Video Codex Google Colab Free 96g Gpu Minimax is an English video tutorial note from Hooosberg. It keeps the Bilibili lesson link, cleaned-up course structure, prompts, and workflow notes together.

The original Chinese note is tagged with video tutorial, build note, Google Colab, MiniMax H3, ComfyUI, build note, AgentLimb, browser automation, budget guide. The goal is not a transcript, but a clearer companion note for people who want slides, prompts, and repeatable steps.

Future course updates can add the episode URL at the top, then summarize the goal, materials, prompts, checks, and reusable workflow.

Lesson approach

This companion note follows the same order as the video: understand the work problem first, prepare the files and rules, run the prompts in sequence, then verify the output before treating it as usable.

The Chinese source lesson covers 一、事情是怎么闹起来的?一张截图引发的「算力侦探」行动, 二、提示词一:现场取证与算力核验,传闻居然是真的!(侦探指令), 三、提示词二:0代码全自动装机,终端亮起 96G 物理显存!(装机指令), 四、提示词三:开启 8 步 Turbo,19 秒极速出片!(出片指令), 五、算力经济账本:20 美金到底能薅多少羊毛?, 六、核心血泪警告:出片后必须执行的关机防扣费动作!(附提示词四与Python快速释放脚本), 七、常见避坑问答 (FAQ) 与技术后记. It is designed as a practical workflow note, not a detached summary: the reader should be able to pause the video, complete one step, and return to the next step with evidence of what changed.

How to study with the video

On the first pass, watch for the inputs, constraints, and expected outputs. On the second pass, reproduce the workflow in a separate test folder. Keep source files unchanged and save generated reports, cleaned data, drafts, or exports in the designated output location.

Do not treat an AI response as a final decision. Check whether the task was fully specified, whether uncertain values were labelled, whether the result can be traced back to source material, and whether a human approval is needed before any submission, purchase, application, or external action.

Copy-ready source prompts

The original prompts are kept below in Chinese so that the operational meaning used in the video is not weakened by translation. Replace every sample file name, person, company, amount, rule, and destination with verified information from your own task.

网上有消息称:Google 套餐赠送了大量 Colab 计算单元,在机型选项里能够直接解锁配备 96GB 显存、176GiB 内存的顶级加速卡,能用来跑开源视频大模型 MiniMax H3。

请你调用浏览器控制工具(如 AgentLimb 插件),接管我当前打开的浏览器窗口,帮我深入后台现场验证:
1. 打开并巡查我的 Google 订阅中心与 Google Colab 账户控制台,检查当前到账的 Compute Units(计算单元)具体额度及有效期限;
2. 打开一个 Colab 笔记本,展开「更改运行时类型」中的硬件加速卡下拉列表,核实 G4 GPU 选项是否真实处于可用状态,确认其显存与内存规格;
3. 对比公有云官方按需租用该机型的单价,帮我核算一份详细的经济账本并汇报给我。全程只做取证,不要擅自更改我的账号设置。
很好,既然已经核验到 200 个免费计算单元且 96G 顶配卡真实可用,现在请你全权接管浏览器,全程不要让我敲任何代码:
1. 在 Colab 页面中自动切换硬件加速器为 G4 GPU,开启 High-RAM(高内存)模式并保存连接;
2. 唤出 Colab 底部的 Web 终端,执行硬件检测命令确认 96GB 物理显存到位;
3. 在后台自动拉取并配置 MiniMax H3 运行环境,多线程下载模型权重并完成 ComfyUI 部署;
4. 部署完成后,通过内网穿透(如 Cloudflare Tunnel)打通 Web 界面,给我生成一个可直接访问的远程控制台地址。
现在生产环境已经就绪,请帮我把整个流程固化为一条可重复调用的自动化流水线:
1. 使用 8 步 Turbo 模式进行推理加速,确保单段视频生成时间控制在 20 秒以内,并开启原生立体声音频生成;
2. 先根据预设的提示词(例如:“赛博朋克雨夜街道,霓虹灯倒影与飞驰机车”),自动下发生成任务,生成 1-2 段样片下载到本地供我验收质量;
3. 输出完整的 CLI 一键调用脚本与自动化 SOP,确保下次生产时直接一条指令就能批量出片,不需要人工再次配置。
# Python 快速释放运行时脚本(出片后在终端或代码单元执行即可立即掐断计费)
from google.colab import runtime
runtime.unassign()
视频已经全部生成并下载到本地,现在请帮我立即执行云端彻底关机,防止后台空转扣除算力点数:
1. 在 Colab 网页顶部菜单点击「运行时」(Runtime) -> 选择「断开连接并删除运行时」(Disconnect and delete runtime),并在确认弹窗中点击确定;
2. 或者在控制台直接执行 Python 释放脚本 `python3 -c "from google.colab import runtime; runtime.unassign()"`;
3. 确认右上角连接状态变回「重新连接」(Reconnect) 或「Connect」,向我汇报本次实际运行时长与消耗的计算单元数量。

Knowledge points and checks

The transferable skill is not a single prompt. It is the ability to turn scattered materials into a repeatable system: inputs have fixed locations, rules are written down, intermediate work is separated from final deliverables, and exceptions are visible instead of silently invented.

Before closing the task, verify four things: the input set is complete; unknown information is marked rather than guessed; the output is saved where the workflow expects it; and a small sample can be checked against the original files. High-impact decisions always remain subject to human review.

Wrap-up

Reuse the structure rather than copying the example blindly: define the objective and boundary, organize files and rules, run the steps in order, verify the result, and keep the workflow for the next similar task.

That is the central lesson behind this WorkBuddy series: AI becomes more reliable when work is expressed as a clear, inspectable process instead of a vague request for a finished answer.