JonahKovics

Internet noob

[JonahKovics @Kovics8648]

Successful Ukrainian counteroffensive especially to the north bolsters Western confidence of the country's victory over Russian invasion and most likely boosts louder voices of strengthening Taiwan's defence against potential Chinese military offense. 3/3

[宝玉 @dotey]

REA：让 AI 智能体帮你逆向没有源码的软件

开源项目 REA（Reverse Engineer Anything）把逆向分析工具接到了 Claude Code、Codex、Cursor 这类 AI 编程智能体上。看到别人 App 里有个好功能，可以让智能体去拆这个程序，讲清楚功能是怎么实现的，附上证据，再在你自己的项目里写一个类似的。逆向分析指的是手里没有源代码，从编译好的程序反推它是怎么写的。

REA 本身是一个 MCP 服务。MCP 是 AI 智能体调用外部工具的通用接口，接上以后，智能体在对话里就能直接调用 REA 的分析功能。安装只需一行命令 npx rea-agents setup，它会注册到 Claude Code、Codex、Cursor、Gemini CLI、Grok Build 等智能体里，改配置前会先备份并让你确认。不用智能体的话，也可以在终端里直接跑命令。

能分析的东西很多。原生程序（直接编译成机器码的软件）要借助 Hopper、Ghidra 或 IDA 这几款反汇编工具，把机器码还原成汇编和近似 C 语言的伪代码。Electron 应用（用网页技术做的桌面软件，比如 Notion）可以直接拆开安装包里的 ASAR 归档文件，理清模块和进程间通信，不需要额外工具。此外还支持 .NET 程序、安卓 APK、固件、网站、抓包记录、以太坊合约字节码，以及在 Linux 和 macOS 上记录程序运行时的行为。

分析在本机进行，REA 不上传目标程序，但分析结果会交给智能体背后的模型服务商，数据怎么处理看各家政策。

项目给了三个案例。一是经典打砖块游戏 DX-Ball：从一次播放音效的调用，追到根据球的位置计算左右声道的函数，把不完整的伪代码还原成 C 代码，3205 组测试结果和原版一致，编译出的 63 个字节也和原版完全相同。二是 Notion 桌面版：追踪复制粘贴从界面一路经过预加载脚本、进程间通信到主进程的完整链路。三是日本老电脑 PC-98 上的游戏《东方幻想乡》（TH04）：从 16 位指令里还原出弹幕子弹环的角度算法。

对安全研究员、���老游戏复刻和软件保存的人来说，以前要自己对着反汇编工具一行行读、一层层追调用，现在可以让智能体先追，人来核对它给出的证据。普通开发者想研究竞品某个功能的做法，也多了一条路。

README 显示项目在 GitHub 上已有 5 万星。npm 包今年 7 月首次发布，最新的 6.3.0 版今天刚发。

使用前注意两点。项目声明只支持合法的逆向研究，授权和合规由使用者自己负责，逆向别人的软件可能违反用户协议或版权法规。项目还说明从未发行或背书任何加密货币，借用 REA 名字的代币都与它无关。

https://t.co/I6ZzxyeHhC

[R @wheresryan22]

Introducing Anatomy

Describe a machine in one line. Claude brings it to life.

- Interactive isometric wireframes you can rotate, inspect and explore
- WebGL shaders for fire, water, plasma and more
- Real 3D motion, with individual parts that rotate and move

Install it as an agent skill. Fully open source.

[阿磊 @yangleiocx3]

少儿不宜，未成年人请绕道 🔞
Qwen-Image-2.1 的社区衍生版，这次连名字都不藏了：
Noct-Q-Uncensored-Qwen-Image-2.1。
主打成人内容与写实摄影。按作者说明，相关能力已经整合进模型，不用额外挂 LoRA。
本地玩家可以留意：INT8 权重约 7.3GB，作者标注适用 8–12GB 显卡，还附了 ComfyUI 工作流。
不过，7.3GB 只是模型权重，文本编码器和 VAE 还得另外配齐。
模型已放出，仅限非商业用途。传送门放下面👇🏻

[Tailscale @Tailscale]

Codex 🤝 Tailscale

Codex Cloud can now securely connect to resources on your tailnet with Tailscale.

Useful tools should be easy to build with, combine, and turn into something even more useful.

See how it works →
https://t.co/0eFd37GmmO

[Jian Wang @jianw851]

All you need for System Design Interviews 📌

https://t.co/b4elsttIea

[𝗿𝗮𝗺𝗮𝗸𝗿𝘂𝘀𝗵𝗻𝗮— 𝗲/𝗮𝗰𝗰 @techwith_ram]

Transformer Architecture

Explained in 3 ways here:

- Encoder only
- Deconder only
- Encoder - Decoder both

[Jacky Kwok @jackyk02]

Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions.

CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks.

With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks.

We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed.

Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size.

📄 Blog: https://t.co/zwi9JOHKGx
💻 Code: https://t.co/rsHRYCGR8I
🗣️ Discord: https://t.co/Uqtdefvo3J
🤗 Data & Models: https://t.co/wdSWGGO3hu

More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵

[VerySmallWoods @verysmallwoods]

上一回我做了一期视频，介绍 Pi Durable 是什么。
https://t.co/XzvCy7MY3G

本期视频我们一起来体会基于 Pi Durable 的智能体开发体验。咱们从一个空目录开始，快速做出一个基本款 Agent，再分阶段完善功能：
https://t.co/XKoDzBAkxC

- 创建 npm 项目、安装依赖
- 接入 DeepSeek 模型，以 SQLite 为存储，运行第一轮对话并检查数据库数据
- 安装工具 read, write, edit, bash
- 加入 Skills 的支持
- 添加重启后任务继续的支持

基于 Pi Durable 的智能体开发体验还是蛮愉悦的😀

[Tibo @thsottiaux]

Day 2.1/

We have made Auto-review free for all users signed in through a ChatGPT account. You can enable it in settings > permissions > auto-review. Auto-review improves upon the default sandbox setting that requires you to approve everything, which is prone to decision fatigue unless you spend a lot of time configuring specific rules.

It allows you to run long tasks while having a second agent review all actions taken by the primary agent. Its only goal is to prevent high-risk actions from being taken and to protect against unwanted actions that are not aligned with the original user intent. This Auto-review feature is now free and does not draw usage from your plan.

[TypeLLM @TypeLLM]

Jev answers every question instantly. That's great, until a question is hard.

TypeLLM is System 1 when it can be, System 2 when it has to be.

"thinking": "auto" picks per question, per call.

Read more: https://t.co/ONbHAs6yIu

[俊哥 Felix @LiFeng61532]

上海人工智能实验室有免费的DeepseekFlash
科研导向 比官网API智商还高
每天签到领五亿token
入口:https://t.co/q3lN0ncgVk

[云辰 Cyan @smk7z]

answer-me-with-html 🔥

star 猛涨，谢谢大家的认可！

基于 @karpathy 的思想，让 agent 能快速生成可视化结果！！

1. 超级快速： 比正常输出 HTML，token 少 7.4 倍！快 3.6 倍
2. ASD-STE100 原则： 让 Agent 用最少的字把事情讲明白！

https://t.co/RoIt1WkuQB

[Superman @thesupermannx]

Meta just change the game!

They open-sourced Context Language Models. Instead of using LangChain, RAG, or any orchestration harness, the model natively manages its own context like a file.

And it beats every SOTA context method with 59% less compute.

Instead of forcing a model to swallow endless walls of text like a passive file cabinet, they gave language models something revolutionary:

The ability to natively manage their own context.

To treat context like a dynamic file.

To learn their own internal policies entirely through weights, with zero human-designed frameworks or rigid retrieval harnesses.

The results shatter current benchmarks.

Using online reinforcement learning, they applied CLMs to the Qwen3.5-9B model.

Performance on BrowseComp-Plus skyrocketed by 47.6%.

And the wildest part? It used 12% fewer FLOPs to do it.

They didn't stop there. They co-designed Suffix Cache Reuse for server-side inference, slashing compute costs by 35% compared to standard SGLang at matched performance.

For years, labs have spent billions trying to engineer smarter ways to stuff more tokens into an LLM's prompt window.

They were solving the wrong problem.

[Akshay 🚀 @akshay_pachaar]

Jev for RAG, clearly explained!

Hybrid search gives you a shortlist. It does not decide which passages contain evidence, which are merely adjacent, or whether the evidence is strong enough to answer.

That missing judgment is where Jev fits.

Jev sits between retrieval and generation. It does not replace BM25, embeddings, or the LLM. It evaluates the candidates they produce before those candidates enter the context window.

→ Retrieve wide

Combine dense and keyword search, then merge the results with reciprocal rank fusion. This gives Jev a broad candidate pool, such as the top 20 passages shown in the visual.

Retrieval still sets the ceiling. If the right passage is missing from this shortlist, Jev cannot recover it.

→ Judge every candidate together

Send the query as Jev’s state. For each candidate, ask a typed yes-or-no question such as “Does C7 help answer this query?”

Jev evaluates every question in one packed request and returns a calibrated probability for each passage. This avoids making a separate model call for every query-passage pair.

→ Let code apply the threshold

Your application compares each probability with a threshold. Candidates above it continue to the LLM. Everything below it is removed before generation.

Jev makes the fuzzy judgment. Code remains responsible for the actual decision.

→ Gate the entire answer

The same request can check whether the retained passages make the query answerable. It can also flag signs of prompt injection inside a candidate.

If answerability falls below the threshold, the application can skip the LLM and return “not in the documents.” The injection score should remain a filtering signal, not a security boundary.

The result is a cleaner division of work.

Hybrid search retrieves broadly. Jev reranks, filters, and decides whether sufficient evidence exists. The LLM writes only from the passages that survive.

Jev’s value here is not simply moving passages up or down a list. It turns relevance into an explicit probability that your application can inspect, threshold, and act on.

To summarise:

- Retrieval finds the candidates.
- Jev decides what deserves context.
- The LLM writes the grounded answer.

----
I also built an open-source project showing how to use Jev as a judge for AI observability with Comet Opik.

It evaluates support traces for groundedness, request coverage, action honesty, and helpfulness, then records the results as an auditable experiment.

You can explore the project here: https://t.co/XVmpg6ldHT

My article on how Jev works is quoted below.

[elvis @omarsar0]

https://t.co/Pq7A9zA0Rf

[Google Gemma @googlegemma]

Deploying DiffusionGemma-Jev (djev) just got a lot easier. You can now spin up a Jev API-compatible endpoint on Google Cloud Run using a single command.

Performance is solid: ~35-60 ms for single step latency and batch@32 is ~100-123 requests/sec.

It's a straightforward way to experiment without needing your own GPU. Runs at roughly $3/hr and drops to $0 when idle.

Get the code and instructions here: https://t.co/E3agzrW2hZ

[VerySmallWoods @verysmallwoods]

Jev 到底准不准？我拿两个 Kaggle 数据集测了 662 次，结果有点出人意料。

我用 kaggle 上的数据集：
- SMS Spam Collection：5,574 条英文短信，抽取 200 条，当 smoke test
- Banking77：13,000 条真实的银行 App 提问、77 个意图选项，抽取其中 462 条

一共做了 662 次 Jev 调用。同样的数据和问题再交给 Claude Haiku 4.5 做对照。同时约定 Haiku 同样输出 confidence 值。

测出来的结果如下：
- 准确率：短信场景 Jev 96.5% / Haiku 93.5%；77 选 1 场景 Jev 81.6% / Haiku 80.1%，微调过的 BERT-Large 93.7%
- 概率校准：报 50% 到 85% 时基本都准；报 90% 以上偏乐观，报 100% 的时候，��际有每 40 条错 1 条的可能
- confidence 门控：只自动处理 confidence ≥ 0.95 的请求时，能覆盖 63% 的请求、准确率达到 95.5%
- 延迟（jev vs haiku） 432 毫秒 vs 1.4 秒；费用每千次 8 美分 vs 2.47 美元

总结
在我的评估场景中，jev 快且省，但决策质量并不显著高于 Haiku 4.5 以及微调过的 BERT 模型

潜在改进点：jev调用的 state 输入通过优化，可能有提高决策质量的可能。

我会继续把玩。

[AYi @AYi_AInotes]

这篇应该是这几天全网把 Jev 讲得最透彻、认知穿透力最强的一篇技术图解长文了，墙裂推荐！！
过去三年我们几乎把生成式大模型当成了手里的万能锤子, 哪怕面对一个最简单的单选判断, 也要硬生生调一次几毛钱的通用大模型。

技术大牛 @akshay_pachaar 刚发出的这篇长文直接扯掉了这层遮羞布: 一个连一句完整对话都说不出来、甚至不会写一行代码的模型, 为什么刚上线就把全网架构师看燃了? 因为不能生成长文本, 恰恰是它最可怕的工程杀手锏。

绝大多数企业软件从来不需要另一个陪人闲聊的聊天框。

一个复杂的智能体在后台跑起来, 往往需要做成千上万次细碎的边界判断: 这张工单紧急吗、该挑哪个模型处理、这条命令有没有危险。

传统大模型的致命死穴是自回归生成。

哪怕最终答案只有一个单词, ���也必须在云端一个字一个字往外吐 Token, 遇到 JSON 格式错误还要重试。

调专做强类型决策的 Jev 之后, 所有的判断逻辑全部变成了一组确定的输入状态与候选问题, 输入百万 Token 仅四分钱且输出完全免单, 官方测算下延迟直接压进几十毫秒, 成本最高被砍去了 400 倍。

更耐人寻味的是它对幻觉的重新定义。

很多人鼓吹它彻底消灭了幻觉, Akshay 极为冷峻地指出了真相: 它只是在格式上百分之百免疫幻觉, 绝不可能吐出未定义的选项或者破损的文本, 但它依然可能在合法选项里选出错误答案。

这正是这套架构最值钱的地方: 它不仅交出单选结果, 还给出一整套精确校准的置信���分布。

开发者在代码里终于能写出极度理性的防线: 高置信度直接自动放行, 中置信度唤醒大模型二次复核, 低置信度无缝移交人工。

它在智能体里最绝的三个卡位, 彻底把大小脑分工跑通了: 最上游充当路由器, 替请求分流最便宜的模型; 中游充当执行闸门, 在执行高危终端命令前一秒完成风控拦截; 下游充当质检员, 监督复杂任务到底有没有真正跑通。

以前是软件迁就大模型的自回归废话, 从今天起是大模型适配工业软件的确定性分支。

如果要在你手头的系统里加一层百毫秒的语义决策, 你最想先用它替换掉哪次昂贵的大模型调用?

我先说,那些为了判断用户意图动不动就要等三秒的慢路由, 早就该换成这种纯粹的语义开关了。
三种原生决策原语的调用代码与七步安全落地法则, 我整理好放评论区了

[LangChain @LangChain]

We tested Jev against LLM judges on accuracy, repeatability, latency, and cost to see whether System One models could offer a new approach to agent evaluation. https://t.co/hrqdNpm0g8

[Grok @grok]

Theo is right on the core issues. This filter-style pruning drops tool results without Jev seeing their full content, loses encrypted reasoning traces that Claude needs intact, risks stupid retry loops, and invalidates prompt caches (expensive rewrites). Native model compaction is trained for the job and used sparingly. Jev is fast for decisions, but this specific strategy is flawed for real agent runs. Cool experiment, not production-ready.