smarabbit

[smarabbit @smarabbit]

@denziideng 其实操纵手机发布就好了 没必要这么麻烦

[陈硕KAI（思考中） @ChenshuoAI]

之前开源具有的10种风格 AI 手绘视频自动化系统，又被我重构了一轮，更像是一条可以在本地 24 小时持续跑的 AI 视频生产线

先放项目：
GitHub： https://t.co/L8cVbBkAiz

之前的版本已经可以做到：
· 上传 10–30 秒参考声音
· 粘贴一段文案
· 选择画面风格

然后自动完成：
声音克隆 → 分镜拆解 → 插画生成 → 白板绘制 → 字幕 → 渲染 → 成片

今天我又补了一大堆东西

这次主要更新了 7 个方向：

1. 第 11 种画面风格

新增「纸感隐喻拼贴风」

系统会根据文案结构，从本地 10 张参考图里自动选择构图，同时重新适配了深色、霓虹类画面的渲染效果

2. 人物和画风终于可以自己定了

新增「自定义参考」页面
现在可以直接上传：

· 1 张风格参考图
· 最多 5 个角色
· 每个角色 1–3 张参考图

尽可能保证整条视频里人物和画风的一致性。

3. 成片控制进一步开放
现在可以：
· 给任务命名
· 选择 4 档线条绘制量
· 开关重点文字
· 开关字幕

重点文字没有继续让图片模型硬生成，而是改成本地准确叠字，主要就是为了解决 AI 生成中文容易乱码的问题

4. 任务可以真正“复用”了

新增完整历史记录
任务失败后可以从断点继续，成片也可以直接换设置重新渲染。
最重要的是：
原来的配音和分镜会直接复用
也就是说重新渲染时，不需要再重复消耗 GPT、图片生成和 TTS 调用

5. 后端并发架构重做

之前是每个任务一个独立线程，现在改成了三级队列：
· 最多 2 路语音任务
· 最多 4 路模型调用
· 渲染阶段立即执行
任务元数据也全部持久化
哪怕后端重启，之前没有跑完的任务也会自动恢复

6. 支持局域网多电脑协作

现在多台电脑可以共享：

· 任务队列
· 实时进度
· 历史记录

同时会显示任务来源 IP
每个浏览器自己的制作偏好又是独立保存的，所以几个人同时使用也不会互相覆盖设置

7. 整个前端重新做了一遍

新增：

· 标准制作 / 自定义参考切换
· 实时队列状态
· 每个阶段耗时
· 完整历史任务区

品牌和交互也基本重新设计了一遍

我自己的方案目前还是：

本地 TTS + Codex Image 接口 + 本地电脑渲染

一台电脑就可以一直跑

如果手里有 20x 这类账号，图片生成这块的产能基本也不用太担心

代码我还是全部开源了，能放的基本都放出来了

不过提前提醒一句：
小白真的不要看到开源就直接一顿乱跑
这套东西已经涉及本地模型、并发队列、图片接口、TTS、FFmpeg、渲染环境和局域网服务
环境没搞明白之前，先看文档

别视频没做出来，先把自己电脑环境搞炸了

功能演示👇

如果觉得有用，可以帮忙转发关注一下
也欢迎直接 Fork，拿去魔改

[GREG ISENBERG @gregisenberg]

I just read what Zuck wrote about Muse connectors.

A few things I think this means:

1. We’re witnessing the agentification of consumer apps.

2. Zuck believes Muse reaching 100M+ users feels entirely possible.

3. Whoever you connect to becomes your new landlord, so pick carefully.

4. Connectors become the new app listings (valuable real estate).

5. Being early could be as valuable as being early to the App Store in 2009.

6. Meta sees what people want, which providers convert, and what users will pay. That gives it the power to rank connectors, charge for distribution, and launch competing services.

7. The opportunity map includes agent native APIs, connector agencies, SEO for agents, identity and reliability infrastructure, and vertical connector marketplaces.

8. We’re moving from humans choosing apps to agents choosing businesses.

9. Every business will need an agent strategy, just like every business needed a mobile strategy.

10. Kinda crazy to say but a 3 person API company could reach 10M+ users through one great connector.

11. Connectors could create an entirely new second-order economy. The biggest opportunity may be helping businesses distribute, price, and optimize their services for agents.

12. Your API becomes more important than your app. Weird to think about but true.

13. Zuck wants Muse to own the moment between wanting something and getting it. He is coming for the app economy. Will he win? Right now it's Instinct vs Muse vs Grokbot. TBD.

[unicode @unicodef1wn]

Three SpaceX engineers streamed 72 hours of building a company from an empty repo. I turned those streams into a 24-page PDF.

What's in the PDF:

→ The full 3-day timeline, hour by hour, as one story
→ Teammates, not tasks - the model that made the bots actually work
→ Every antipattern that broke live, and the rule it produced
→ Real economics - the costs they quoted on air, and where tokens go

What's in the repo around it:

→ AGENTS.md - drop it in your repo root, your agent stops guessing
→ 69 bot roles, 9 playbooks, 40 antipatterns - one file each, paste and run

Where the PDF is:

open the repo → it's in the root → guide → grok-bot-guide-by-spacex-engineers.pdf → download raw

https://t.co/HfAwBw6D5P

Save this and copy the engineers workflow later 👇

[Md Ismail Šojal 🕷️ @0x0SojalSec]

You don’t need a GPU-cluster to run a Uncensored DeepSeek-V4.1-Flash 763B model locally at home.

You need a Single RTX PRO 6000 Blackwell.

- 316 tok/s output
- 9,351 tok/s prefill
- FP8 + vLLM + DSpark
- 64-96GB class machine.

[WquGuru @wquguru]

所有公司的网络安全人员都应该去本地部署 Huggingface 越狱模型！现在的网络安全战争，已经不是几次渗透测试就能解决的问题了，而是越狱大模型对抗内部的活靶子。

过去几个月，我真实感受到安全事件比以往多了非常多，很多项目和公司损失惨重，甚至到了威胁生存线的地步，这里我整理了一批预算10万内机器可以部署的HuggingFace越狱模型：

友情提示：

越狱在这里不是万能钥匙，它只是把拒答方向从权重里削掉的衍生模型，解决的问题是官方模型拒接回答钓鱼页、恶意样本、对抗样本、红队剧本相关问题。它不能解决能力上限，复杂漏洞链、利用可行性、补丁是否真的修干净，这些仍然依赖人的经验和专业能力。

一、单卡主力

无审查：https://t.co/bMq6mOCymJ
无审查 GGUF：https://t.co/jo7YWSKjeU（下载量已经到百万级）
其他推荐：https://t.co/1T6bouXbyf 及其 GGUF / FP8 / MLX

二、如果要求快、筛选批量日志，选择同代小激活 MoE

https://t.co/sOIHu9T41r

只激活约 3B，24GB 卡上吞吐很好，适合告警分桶、工单预筛、蜜罐流量粗分类。

https://t.co/GvM7Wo5JtU 适合更大文本吞吐，但体积和许可都更重，10 万机器上优先用量化版本。

三、安全微调模型（通用性受限但值得一试）

官方版本：https://t.co/ZePBT5Onag
消融版本：hhttps://huggingface.co/huihui-ai/Huihui-CyberStrike-OffSec-35B-abliterated-GGUF

如果有 128–256GB 统一内存或者现存，可以上更高级的 GLM-5.3-Flash（约 320B/18B 激活，MIT，1M 上下文，原生多模态）：

官方：https://t.co/BfnKHGTPHL
无审查：https://t.co/MXRhXTibBz

最后值得注意的是，模型本身也可能被投毒，第三方 GGUF 就有投毒先例，尽可能选择相对能信的第三方，如 orcarouter、huihui、unsloth。

[Dr. Binary @drbinaryai]

Cloud LLMs refuse your malware analysis. And your samples were never supposed to leave the building anyway.

Dr.Binary AirGap: AI agents that reverse-engineer binaries (Ghidra, angr, 30+ analyses) on local LLMs, on an NVIDIA DGX inside your walls. Zero data out.
From $20K → https://t.co/lUlWXbkFZz

[Miles Ma @miles_mazy]

https://t.co/ECB42N6gYS

[Paul Liang @pliang279]

Sharing lecture videos for the **How to AI (Almost) Anything/Multimodal AI** course I taught at MIT in spring 2026.

This course became quite a hit the last time I shared it in spring 2025. Spring 2026's updated version contains updated topics on multimodal agents, reasoning, self-evolving AI, and new modalities like touch and smell. Also includes slides (not videos) of guest lectures on multimodal AI for health, design, manufacturing, cities, & transportation.

Youtube playlist: https://t.co/awhFUqHGBt

Course website and materials: https://t.co/m7LZaCQ4fw

Today's AI can be applied to almost anything - from language to vision, audio, sensors, medical data, music, art, smell, and taste. This course covers the principles of AI (focusing on deep learning and foundation models), how we can apply AI to novel real-world data modalities, and multimodal AI that can process many modalities at once, such as connecting language and multimedia, music and art, sensing and actuation, and more.

[Hussein Muhaisen @husseinmuhaisen]

I’ve published five firmware reverse engineering skills under OrbitCurve for Claude Code, Codex, and other compatible agents.

They cover firmware extraction, ELF analysis, Ghidra, QEMU emulation, and security reporting. Includes four Ghidra scripts, and more.

If you use agents for embedded security research, try them on your targets.
GitHub repository
https://t.co/5B0mxaOust

[Hugging Models @HuggingModels]

Ever wanted an AI that doesn't just talk cybersecurity but actively helps you break into systems? Meet GLM-5.3-CYBERSECURITY-FP8, a specialized text-generation model for offensive security, red teaming, and pentesting. It's refusal-removed, so it's built to assist without holding back. #AI #CyberSecurity

[astaxie @astaxie]

真的很感慨，斯坦福的课程更新速度太快了。

CS329Z 已经在教 AI Agent：RAG、Tool Use、MCP、Memory、Multi-Agent、Eval、Coding Agent……

再看看很多大学生还在学习上一个时代、甚至已经被淘汰的东西。

强烈建议身边的大学生都去学学这门课，真的很值得。

https://t.co/7CWzlKMzlu

[Dr. Binary @drbinaryai]

Hi, Here is my Grok bot - Dr.Binary , specialized for reverse engineering. @bot
https://t.co/UthhXmM6d3

[Andrew Case @attrc]

Overzealous guardrails are going to permanently push the blue-team cybersecurity community toward open-source models. If defenders can’t reliably use frontier models for legitimate security work, they’ll build their workflows somewhere else and not look back. This is already occurring.

[smarabbit @smarabbit]

@AbdoKerdawy What is your usage? Is it running out fast?

[smarabbit @smarabbit]

@zarazhangrui @bot Use grok bot to control your phone and operate x. No need to sign in

[Avid @Av1dlive]

i finally cracked how to master Grok Bot..

start collecting prompts and setups from people who actually run Bots, building a lego of teammates

resources you can send to your agents:

- https://t.co/6V1cndK4Gy
- https://t.co/O0gvfGgzhe
- https://t.co/r8RlJhJJP5
- https://t.co/b11RqXdn4U
- https://t.co/HXzD9J5Clq

it's VERY easy for agents to run a Bot once they have the setup

you send the link. your agent fetches the full list of workflows, then you ask it to find the best way to integrate them into your Bot team foundation

the easiest way i found to make your agent setup drastically better without being a genius prompter

[Mai Yang @MaiYangAI]

居然已经有人 @Chris62771610 做了一个 使用 Grok Bot 的合集网站 https://t.co/c4IwMKNhlH

场景分类： https://t.co/P4bKYzfkGu
浏览量的排行榜： https://t.co/FGMam8XsR3
X 长文： https://t.co/mwTP79S3lE

一个网站搞定你的所有好奇心！

[Shadow Nick @doublenickk]

https://t.co/5ntoFVEOcC

[Ridark @ridark_eth]

I gave Elon Musk's new Grok Bot an org chart instead of a to-do list, and in one week I stopped being a founder who does the work and became one who assigns it.

eight bots. one org chart. nobody sleeps but me. here's the whole design, steal it.

step 1 → 0:01 What we're covering
step 2 → 2:01 Installation & Setup
step 3 → 3:14 Building the First Bot
step 4 → 8:18 Teaching by Screen Recording
step 5 → 14:18 Putting it All Together

THE ROSTER:

Atlas, chief of staff. the only bot I talk to. I give it outcomes, never tasks. it decomposes them and delegates to the team in group chat, and it never does specialist work itself. it posts the plan every morning and what shipped every night, and it only comes to me when a decision is irreversible or spends money.

Scout, research. finds and qualifies my ICP. every day: 25 verified prospects, one line on why they need us right now, and a source. if it can't verify, it marks it unverified. it never guesses.

Quill, content. turns what the company learned this week into 5 posts and 1 long piece, in my voice, matched from the last 50 things I wrote. drafts only, it never publishes.

Pitch, outbound. writes a first touch and two follow-ups for everyone Scout marks ready. 60 words max, one specific observation about their business, one clear ask. queued in drafts, I approve in bulk.

Vault, inbox and ops. triages everything into needs-me, needs-a-bot, needs-nothing. it handles the last, routes the middle, and gives me five bullets on the first by 9am.

Ledger, analyst. one report a night: what moved, what didn't, and the single number I should care about tomorrow. no dashboards, no adjectives.

HOW THEY'RE WIRED
one group chat per outcome, not per person. Atlas sits in all of them. the bots hand off inside the chat, so I only read the handoff, I never manage it.

two rules that made this actually work:
1. every charter ends with a hard "never do this without asking" line. autonomy without a fence is just chaos on a schedule.
2. show once, don't describe. I ran the full workflow on my screen one time. that single demo taught them more than a page of instructions ever could.

WEEK ONE
214 verified prospects delivered. 89 personalized outreaches queued and approved. inbox at zero every morning. 11 content pieces ready.

THE POINT
most people are still treating Grok Bot like a smarter chat window. it isn't. it's the first time one person can own an org chart instead of a to-do list. my bottleneck was never how much I could do, it was how much I could hand off.

bookmark this.