耗时减少 68%。
Simon Willison 评估 smolmachines/smolvm 作为安全沙箱运行不可信 Python 和 JavaScript 代码，限制资源、无网络，仅访问指定文件；因 web 环境限制，最终在 GitHub Actions 中安装 smolvm 进行测试。
I tasked Claude Fable 5 running in Claude Code for web with the following research task:
Put https://smolmachines.com through its paces as a fast secure sandbox. Explore what it would take to use this to run untrusted Python and JavaScript code in a way that is limited in what RAM and CPU time it can take up (protection against "while true") with no network access and filesystem access only to designated files
Goal is to be able to use this to execute user-provided tasks for things like data transformations
It quickly ran into a problem: the Claude Code for web environment can't run smol machines. Quoting the notes it wrote:
And Plan B is what it did, installing smolvm and running these tests directly in a GitHub Actions runner against that branch.
That was a creative solution to the environmental limits posed by Claude Code for web. Another example of Fable being relentlessly proac…
Simon Willison 引用 Jeremy Morrell 观点：LLM 大幅降低扩展开发成本，现代沙箱降低部署成本并提供安全边界，为网络可扩展软件创造新机遇，让用户安全扩展应用。
My hypothesis is that there is a new opportunity for Extensible Software on the web. LLMs radically lower the cost of authoring extensions, and modern sandbox primitives lower the deployment cost and provide good security boundaries. We can build our app as a solid, accountable core, and allow users to safely extend it in many directions by having LLMs fill in the missing pieces. We can give our users super powers.
— Jeremy Morrell, Extensible Software in the age of LLMs
This is a quotation collected by Simon Willison, posted on 19th August 2026.
Simon Willison 在播客中讨论 AI 如何改变软件开发，认为在编码代理时代，用代码行数衡量生产力有一定意义，因为过去工程师每天仅能产出数百行生产代码，而现在有了硬限制的突破。
Last week I recorded an episode of the Talking Postgres podcast with Claire Giordano on the subject of “How AI is changing software development”. We had a really great conversation. Here are a couple of my highlights from a lightly edited transcript (prompt to Claude: “very minor edits to remove disfluencies”).
This is the latest version of an argument I’ve been trying to build about why sometimes it does make sense to talk about lines of code as an indicator of productivity with coding agents, at 35:01:
A lot of people will tell you it makes no sense to measure productivity in lines of code. I’d actually disagree, because there’s a hard limit. In the before-times, a software engineer could produce a few hundred lines of production-ready code per day — and 200 lines of working, debugged, production-level code is an incredibly good day. Most days you’d produce 50 or 60.
If agents let y…
OpenAI 重申为符合条件的 API 客户提供零数据保留，并预览私有安全处理功能，在不损害数据隐私的前提下实现高级 AI 安全。