📈 社区总体趋势 · LLM 综述 · deepseek-v4-flash
Sam Altman 公开附议 Dario Amodei 的「为前沿模型降速」，并承诺 OpenAI 引入具备员工级权限的独立评估者——Tomek Korbak 称这类审计已从空想变为行业标准。安全阵营随即出现分歧：Eric Ho 把可解释性列为对齐的首要瓶颈并预告开源对齐栈，Micah Carroll 则要求第三方员工级审计不打折扣。同时，Agent 工程的重心继续从模型转向外围：AI Engineer 的 Harness Engineering 专场直接下判断——生产环境里 Agent 崩溃，出问题的很少是模型；日本法律团队自建 MCP 把 Claude Code/Codex 改造成法务专用，国内某平台则要求 System Prompt 必须写「You are Claude Code」才放行。另一条线上，评测与变现都被重估：ApprenticeBench 发现后端 API 要包成 MCP 才勉强追平原生 GUI；同预算下 token 并不等价，加一个「顾问」角色能把金融分析准确率从 76% 提到 89%；1Password「AI 只能修 26% 漏洞」被 Trail of Bits 质疑为评分口径问题。CJ Zafir 则押注 0.5B–12B 小模型微调授权取代软件变现，宇树 G1+ 把肩腰扭矩翻倍、负载提到 3kg。
今天 13
Wao, this post blew up, and I triggered some Saastophers.
But let me tell you more about SLM fine-tuning.
In April-May this year i fine-tuned a 6.5B model: Mac-1 (it controls 487 Mac native apps)
My goal was to build a better Siri (and i achieved that).
Here's how i
解读自述 6.5B 微调模型 Mac-1 控制 487 个 Mac 应用，胜过 Siri
Images 2.5 makes perspective part of the design!
I tried four different ways to build depth with people and architecture.
1. FRAME SHIFT, nested frames
2. VANISH POINT, converging lines
3. LEVEL SHIFT, stacked levels
4. STRATA, layered slabs
Great for editorial posters,
解读Images 2.5 把透视变成设计工具，四种造层次手法实测
「AIは脆弱性を26%しか直せない」
この数字だけ一人歩きすると危ない。
1Passwordは6 CVE・6,080個のAI生成パッチを評価し、
完全修復できたのは26%と報告。
でもTrail of Bitsが突いているのは、
AIの性能そのものより「採点方法」。
セキュリティパッチは、
・PoCが止まった
・テストが通った
解读AI 修复漏洞仅 26% 引争议，Trail of Bits 指评分方法有问题
となりのトトノLocal LLM | Tonoken3 @Tono_Ken3 @Tono_Ken3 工具生态3小时前
我想用本地 LLM 机器人来管理自己的硬件和数据库，并保持其干净整洁。
我想赋予它们像人类一样具有自清洁功能和自主能力提升机制。
如果可能的话，我希望仅依靠太阳能系统的电力来100%满足运营能源需求，因此无法构建大规模系统。
Lna-Lab 的电量预算是 20 Kwh/Day。
解读构想本地 LLM 自管硬件与数据库，日耗电预算 20kWh
Unitree G1がG1プラスにアップデート！
見た目はG1だけど、中身はけっこう大きく変わっている。
首が2DoF追加されて、カメラだけ対象に向けられるようになった。
肩・腰のピークトルクは約2倍、腕も強化。
腕の可搬重量も標準で2kg→3kg
センサーも
Stereoカメラ、広角カメラ、3D
解读宇树 G1+ 升级：颈部加 2 自由度，肩腰扭矩翻倍，负载 3kg
となりのトトノLocal LLM | Tonoken3 @Tono_Ken3 @Tono_Ken3 工具生态4小时前
地元のLLM的机器人被指示使用订阅版的GLM-5.3-Flash作为Ahigaru-Search的足轻
用于拉取信息并进行检疫的话，使用订阅版的风险较小
解读本地 LLM 借订阅版 GLM-5.3-Flash 拉信息，称风险更低
LLM 的速度，有两种体验：等多久开始回答，回答时是否流畅。 两种体验，与 Prefill 和 Decode 密切相关。
从发出请求到收到第一个 token，这段等待叫 TTFT，首 token 反应时间。
模型在其中进行 prefill，处理输入内容
开始回答后，模型通过 decode 逐步生成后续 token，相邻两个 token 到达的间隔叫
解读科普 LLM 两种延迟：TTFT 看 prefill，流畅度看 decode
Amazing news. Congratulations to Dave, to
@OpenAI
, and
@RonnieChatterji
. We need more talented, ambitious researchers studying the economics of AGI, and I can't wait to see what Dave's group produces.
解读OpenAI 招揽研究者做 AGI 经济学，学术界跟进
we try to publish and open source as much of our research and alignment tech as possible. it's really important to have an open alignment stack
we'll be publishing a ton in the next month
解读承诺尽量开源对齐技术，下月集中发布成果
it's more urgent than ever to solve alignment, and more people should think of it as a solvable technical problem
i'm confident we can do it, but we first need to solve interp which i think is the primary bottleneck
解读把可解释性定为对齐首要瓶颈，主张这是可解的技术问题
Tokens aren't fungible. Same 600K token budget, same financial analysis bench: plain execution hit 76% accuracy. Execution plus an advisor hit 89%.
Watch
@katelyn_lesse
and
@angjiang
walk through advising, grading, and dreaming as jobs for tokens.
解读同预算 token 不等价，加「顾问」角色准确率 76%→89%
“While I agree with those who downplay the [Hugging Face] attacks by claiming that there are basic measures that could have prevented them, I think the important lesson is that even knowing all the mistakes that were made, one would not have predicted that the agents would behave
解读复盘 HF 安全事件：即便知道全部失误也难预测 agent 行为
Live now: our Harness Engineering Track from AI Engineer World's Fair 2026.
Ports, proofs, kill switches, and what is left of an agent when you remove the model.
Thesis: when an agent breaks in production, the model is rarely the part that failed.
https://
youtube.com/watch?v=PXj0p_
mW9nI&list=PLcfpQ4tk2k0WzqWDdWkN2DnZOhtYI9jyI
…
-
解读论断：Agent 生产故障很少出在模型，多在 harness 层
昨天 6
Monetization of software is dead now.
Anyone can clone your app in few hours, copy your ads, ditto your positioning.
I see a new opportunity now.
Monetization of SLMs (small language models)
You can fine-tune 0.5B to 12B models for specific use cases.
Then sell licenses to
解读断言软件变现已死，押注 0.5B–12B 小模型微调授权
Sam Altman @sama @sama 安全对齐1天前
There are two ways AI progress could go very badly and that we must avoid.
First, we could lose control of the future to AI. This is unacceptable; we are unapologetically on Team Humanity, and AI must always serve people. To ensure that, we need ways to ensure that alignment and
解读Altman 再划红线：失控于 AI 不可接受，对齐是前提
鹈鹕骑车又刷屏？
试试鹈鹕骑车背后的作者另外一个benchmark！
同一句话丢给 6 个大模型：
"Write an HTML and JavaScript page implementing space invaders"
规则很简单：
每家只写一次,第一次写成什么样就什么样：
不允许重试,不改一行代码。
结果 6 个全能玩，但
有的 40 秒交卷、极简草草了事；
解读六个大模型一次性写太空入侵者，都能玩但速度风格分化
But China... may also be down for coordination!
I appreciate that
@sama
isn't doubling down on China hawk positions – races to the bottom with China are not inevitable.
Catastrophic risks would be a lose-lose for everyone.
解读指出中国也可能愿协调，赞 Altman 不搞对华鹰派竞赛
Important result from our Google AI x Econ team. Field evidence for pre existing expertise favoring the ability to learn from AI assisted work.
解读实测：原有专业能力越强，越能靠 AI 辅助工作获益
MidoriKochiya0/0/1 @enp39s0 @enp39s0 Agent工程1天前
我司的大模型平台中转和部署了一系列的模型，一定的额度内免费使用。然而除内部工具外，只支持Claude Code一个第三方Harness，Codex、Pi、DSH等用正确的API也无法使用。
怎么做到的呢？System Prompt的第一句必须是
You are Claude Code, Anthropic's official CLI for Claude.
否则就会400，绝了。
解读平台靠 System Prompt 硬编码识别 Claude Code，暴露 harness 锁定
本周早些 6
Weak sauce. Employee-like third-party auditing or bust
解读批评 Altman 表态力度不足，坚持第三方员工级审计
朝戸統覚(Asato Noriaki)@Legal Agent @NoriakiAsato @NoriakiAsato Agent工程2天前
Legal Agent では、MCPを自作し、ClaudeCodeやCodexを法务业务に特化した形で使用しています（ClaudeCodeやCodeは、そのままだと法务业务との関係では使いにくい）。
そのMCPの開発を一人で担当し、大活躍されている、LAのエンジニアの博客です！
ぜひご覧ください！
フィールドレポート
解读日本法律团队自研 MCP，把 Claude Code/Codex 改造为法务专用
If you showed me Claude Code today in 2018, I would have thought it was AGI.
We have already absorbed a dramatic amount of change in the profession of software engineering and society as a whole. But still, we’re starting to see cracks.
Things are accelerating faster than I can
解读称 2018 年看到 Claude Code 会当作 AGI，但工程界已现裂痕
It’s kind of remarkable how quickly “independent auditors with employee access” has gone from a pipe dream to an industry standard. More next?
解读感叹员工级独立审计一年内从空想变成行业标准
Sam Altman @sama @sama 安全对齐2天前
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks.
Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.
解读OpenAI 将引入员工级权限的独立评估者，前沿降速落到机制
nice writeup,
@kohjingyu
.
Fun fact: when building ApprenticeBench, we wanted to do a rigorous between GUI and API.
The full set of backend APIs is available. We just needed to wrap them into MCPs to reach feature parity with the native GUI. But that turned out to be quite
解读ApprenticeBench：API 包成 MCP 才勉强追平原生 GUI