Oliver Perez

[International Cyber Digest @IntCyberDigest]

‼️ A hacker says an 8GB Nvidia crypto-mining card hides 80GB of memory. She unlocked it, and in her tests it outran Nvidia's A100 AI chip.

The CMP 170HX uses the same chip as the A100, with the extra memory locked off by fuses. Kate broke the lock by glitching its voltage with a homemade STM32 board that she says an AI model designed.

That switched on up to 126 compute units, more than the 108 on any A100. It hit 85.9 TFLOP/s in FP16 against 72.5 for a retail A100, with all 80GB usable.

Each card took 1 to 48 hours of glitching. One of her 10 only gave up 64GB.

[Shlcy @CY160909]

挖到一个硬核的开源Skill

它能把手绘草图、扫描图纸，甚至 CAD 截图，重新绘制成可编辑的 DWG。

而且它是不只负责画，还会检查重绘结果

以前遇到没有源文件的老图纸，往往只能打开 AutoCAD，一根线一根线重新画。

这个项目直接把整套流程交给 AI：

识别图纸中的线条、圆弧和尺寸信息
自动生成 CAD 绘图数据
调用 AutoCAD 绘制可编辑 DWG
整理图层，导出 DXF 和预览图
对比原图，检查重绘偏差

最让���感兴���的是它的尺寸处理机制。

哪些尺寸有明确标注，哪些是按比例推算，哪些只是 AI 推断，甚至哪些根本无法识别，都会区分处理

这比单纯把图片转换成线条实用得多。

需要整理老图纸、重建 CAD 文件的人，这就是神器

我建议关键尺寸最好人工复核，别当成工程级精准转换。

GitHub：
https://t.co/MGyihzhq3v

[小墨同学 @xiaomovps]

DeepSeek Harness 现在已经不满足于做 Coding Agent 了，开始尝试让 Agent 自己接活、干活了！

最近发现 DeepSeek 团队开源了一个很有意思的项目：dsh-superagent，目前 GitHub 已经超过 2.2 万 Star

以前我们使用 Coding Agent，基本都是自己提出需求，让 AI 写代码、修改文件，然后等待结果

但是这个项目想解决的是另一个问题：能不能让 Agent 直接参与团队的日常开发，而不是每次都需要有人打开聊天窗口给它安排任务？

我看了一下它的设计，主要有三个比较���得关注的地方：

1、直接对接 GitHub 和飞书

Agent 可以监听 GitHub 中的任务和讨论，也能选择接入飞书

团队成员提出需求后，经过权限和规则检查，就能创建对应的工作任务，不需要每次手动打开 DSH

2、Router + Worker 分工协作

Router 负责判断和分配任务，Worker 负责具体执行

每个任务都有独立的 Agent 会话和工作目录，还能按照并发限制处理多个任务，避免所有工作都挤在一个对话里面

3、尽可能减少人工干预

Agent 可以自己处理任务、修改代码，并通过 Pull Request 交付结果

只有需要人工审核、补充信息或者遇到问题的时候，才通知开发者

我觉得第三点才是这个项目最有价值的地方

以前的 Agent 更多是我们主动找它帮忙，现在开始尝试让 Agent 融入原本的开发流程，���己处理被授权的工作

当然，目前它还只是一个研究原型，主要围绕单个 GitHub 仓库设计，需要自己部署和配置，距离成熟的生产环境还有一定距离

但是从这个方向来看，DeepSeek Harness 未来可能不只是一个写代码的工具，而是一套可以持续协作的 Agent 系统

项目地址：
https://t.co/tSPsyRu9AM

[Stanford NLP Group @stanfordnlp]

The Playlist page for CS329Z: Engineering AI Agents is now up! Bookmark it!

𝍣 With 4 lectures now available!!! 𝍣

https://t.co/zKZScE3HO5

[alphaXiv @askalphaxiv]

"The Surprising Effectiveness of Deleting Weights in LLM Post-Training"

LLMs can learn entirely new capabilities by simply deleting less than 0.05% of their pretrained weights, without modifying any of the remaining parameters...?

This new paper proposes Deletion Tuning, which learns a binary mask that determines which weights to remove from the model's feedforward layers. Unlike conventional fine-tuning or LoRA, it doesn't learn new weight values, but instead learns which existing connections to disable.

Across RL, SFT, and distillation, Deletion Tuning matches full fine-tuning at the largest scales tested.

On a 30B-parameter model, it achieves 58.54% vs 58.43% for full fine-tuning, with an adapter roughly 4,000x smaller.

It also preserves previous capabilities better, while conventional fine-tuning methods lose up to 5.4% of prior-task accuracy.

The reason this might work is because full fine-tuning tends to preserve the model's main weight structure while making changes outside its dominant components. So selectively deleting a tiny fraction of weights can produce a similar pattern of changes without disturbing the rest of the model.

https://t.co/UB4aoYcFGW

[PapersWithCode @paperswithcodee]

Congrats @thinkymachines: Inkling 2 Small is the best open model on LongBench v2! 🏆

Now that its weights are out under Apache 2.0, it scores 70.6, ahead of Beam (67.5) and Hy3 preview (65.4).

It's also the top open model on JobBench and FrontierCode 2.0.

[Erick @ErickSky]

IDM quedó obsoleto.

Este download manager open source escrito en Rust es 5 veces más rápido, gratis para siempre y no tiene ads ni trial.

• 1 GB en 1 segundo (IDM: 5.5 s)
• Soporte completo de torrents + magnet
• Descarga videos con un click (yt-dlp)
• Se integra con Claude, Cursor y VS Code vía MCP
• Windows + Linux + Mac (experimental)
• 86 idiomas

Cuando lo tienes instalado te sale un botoncito en los videos de youtube para que lo puedas descargar facilito :)

REPOOO👇

[NIK @ns123abc]

If it weren't for Jensen, OpenAI (and by proxy Anthropic) wouldn't even exist.

[Plato | Seller FDE @0xPlato]

Cloudflare 把 Kitesurf 开源了 🐙

Kitesurf 一个专为 AI Agent 造的浏览器引擎。

现在可以：
📸 给 Agent 配一只"眼睛" — 丢个网址进去，拿回一张截图或一个 PDF。日报自动生成、页面监控、给多模态模型喂图当上下文。

📄 批量抽取网页内容 — 抓取、竞品比价、资讯聚合、每天扫几百个页面只取正文。这类"打开—取走—关掉"的活儿，它最擅长：无状态、可并发、账单便宜。

🔁 换掉现有的浏览器后端 — 原本用 Puppeteer / Playwright / MCP 的脚本，连接地址后面加个 browser=kitesurf 就切过去了

https://t.co/TdZv1EZc9i

[Oliver Perez @OliverPeredckc]

@MatthewWrifovz This look — that's the whole post.

[何笠聊 AI @HeLi8866]

简直牛逼克拉斯��！

GitHub 上这 10 个开源仓库，够你省去很多订阅费了！

1. TradingAgents
让好几个 AI Agent 分工协作来做交易分析
https://t.co/NMcQnKOdvp

2. LibreChat
ChatGPT、Claude、Gemini 这些模型放在同一个界面里用
https://t.co/CvSmgVfY3i

3. HyperFrames（HeyGen 开源出来的视频引擎）
https://t.co/QtJPrAKL7f

4. Fincept Terminal
开源的金融数据终端，不少人拿它对标彭博
https://t.co/F5NkZLMawz

5. MoneyPrinterTurbo
给个主题，AI 就能产出一条短视频
https://t.co/4lU8Hn0fZw

6. Agentic Inbox
Cloudflare 开源的 AI 邮箱助手
https://t.co/UxuHHdprsM

7. VoxCPM
能把文字转成语音，也能克隆音色
https://t.co/EtW9cLJulU

8. Flowsint
专门梳理公开情报（OSINT）的工具
https://t.co/zgYx8S4Oyq

9. agent-skills
一整套给 Claude Code 用的技能包
https://t.co/LO6O5iUvCB

10. Nango
帮你把各家 API 接到一起的开源平台
https://t.co/fmV63rg1Ev

这些都是正经能干活的项目，不少每个���要掏��的软件，在这里都能找到开源的替代。

与其天天逛 AI 工具导航站，不如直接上 GitHub 翻翻，好东西其实都在这。

[Tom Dörr @tom_doerr]

Compiles raw documents into a structured wiki using reasoning-based retrieval without relying on a vector database.

https://t.co/WU9gioojt2

[Mike Investing @MrMikeInvesting]

This is absolutely absurd…

Director at Nvidia for 25 years, Mark Stevens sold $947M of $NVDA stock exactly near the top.

Since then Nvidia is down beyond -30% & counting.

Most interesting part is within his 25 years at Nvidia, he’s never sold his stake till just last month.

Talk about insider trading taken to the next level.

How this is allowed is beyond me…

[Nic Vandewetering @nvandewetering]

ChatGPT 6.1 Terra will get 90% of your work done. Stop wasting so much money about the big models.

Everyone is waiting for the next frontier model. Meanwhile GPT 6.1 Terra can use my computer in ChatGPT for a fraction of the usage. I think cheap models with computer use are going to do more real work than the expensive ones.

Am I wrong? What have you actually gotten Terra's computer use to finish?

[Codex Releases @CodexReleases]

Codex CLI 0.168.0 is out!

Highlights:
- Browse and set a default model provider for new chats under Sessions → Model providers
- Enable https://t.co/aiCUF2DHJf_workspace to view a side conversation beside the main chat, with resizable panes and a zoom shortcut
- Attach recordings when using an audio-capable model; Codex skips confirmed rejected recordings on later turns and uses clip duration to budget short clips in context
- Enable features.image_generation_animation to generate or edit animated GIFs with optional frame count and timing controls

Complete details in thread ↓

[min heo @minheo93698092]

클로드가 IPO전에 엄청난걸 쏟아내고 있습니다.

이거 대박입니다.
회사들이 그동안 가장 구축하고 싶었던 온프레미스 환경을 모델과 하네스 끝판왕인 클로드에서 가장먼저 공개했습니다.

Anthropic이 이제 Claude Managed Agents를 회사 자체 서버나 클라우드 안에서 직접 운영할 수 있게 해줘서, AI 에이전트가 회사 내부 코드·시스템과 민감한 데이터를 다루면서도 모든 것이 외부로 나가지 않고 안전하게 자동으로 일을 할 수 있게 됐습니다.

Anthropic 문서를 보니 에이전트 서버, 샌드박스, 인증 정보, 세션 기록까지 전부 우�� 쪽에 둡니다.
모델 ��청은 Anthropic 직접, Bedrock 같은 클라우드, 자체 게이트웨이 중에서 고를 수 있어요.

다만 기본 설정에서는 모델을 부를 때마다 대화 기록과 도구 결과가 Anthropic으로 간다고 문서에 적혀 있어요.

여러분은 에이전트 돌릴 때 직접 호스팅하는 쪽인가요, 그냥 맡기는 쪽인가요?
https://t.co/Gfu1Gdq8X0

[Oliver Perez @OliverPeredckc]

Got more better quality videos. Check the bio link.

[Oliver Perez @OliverPeredckc]

More coming soon. DM me if you want.

[Anthropic @AnthropicAI]

Our initial public offering: https://t.co/Ah4BkI2loD

Legal disclaimer: Anthropic has filed a registration statement (including a prospectus) with the SEC for the offering to which this communication relates. Before you invest, you should read the prospectus in that registration statement (available at: https://t.co/Z44axxzf9b) and other documents Anthropic has filed with the SEC for more complete information about Anthropic and this offering. You may get these documents for free by visiting EDGAR on the SEC Web site at https://t.co/94ojsWZTKi. Alternatively, copies of the preliminary prospectus may be obtained from: Goldman Sachs & Co. LLC, Attention: Prospectus Department, 200 West Street, New York, NY 10282, by telephone at [phone], or by email at [email protected]; or Morgan Stanley & Co. LLC, Attention: Prospectus Department, 180 Varick Street, 2nd Floor, New York, NY 10014.

The securities may not be sold, nor may offers to buy be accepted, prior to the time the registration statement becomes effective. This post is not an offer to sell, or a solicitation of an offer to buy, the securities in any jurisdiction where the offer or sale is not permitted.

[shadcn @shadcn]

We've built some really amazing (agentic) features into UI SDK. We'll show you in the coming days.

https://t.co/uygHmhr1Wa