CyberByte retweeted
AI 圈已经有人开始为「AI 断供」做末日准备了,甚至囤出了 10.5TB 的 AI 战略储备😂
黄仁勋离开 NVIDIA 后,Ruby on Rails 创始人 DHH 突然发了一条意味深长的推文:
「你最好提前备份几份开放权重模型,再准备一些能运行它们的基本设备。没人知道 NVIDIA 到底发生了什么,但这个世界刚刚失去了自由开放 AI 最重要的支持者之一。」
Reddit上,已经有人认真讨论起了「如果有一天开放权重模型不再发布,我们该提前囤些什么?」
结果评论区直接变成了 AI 末日生存物资展示大会。
一位叫 Kahvana 的玩家晒出了自己的 2TB NVMe 硬盘。
里面不只是几个聊天模型,而是囤了 Qwen3.8、Gemma 4、DeepSeek V4 Flash、Nemotron 等多个系列,连不同精度的量化版本都准备好了。
而且考虑得相当周全:
不仅有负责聊天和写代码的大语言模型,还有语音识别、语音合成、图片生成、Embedding、Reranker,以及 ComfyUI、MCP 工具、离线维基百科。
甚至还专门把最重要的模型复制到 NAS,并备份 llama.cpp 和 SillyTavern,确保不只是有模型权重,还能真正运行起来。
但这还不是最夸张的。
另一位叫 TheBigHODL 的玩家直接晒出了一份 10.5TB 的 AI 储备清单。
中国的 Kimi、DeepSeek、Qwen、GLM、MiniMax,美国的 Llama、GPT-OSS、Gemma,以及 Mistral 等模型,几乎把主流开放权重阵营都囤了一遍。
还不止如此。
从 Whisper 语音模型、FLUX 图片模型,到训练数据集、评测数据、完整维基百科、推理框架、量化工具,甚至一些开发工具的源码,都在他的存档范围内。
这已经不是简单地下载几个 GGUF 文件了,而是在试图给自己保存一整套能独立运转的 AI 工具箱。
他们担心的也很现实:
万一未来中国模型受到更多出口或访问限制,美国 AI 实验室又逐渐停止开放权重,今天随手就能从 Hugging Face 下载的东西,明天可能就没那么容易拿到了。
当然,这只是社区对未来的担忧,并不代表相关禁令已经出台。
但这种变化确实挺有意思。
以前本地 AI 圈比的是谁的显卡多、显存大、Token 跑得快。
现在有人已经开始比谁的硬盘大、谁的模型囤得全、谁断网之后还能继续用 AI。
以前末日生存玩家囤粮食、柴油、发电机,现在 AI 玩家开始囤模型、算力和离线知识库了。
而且仔细想想,这事还真不完全是杞人忧天。
云端 AI 再强,使用权终究取决于服务商;本地模型即使落后几个版本,只要权重、软件和硬件都在,至少还能自己运行。
CyberByte retweeted
After Andrej Karpathy released autoresearch, where an agent runs experiments in a loop on its own, I packed his method into one prompt that turns Opus 5.5 into a graph and can carry almost any task to the finish.
→ writes a check that can actually fail, first
→ splits the task into parts
→ runs the independent parts in parallel
→ loops each one until it passes the check
→ sends back only the part that failed
→ turns every mistake into a rule for next time
→ calls you only for the step that can’t be undone
Throw it at Opus 5.5 with the task you’ve been putting off for a week, and you’ll see the difference in one run.
The full prompt is in the comments.
How to tune it for your own work is in the article below. Save it before it gets lost ↓
CyberByte retweeted
The video shows an ‘unreleased’ humanoid at the 2026 robot exhibition. The design looks flawless and the movements are smooth, which is typical for such models.
Scenes like these demonstrate the sophistication of the models and attract a great deal of attention. It is harder to imagine a robot capable of operating for long periods without constant human intervention and carrying out tasks for which someone is actually paying.
Impressive demonstrations and real-world hours of unsupervised operation are still two different things.
#AI #Robotics #HumanoidRobots #PhysicalAI #Automation #NewTechnologies #AIRobotics
CyberByte retweeted
I built an absolute monster. My crazy idea to merge all the most powerful AI tools together made me $832 while I was asleep
opus 5.5 + REA + Jev + Dots + Grok Bot + Obsidian
Half a dozen utilities. A single unified ecosystem. Zero human intervention
Opus 5.5 -> acts as the core intellect, designing every assignment
Grok bot -> monitors X around the clock to spot emerging patterns
REA -> conducts deep analysis on any subject within a four-minute window
Jev -> executes concepts autonomously, turning them into reality
Dots -> weaves all six components into a seamless workflow
Obsidian -> functions as the archive containing 1,260 records with zero data loss
Before heading to bed, I assigned it a single objective: discover a profitable market gap and launch a working product by sunrise, keeping the expenses under $15
I went to sleep at 11:40 PM, and here is what unfolded over the course of seven hours:
12:12 AM: Grok bot uncovers 1,900 distinct posts from delivery workers venting about a shared frustration: tracking their income across three separate platforms
1:05 AM: REA evaluates the competitive landscape, finding 14 existing paid solutions starting at $19 monthly, with users universally criticizing the exact same two software glitches
2:30 AM: Opus 5.5 drafts an application architecture designed to resolve both flaws, consisting of nine screens and a single database, without asking me a single question
4:10 AM: Jev programs the software, deploys it live, and writes the promotional landing page.
5:45 AM: Dots generates a 30-second promotional animation and publishes it alongside three forum threads
7:20 AM: Obsidian logs every single action, financial expense, and misstep along the way
Upon waking up, I discovered a fully functioning application, 243 registrations, and 31 paying customers subscribed at $9 each, bringing in $832 before my morning meal
The entire nightly overhead amounted to $5.10 for API tokens plus a $4 web domain, totaling $9.10 -> yielding a 91x return while I slept
The system even exercised self-control: when Jev attempted to purchase a $55 web address, it hit the preset $15 threshold and automatically opted for a $4 alternative instead
The primary misconception is that individuals turn to AI merely for answers: instead, I provided mine with a clear objective, a strict timeline, and a financial ceiling
I did not author a single line of programming code, nor did I open a single browser tab: I simply went to sleep
While everyone else is holding out for artificial general intelligence, I already have six autonomous utilities pulling the night shift
CyberByte retweeted
分享一套由架构师 Brij Kishore Pandey 提出的标准,
是目前避免 Agent 上下文膨胀最优雅的工程解法。
90% 的人都把 CLAUDE.md 当成普通的提示词文件,难怪你的 AI 永远像个乱改代码的弱智实习生:
把代码规范、API 格式、业务逻辑全往根目录死命塞, 模型还没动工,注意力就被几千行无用说明给稀释光了。
想让 Claude Code 真正变成常驻在你代码库里的资深架构师, 你需要的不是写更长的 Prompt,而是这套把上下文物理切开的项目解剖学:
1️⃣ 根目录只放极简骨架(四维契约): 只写系统是做什么的目标、文件在哪的地图、允许禁止什么的死规则,以及怎么完成任务的工作流,绝不堆砌具体代码;
2️⃣ 业务上下文下沉到子目录: 很多人不知道 Claude 会自动读取子目录的独立记忆,在 src/api 和持久化目录下分别建局部 CLAUDE.md,模型钻进哪个模块才读哪份手册;
3️⃣ 工作流封装进 .claude/skills: 代码审查、重构、发布流程不要每次口头重打,全部固化成可复用技能,用完即走;
4️⃣ 刚性护栏交给 .claude/hooks: 关键改动前强制触发自动化检查,不符合规范直接代码级打回,绝不靠模型的自觉。
提示词做减法,工程结构做加法, 保持上下文的最小化与精确性,模型才能在你的代码库里大杀四方。
你现在的项目根目录,还塞着一份几千字的大杂烩提示词吗?
《九霄云外》
Beyond the Mist
三十三重碧落外,缥缈烟霞隐圣贤。
Beyond the thirty-three realms of azure skies,
Where mists and starlight weave and rise,
本是凡尘一粒种,风骨错铸化云烟。
The gods are but mortals who dared to wake,
Carved from the dust for spirit's sake.
叩问南天迷雾里,依稀错认白玉京。
Lost at the South Gate, wrapped in white,
Is this the Jade Capital bathed in light?
In clouds where earth and heaven blend,
I lose my way, where worlds transcend.
🤖 Made with AI
Upgrade your @Grok app
CyberByte retweeted
GPT-5.2-Codex launches today.
It is trained specifically for agentic coding and terminal use, and people at OpenAI have been having great success with it.
CyberByte retweeted
Replying to @starzq @newton_xyz
完全同意!大部分项目都在卷基建,但用户只关心实际价值。BasisOS的更新确实给应用层带来了新玩法,特别是内容挖矿模式很有潜力。
AI Agent必须能帮用户赚钱或省钱,否则就只是个花哨的聊天工具
CyberByte retweeted
Replying to @AMAZlNGNATURE
Seal bruh is not ready to take a small one, negotiating for a big fish.





























