Darrell Dai

Robot Research Engineer @NVIDIA | Ex-EngineAI | Ex-LimX Dynamics | , M.S. from @DukeU, B.S. from @Purdue.

[Darrell Dai @darrell_dai]

I’ve decided to leave NVIDIA. There are a few things I want to say.

As a robotics engineer at NVIDIA, I have seen Jensen as much more than the company’s CEO.

I admire his willingness to invest in a future that was not yet obvious. He committed to CUDA long before the deep learning boom and positioned NVIDIA for AI before it became everyone’s priority. Those decisions helped build the infrastructure that powers today’s AI.

I have seen that same vision in physical AI: investing in robotics, sharing open-source tools and projects, and building computing platforms that bring AI into the physical world. This is the work I do every day—and a future I believe in.

But what matters most to me is the culture he helped build.

Engineers are encouraged to explore, take risks, and learn from mistakes. Jensen reads employees’ Top 5 emails, follows their work, and engages directly. In a company this large, that connection matters. It makes an individual engineer feel that their ideas can still be heard.

That is why this has been so disappointing.

I do not have all the facts surrounding the allegations, and I cannot prejudge the investigation. My admiration for Jensen does not put him above scrutiny.

What I find difficult to accept is having to come to terms with his departure before receiving an explanation that helps me understand it. What this means for our direction, our culture, and the people doing the work remains unclear to me.

I have felt that engineers’ perspectives matter here. On this decision, I have not felt heard. That gap has been more painful than I expected.

I understand that some information cannot be shared immediately. But that does not erase the feeling of being left outside a decision that affects the company we help build every day.

A company’s greatness comes from its leadership, its governance, and the people who turn its ambitions into reality. Employees’ trust and sense of belonging are part of that, too.

Jensen’s vision and the engineering culture he shaped were important reasons I wanted to stay. His departure led me to decide that my own chapter here should also come to a close.

I leave with sadness and gratitude. The colleagues I have worked with, and the work we believed in together, will always matter to me.

These are my personal views.

[Darrell Dai @darrell_dai]

Today I tried building an auto research agent with Codex. I wanted each agent to have its own memory and skills, with a Hermes-like setup to auto-log them.

Memory is loaded on `SubagentStart`, and a background process organizes it on `SubagentStop`.

Problem: in Codex, the parent’s task instruction to the child agent is encrypted, so the cleanup process has to organize memory without knowing the task — which makes it pretty inaccurate.

A simple trick fixed it: on `SubagentStart`, use a prompt to ask what the task is.

#AI #codex #agent

[Darrell Dai @darrell_dai]

🚨 I found a Codex bug (CLI v0.142.4):
built-in Memory analyzed my conversations, but never created the files needed to remember them later. No warning appeared. Its database advanced a “success” marker. Would love a look, @OpenAIDevs. Anyone else have the same problem?🧠

[yibie @yibie]

训练小模型：2026 年最被低估的 AI 技能

2026 年 5 月 11 日，一个叫 CJ Zafir 的人发了一条推文。他想教普通人 fine-tune 开源模型。

2538 个赞，316 次转发，178,000 次观看。这条推文炸了。不是因为他发明了什么新东西——Unsloth 早在 2023 年就开源了，Hugging Face 上的 fine-tuning 教程有几百篇。炸了是因为他在一个所有 AI 公司都在比赛造「更大模型」的时候，反手指向了完全相反的方向：小模型。你自己训练的。在 Colab 上。花几十美元。

两个月过去了。CJ 没有停。

他把这条推文变成了一套完整的方法论：Codex 5.5 做大脑规划工作流，DeepSeek v4 Pro 做肌肉批量生成数据，Unsloth 做训练引擎，Qwen 3.5 4B 做基座模型。然后他用这套方法论，花了 $173 训练出一个 4B 模型，在垂直任务上精度 96%+，击败了 20 倍大的模型。 再后来他做了个 voice agent 的工具调用 fine-tune——成本 $11。

从 $80 到 $173 到 $11。这不是渐进式优化。这是在重置整个 AI 部署的经济学。

一、CJ 方法论：为什么值得认真对待

CJ 不是第一个倡导 fine-tuning 的人。但他做对了一件事：他把整个流程的成本打到了个人开发者可以承受的水平，并且公开了每一步。

工具链

| 角色 | 工具 | 成本 |
|------|------|------|
| 规划（Orchestrator） | Codex 5.5 | 订阅内 |
| 数据生成（Generator） | DeepSeek v4 Pro | API 按量 |
| 训练框架 | Unsloth | 免费开源 |
| 训练硬件 | Google Colab Pro A100 80GB | ~$0.60/hr |
| 推理部署 | llama.cpp / Ollama | 免费 |

数据流水线

CJ 最核心的洞察不是「用什么工具」，而是「怎么造数据」：

「"Low quality data = Low quality model performance"」

他的流水线是自进化的：Codex 设计工作流 → DeepSeek 批量生成数据行 → 每批次过 Quality Gates → Codex 根据上一批结果优化生成规格。结果是越跑越快、越便宜、质量越高。

$80 生成了 1 亿+参数的高质量数据集。 不是简单 paraphrase，是「手工打造」每一行——Codex 设计的多样化 prompt 模板 + DeepSeek 的高速执行 + 可编程的质量门槛。

模型选择

CJ 实测了 Gemma 4 和 Qwen 3.5 全系列后，结论很明确：

「Qwen 3.5 9B 和 4B 是垂直训练的完美基座模型。Gemma 4 在垂直任务上不及 Qwen 3.5。」

他的 Qwen 3.5 4B fine-tune 在精度和延迟上击败了 20 倍大的模型。不是什么小众 benchmark——是他自己在垂直任务上实测的结果。

二、这不是 CJ 一个人的实验——行业数据在同时发声

CJ 的方法论看起来像「个人黑客技巧」，但独立的数据正在从各个方向验证同一个趋势。

市场规模

SLM（Small Language Model，1B-13B 参数）市场 2024 年达到 65 亿美元，年复合增长率 25.7%。SLM 在大多数企业任务上已经匹配或超过 GPT-3.5 级别性能，只需一小部分成本。

性能

阿里巴巴的 Qwen 3.5 9B 在 MMLU-Pro、GPQA Diamond 和 multilingual MMMLU 三个基准上击败了 OpenAI 的 gpt-oss-120B——13 倍参数差距，小模型胜。

另一项独立的 brand normalization 案例中，一个小模型达到了 90.6% 准确率，超越了 GPT-5.2 和 Gemini 3 Pro。

企业已经在省钱——真正的省钱

| 公司 | 方案 | 效果 |
|------|------|------|
| Knowunity | 用 Distil Labs fine-tune SLM | 推理成本降 68% |
| 某客服 agent | 从通用大模型切换到 fine-tuned SLM | 月成本从 $13,000 降到 $400 |
| EliseAI | 专门训练的垂直小模型用于住房和医疗 | 推理成本降 60%，延迟降 80% |
| AT&T | 专用 SLM 做客服和欺诈检测 | 数百万美元节省 |

400 美元一个月。一个客服 agent。这是几个工程师的年收入可以覆盖无限次 API 调用的部署方案。

三、「数据集工厂」才是真正的护城河

CJ 在 5 月 22 日发了一条推文，八个词：

「"The moat might not be the fine-tune itself. It might be the dataset factory."」

这是整个小模型运动最重要的洞察。Fine-tuning 技能本身正在被民主化——Unsloth 已经能在 3GB VRAM 上训练，Colab 免费提供 GPU，CJ 连 prompt 模板都公开了。当人人都能做 fine-tune 的时候，差异化从「会不会做」转移到了「数据从哪里来」。

中文社区的 AKIRAXCLAW 在一篇分析里补了一个关键观察：

「「CJ 的工作流在 2026 年多数是业界共识，但真正难点不是工具——而是你有没有 200 笔干净数据 + 自动评估脚本。」」

在 CJ 的体系里，Codex 的核心角色是 orchestrator，不是因为它比 DeepSeek 聪明，而是因为它能设计数据流水线、构建 Quality Gates、自进化优化生成规格——它在管理数据生产的质量系统，而不仅仅是生产数据。

四、从边缘到生产线：部署已经在发生

SLM fine-tune 的部署门槛已经降到消费级硬件。Unsloth 2.0 支持 GGUF 量化导出 + llama.cpp / Ollama 部署。一个 3B 参数量化到 INT4 后占用 1.5-2GB 内存，可以在树莓派 5 上运行。Qwen 3-0.6B 在 Pixel 8 和 iPhone 15 Pro 上实现约 40 tokens/s 推理。

2024 年你可能需要一台 A100 才能跑 fine-tune。2026 年你可以在 Colab 上训练、在手机上推理。这种从「数据中心才能碰」到「个人设备上运行」的变化，是 Platform Shift 级别的。

而 NVIDIA 在推动的「Data Flywheel Blueprint」——用生产流量持续生成训练数据，自动 fine-tune，自动部署——正在把这个流程变成企业基���设施。NVInfo AI 用这个方案在三个月内把内部知识助手的质量提升到覆盖 30,000 名员工。

五、成本轨迹：从「公司预算」到「午饭钱」

CJ 的公开成本记录是最好的注脚：

| 日期 | 项目 | 成本 |
|------|------|------|
| 05-13 | 1 亿+参数数据集生成 | ~$80 |
| 05-27 | Qwen 3.5 4B 完整 fine-tune | ~$173 |
| 05-28 | Qwen 3.5 tool calling for voice agent | $11 |

$11。这在硅谷大概是一顿午饭。在印度、东南亚、非洲，是一个独立开发者完全负担得起的预算。fine-tune 已经从「大公司才有资格做的研发项目」变成了「个人开发者用午休时间能完成的 side project」。

六、结论：学会训练你的第一个模型

CJ 给的学习路线图异常简单：

1. 从 1B-4B 小模型开始，不要一上来就搞大模型
2. 用 Colab Pro（A100 ≈ $0.60/hr），不要急着买 GPU
3. 先 fine-tune 7-10 个模型积累经验（SFT → LoRA/QLoRA → GRPO/DPO）
4. 用 Codex 做规划，DeepSeek 做数据生成
5. 理解量化（GGUF）、本地推理（llama.cpp）、KV 缓存

他甚至在 5 月 23 日直接发布了一个「复制粘贴就能用」的 prompt——把整个学习路径封装成一段对话，扔进 Codex 或 ChatGPT，让 AI 带你从 beginner 到 advanced。

这就是 2026 年 AI 技能民主化的真实面貌：不是每个人都需要学会训练模型。但学会训练小模型的人，不再需要 10 人团队和 10 万美元预算。

通用大模型的军备竞赛还在继续。GPT-5、Claude Opus 5、Gemini 3 会越来越强。但越来越多的真实场景正在证明：一个花了 $173 训练的 4B 小模型，在你自己的数据上，比你花 $200/月订阅的通用大模型更好用。不是因为小模型更聪明。是因为它只做了你需要的那一件事。

参考来源：CJ Zafir X 推文系列 (2026.05.11-05.28)；AKIRAXCLAW 中文分析；Alibaba Qwen 3.5 benchmark (VentureBeat)；Unsloth 2.0 文档；NVIDIA Data Flywheel Blueprint；AT&T / EliseAI / Knowunity / https://t.co/7r0fApoHPJ 企业案例；Manicode / Arendil 独立 benchmark；SLM Market Report 2024-2026

[Darrell Dai @darrell_dai]

@ericjang11 Thanks for the inspiration you brought to this industry. Also, welcome to visit NVIDIA new office if you visit Shanghai

[Darrell Dai @darrell_dai]

After waiting over a month, I finally got my official employee badge!

Celebrating officially joining the @nvidia as a Senior Research Engineer. I feel incredibly fortunate to work alongside some of the brightest minds in robotics globally.

#nvidia #ai #Robotics #physicalai

[Darrell Dai @darrell_dai]

World First Dancing Robot in Livestream with @IShowSpeedHQ in Public, which is from the company I work for, @engineairobot

#IShowSpeed #Ishowspeedchina

[Darrell Dai @darrell_dai]

Me in @IShowSpeedHQ livestream, witnessing the first robot dancing in livestream in public

#IShowSpeed #ishowspeedchina

[Darrell Dai @darrell_dai]

Workout buddy is right here😄

[EngineAI @engineairobot]

EngineAI Robotics’ mechanical rampage strikes the sci - fi future with the beauty of real machinery!#HumanoidRobot #ArtificialIntelligence #AI #EmbodiedIntelligence #EngineAI

[Darrell Dai @darrell_dai]

@DrJimFan Hi, Jim. Glad you noticed our robot. This video was filmmed when we were doing our outdoor test.

There are many videos on Xiaohongshu and Douyin fimmed by random people from different perspectives, which I believe is the best evidence showing it's real.

[Darrell Dai @darrell_dai]

@TairanHe99 @zhengyiluo Yes. On the final version we use all harmonic.

[Darrell Dai @darrell_dai]

@zhengyiluo Absolutely. We’re discussing the part we can share.

[Darrell Dai @darrell_dai]

I like the latest demo of Optimus, the way it interacts with human is impressive. Imagine if it can walk more elegantly too, it’ll be more likely human think Optimus as a friend. @elonmusk

[Darrell Dai @darrell_dai]

After months of great efforts, we finally made our SE01 robot really walks like a normal human, not Biden.

I truly believe it can be a super model, and we’ll give it gaits of different styles. It deserves to be on the fashion show of @LouisVuitton @gucci @Hermes_Paris etc.

[EngineAI @engineairobot]

Introducing the World's Coolest Humanoid Robot — EngineAI SE01！
It features a set of self-developed harmonic joint module and combines reinforcement learning with imitation learning, achieving unprecedented excellence in natural gait through end-to-end neural network models.