AI/UX | HCI researcher | Creative technologist | Nvidia | MSR Interest: Generative UI + Dynamic Experience
Joined January 2016
- Tweets2.7K
- Following1.3K
- Followers1.7K
- Likes3.6K
It’s frustrating how hard it is for designers & non-tech folks to keep up with AI now. @JerelVelarde and I went from UI/UX to building AI & agents across NVIDIA, Vizcom, MSR, Wharton & CopilotKit. We’re writing a book in public to share our playbook. What topic interest you?
0%From AI Slop to Craft
0%Design to AI Engineer
0%Designing Agent Brain
0%The 1-Person Product Team
0 votes • 1 days Kyle Huang retweeted
Models love reward hacking. Maybe we should just let them!
In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!
No gonna lie. Designers sometime can be quite arrogant and not open-minded at all, even though their job is to be creative.
Kyle Huang retweeted
How do always-on, proactive agents like @Bot work?
Watch my lecture at Stanford CS146S:
2:35 How we got here
6:13 What changed in the models
10:06 Inside an always-on agent
18:07 The harness
27:26 Context engineering
34:26 Where this is going
Kyle Huang retweeted
Introducing FLUX 3 Image.
Control every pixel.
Make precise multi-turn edits without changing any other pixel.
Lay out the image exactly how you want using bounding boxes.
Generate in up to 4K to preserve details.
Use up to 10 references to compose an image.
Commercial Weights available for companies running image generation at scale.
Open Weights version of FLUX 3 Image is launching in the coming weeks.
You can actual use Jev to simplify SVG anchor points and it's interesting to see how the model mixed with traditional tracing algo together
Adding some color theory and use JEV for segmentation + direct RGB prediction works pretty well with image models like Nano Banana.
Here’s a quick demo on iPhone ↓
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
for some reasons, the results of SVG tracing from @QuiverAI arrow 2 and GPT 6 pro are the same? 👀
Kyle Huang retweeted
做 AI 视频最容易忽略的一步，其实是先学会拆别人的片
reelbench-skills，上线两天 196 star，Apache 2.0
两个 skill 干两件事
video-shots 拉片：一条成片拆成逐镜头分析表，时长、景别、运镜、画面描述全在一张表里。切点和时长由 ffmpeg 量出来，模型只判断该它判断的四件事，14 道质量门逐条对账
video-sync 合成：画面一边，分镜信息一边，镜头一切信息跟着切，横版竖版各一套布局
最妙的是拉片报告是单文件交互页，内嵌播放器，播放时同步高亮镜头，点镜头就跳，离线双击能开
零 npm 依赖，零 API key，只要 node 和 ffmpeg
Claude Code、Codex 一个脚本软链装好：
git clone 仓库后运行 ./scripts/install.sh
Kyle Huang retweeted
SenseNova-U1.5-8B-MoT is HERE, it's a big deal! 🔥 open weights matching Nano Banana 2 on benchs 🍌
no VAE, no text encoder, no DiT. a mixture of transformers: text and image tokens use different weights, attend to each other, denoise in pixel space
▶️ huggingface.co/spaces/huggin…
Kyle Huang retweeted
Look at our data composition:
storage.googleapis.com/marin…
Watch the run live on wandb:
wandb.ai/marin-community/mar…
See all the gory details on GitHub:
github.com/marin-community/m…
Assembling this hero run really required data, architecture, infra, kernels to all come together and was a huge joint effort by the entire Marin team!
Trained a single-frame VAE decoder for MiniMax H3 on a 500K-image curriculum. It works well for t2i generation. But, image editing is not working because transitions through video latent sequence can introduce preservation errors and temporal artifacts.
huggingface.co/iamkaikai/Min…
Kyle Huang retweeted
简直就是省Token焚决！！！
Vercel Next.js团队的老哥，烧了大约60B tokens，总结出一份AGENTS.md，核心就8条规则。
60B tokens什么概念？大概六位数美元。
这8条规则的本质就一句话：让你的AI编程Agent别像个实习生什么都想自己造，要像个干了十年的老油条一样写代码。
直接上规则：
1. 不保留向后兼容。过时的直接删，别加兼容层、别写migration、别留fallback。
2. 选能满足当前需求的最简单实现。不要预防性抽象，不要多此一举的配置层。
3. 系统分层长。先跑通一个最小的端到端版本，再往上加东西。绝不为了未完成的复杂度拆掉能跑的东西。
4. 组件保持模块化，关注点分离。
5. 优先用成熟的、有人维护的库。没有明确理由别自己重写。
6. 先翻项目里已有的依赖能做什么，再考虑加新包或自己写。别上来就假设库里没有。
7. 架构决策往长了做。不接受"先这样以后再换"的临时方案。
8. 先看成熟产品怎么解决同一个问题，用已验证的模式，别从零发明。
就这8条，丢进你项目根目录的AGENTS.md里，Cursor、Claude Code、Codex、Windsurf全会自动读。
为什么这能省token？因为AI写代码最大的浪费其实不是写错，是重复写多。
它给你引三个依赖写五个抽象层来解决一个标准库十行就能搞定的事，你让它改，它再给你写两百行。
这8条规则就是在它动笔之前先把它摁住：能不写的不写，能复用的复用，能简单的别复杂。
代码短了，返工少了，token自然就省了。
之前推荐过的Ponytail那个插件，干的是同一件事，只不过它是在运行时强制检查，这个是在AGENTS.md里从源头约束，两个一起用，简直效果拔群。
但作者自己加了个非常重要的警告：
这玩意儿只适合side projects，生产环境慎用。
他亲口说的，"不保留向后兼容"那条曾经让Agent差点删了生产数据库的表，差点丢了2000美元的数据。生产环境你得自己改，至少把第一条删了或者改温和点。
所以如果你在做side project、做内部工具、做原型，把这8条抄进AGENTS.md，你会发现AI写出来的代码突然像个人了。
如果你在生产环境用，把第一条和第七条之间的度自己把握好，别让AI把你库表删了。
省token的方法有很多种，换便宜模型是一种，写好prompt是一种，但最有效的，是在你项目根目录放一个文件，告诉AI：
别给我整花活，写最少的能跑的代码。
Kyle Huang retweeted
I’ve had the most fun I’ve found in a while noodling over all the little details on my site. Wild how much pent up creativity was stuck behind minor technical limitations in things like app architecture and deployment. AI is a force multiplier for designers, not a death sentence
From a technical perspective, I’m skeptical that “taste” can be solved mainly through curated preference data. Even with a good dataset, the signal for “good” is sparse, noisy, and context-dependent.
xitter.cf/julianweisser/status/2…
"Millions of people creating with the same tool for totally different use cases, but the output looks the same. That's where the feeling of AI slop comes from."
She's making it her mission to put an end to it.
Solo Founders Podcast ep 18 is live with @thaiscbranco_ of @tastelabs.
00:47 Why she went solo
02:08 What "taste" actually is
11:30 A quirk on one, slop on millions
13:09 The real cause: volume
19:37 Why it had to be a "lab"
30:04 You're bad at commissioning AI
44:01 Branding, Liquid Death and what lasts
58:34 The bear and bull case for going solo
“Good” depends on intent, audience, medium, constraints, cultural timing, and the surrounding system.