今晚别刷Netflix了。
去看这个SpaceX AI工程师1小时的课程。
这是我见过最清晰、端到端讲透GrokBot到底该怎么真正用起来的课。
从单agent开始,一路到loop、graph、多智能体协作、verification和agent-friendly架构。
不管你是从没认真跑过agent,还是已经天天在用GrokBot,看完都会把很多“原来如此”的点一次打通。
真正核心就这几层:
1. 99%的人只用1个agent,没用loop和graph,只发挥了1%实力
2. 真正的新stack是Chief of Staff + PM + 20+ workers全自主团队
3. 怎么用高质量skills让agent像真工程师一样干活(pstack)
4. 怎么把代码和架构改成agent-friendly(Dune、contract、CI)
5. 怎么用verification让自己真正敢把活交给它们,而不是一直盯着
大多数人花好几个月才拼出的全景图,这节课直接给你画完。
把时间空出来,这可能是你这个月最值的一堂课。
Two Anthropic seniors just made Karpathy's loop 1000x better with "Graph Engineering" - dropped 11-page PDF
the shift: the agentic systems got 1000x better the moment you wired agents into a graph
here's the playbook in 6 steps:
step 1 → build one loop: generate, critique, revise - one self-review cycle beats a smarter model with none
step 2 → add tools: search, code execution, database - thinking without tools is hallucinating
step 3 → go parallel: spin up agents in separate worktrees - same repo, different branches, no conflicts
step 4 → add a graph: agents write findings as typed nodes and edges - not transcripts - every claim keeps its source
step 5 → ground your evaluator: it checks claims against graph edges, not vibes - "Triple not found" beats "seems off"
step 6 → the graph survives every session - your agents stop rebuilding context from scratch
the result: Karpathy ran 1 agent in 1 direction - this system runs 1,000 with shared memory - same model, it's the architecture
read this 11-page PDF and paste it into your Claude - you won't regret it
bookmark - then read the article on building graphs from scratch ↓
Harsh Seth retweeted
/goal is f*cking insane.
You can literally get your AI agents to work for HOURS without manual intervention.
Already active in Claude Code and Codex - you need to use it now.
Use this prompt and your agents will complete any task on autopilot:
Harsh Seth retweeted
Stanford CS336 上,Tatsu 讲了一节 LLM 架构课,把过去 3 年所有主流 LLM 拆开,看它们的共通模板
结论挺爆:90% 的架构选择已经收敛,你随便挑一个开源大模型,它跟其他模型在这些维度上几乎一模一样
讲师的原话
- 2024 年大家都在 cosplay Llama2
- 2025 年的主题是「怎么训得不崩」
- 2026 年的主题是「怎么扛住长上下文」
下面是 2026 年开源 LLM 的标准模板 你训自己的模型可以直接抄
【架构层 已经收敛的 7 件事】
1)Layer Norm 挪出残差流(pre-norm)
原版 Transformer 把 LN 放在残差里 几乎所有现代模型都挪到外面
原因:keep your residual stream clean 梯度反传更稳
2)RMS Norm 替代 LayerNorm
LayerNorm 的减均值 + 加 bias 那部分实际没怎么帮上忙
丢掉之后 flops 只省 0.17% 但运行时省到 25%
(瓶颈在数据搬运 计算反而次要)
3)所有 bias 项全删
跟 RMS Norm 一个道理 系统层省内存搬运
4)激活函数用 SwiGLU 或 GeGLU
gated linear unit 几乎所有现代模型都用
Llama 系 / Qwen / Mistral 用 SwiGLU
Google 系(Gemma / T5)用 GeGLU
区别极小 选哪个都行
5)位置编码用 RoPE
2024 年之后基本统一了
原理:把每对维度按位置旋转一个角度 让 inner product 只依赖相对位置
6)Transformer block 串联(不是并联)
GPT-J / Palm 试过并联 现在基本被放弃
串联的实现优化得太好了 并联省的那点系统开销不值得损失表达力
7)Layer norm 可以「撒」
哪儿不稳就在哪儿加 LN
attention 之前能加 之后能加 两边都加(double norm)也可以
现代模型很多这样做
【超参数 已经收敛的 5 个数】
1)feedforward 维度 / hidden 维度
- 非 GLU 模型:4 倍
- GLU 模型:8/3 ≈ 2.67 倍(因为 GLU 多一组矩阵 要保持总参数量)
- Llama 系:3.5 倍
- T5 1.0 试过 64 倍 后来 T5 1.1 改回标准 别学
2)head 数 × head 维度 ≈ hidden 维度
几乎所有模型都遵守 T5 是为数不多的例外
3)模型纵横比(hidden / 层数)≈ 100
太深 pipeline parallel 难做
太宽 表达力受限
100 这个数字是系统约束 + 表达力的平衡点
4)vocab size
单语模型:30K 左右(早期 GPT-2 那种)
多语 / 通用模型:100K-200K(GPT-4 / Llama 3 / Gemma 都在这个范围)
现代基本都是后者
5)weight decay
仍然普遍使用
但研究发现它在 LLM 里干的事其实是优化器干预 让你最终能收敛到更深的最优点
跟你想的「防过拟合」没什么关系
所以别因为「单 epoch 不会过拟合」就把它关掉
【稳定性 三个救命 trick】
训练大模型最怕中途 loss 突然飙升 然后 NaN 全军覆没
现代模型用三个 trick 防这件事
1)Z-loss
output softmax 的 normalizer 容易爆
加一个 (log Z)² 的正则项 让 Z 始终接近 1
DCLM / Olmo 都用
2)QK norm
attention 的 Q 和 K 在矩阵乘之前各加一个 LN
让 softmax 的输入永远是单位尺度
multimodal 圈先用起来 现在所有大模型都加
3)Logit soft cap(仅 Google 系)
attention logit 用 tanh 硬封顶
Gemma 2/3/4 都在用 但会损失一点点性能 慎用
【Attention 两个新趋势】
1)GQA(Grouped Query Attention)几乎统一
原版 multi-head 推理时 KV cache 会让算术强度崩到 1/h
GQA 共享 K 和 V 但保留多个 Q
表达力几乎不损失 推理成本砍掉 80%
现在所有要做生产部署的大模型 没有不用 GQA 的
2)局部 + 全局 attention 交替
处理长上下文的新方式
Cohere Command A 起头 现在 Llama 4 / Gemma 4 / Olmo 3 全在用
比如每 4 层有 1 层 full attention 其他 3 层是 sliding window 只看附近的 token
比纯 SSM 更稳 比纯 full attention 便宜得多
(Qwen 3.5 做了变体 把 sliding window 那 3 层换成 SSM)
收尾一句
如果你正在训自己的 LLM,上面这一套就是 2026 年的「默认配置」 不需要重新发明,直接抄
如果你只是想看懂 GitHub 上那些 modeling_xxx.py
这一份足够你不再被术语吓住
Anthropic pays $750,000+ a year for engineers who can build LLM architectures from scratch. Stanford taught the entire thing in 1 hour lecture & released it for free.
Bookmark & watch this today before someone takes it down.
Readers added context they thought people might want to know
The video is Lecture 3 on architectures and hyperparameters from Stanford's 19-lecture CS336 course "Language Modeling from Scratch," which covers the full process including data, training, and evaluation; one lecture does not teach building LLMs from scratch. youtube.com/watch?v=ptFiH_… stanford-cs336.github.io/spring2025/
Harsh Seth retweeted
Interesting post on Hacker News sharing tips, insights, and practical code on how to better fine-tune models like Llama 2.
The best part is that they have released a nice set of notebooks.
news.ycombinator.com/item?id…
Harsh Seth retweeted
The deeper I go into LLM use cases, the more time goes into prompt optimization.
At first, it was possible to simply document and manage a spreadsheet of prompts and generations but as you scale experiments, compare different models, document parameters, develop different prompt versions, and track other metadata, this approach becomes unsustainable.
Even when I am fine-tuning LLMs, which also requires a lot of tuning on data, I need to have a clear picture of what prompts are working and which ones aren't based on different settings.
As LLM experiments scale and use cases become more complex, the more you will rely on tools to track and debug prompts. You will need a solution that logs prompts, allows you to store metadata, and easily visualizes, navigates, and search prompt results.
I have been really impressed with the new prompting tools by @Cometml. I already use their tools for tracking and managing my fine-tuned LLMs, so it's cool to see that they also enable LLMOps and prompting tools to easily track and debug prompts at scale.
It may not be obvious right now but as we continue to improve and build on top of LLMs, you will require tools like this to manage your experiments. There is a huge good opportunity to upskill in this area of LLMOps as not too many have this skillset. It's also a good time to learn about this topic.
If you are interested, here is a nice detailed Colab notebook showcasing how to log different types of prompts like few-shot and prompt chains using Comet: colab.research.google.com/gi…
Harsh Seth retweeted
If you’re a starting entrepreneur, your first business should cash flow.
Then you can go big.
A particular type of small business is perfect for this:
“Nice Life Businesses.”
Here are my favorite 7 examples of them:
Harsh Seth retweeted
Now that ChatGPT has rolled out custom instructions to most users, try out this instruction -- it makes GPT 4 far more accurate for me: (Concat the rest of this 🧵 together and put in your custom instruction section)
Harsh Seth retweeted
How much does it ACTUALLY cost to start & scale a $20k/mo business?
A Reddit user broke down how he did it with "inventrepreneurship."
Here’s how it went down:
Harsh Seth retweeted
You can now connect Jupyter with LLMs!
It provides an AI chat-based assistant within the Jupyter environment that allows you to generate code, summarize content, create comments, fix errors, etc.
You can even generate entire notebooks using text prompts!
You can also pass it documentation which it can use as a knowledge source to answer questions. It uses RAG it seems.
There is also %%ai magic commands available to connect to different LLM within the cells and pass prompts that generate code directly in the notebooks.
Seems like LangChain is used to power these new AI-powered capabilities. Supported models include LLMs from OpenAI, AI21, Anthropic, Cohere, and more. Local models are also possible to use.
This is really cool stuff!
(blog announcement in the replies)
Harsh Seth retweeted
Knowing how to analyze a balance sheet is a must to make good investment decisions.
You want to invest in companies that are in good financial shape.
Learn everything you need to know about a balance sheet this article.:
Harsh Seth retweeted
🏆 This portfolio massively outperformed the S&P500
It probably will continue to outperform over the next decade too
Here are 15 Quality stocks you've never heard of:
Harsh Seth retweeted
From word vectors to Reinforcement Learning from Human Feedback...
Stanford's "Natural Language Processing with Deep Learning" course is one of the most relevant and best AI/ML courses today.
It's just amazing how much knowledge and content this course pushes out every year.
It's hands down one of my go-to resources to catch up on everything to do with NLP every year.
It contains notes, suggested readings, slides, tips, exercises, and so on.
I remember watching the 2017 lectures and just falling in love with the course. I have studied the course material since then and it has tremendously helped me to keep up with things on NLP.
It can feel like an advanced course for people that are just beginning but it's still an exceptional reference to keep up with research and topics.
(links in the replies)
Harsh Seth retweeted
Passive income is one of the best ways to build wealth.
ChatGPT allows anyone to start building passive income fast.
Here are 9 ways to start building passive income with AI 👇
Harsh Seth retweeted
🚨 BREAKING: Code Interpreter is FINALLY rolling out to all ChatGPT Plus users.
It's the most powerful feature OpenAI has released since GPT-4. It makes everyone a data analyst.
Here are 15 mind-blowing use cases of Code Interpreter:






























