Building stuff / ML infrastructure at NVIDIA, Alibaba and Microsoft
San Francisco, CA
Joined September 2012
- Tweets2.3K
- Following890
- Followers1.9K
- Likes3.8K
Spent days playing github.com/Gen-Verse/OpenCla… , it is very interesting.
Basically your next message is the reward signal for the last one. "Perfect": +1. "Redo": -1. Topic switch: 0.
OPD part is interesting too, have the judge extract a hindsight hint from user feedback, feed prompt+hint through a teacher, use teacher's per-token logits as supervision. Smart.
Risks are obvious. Judge isn't trained, so reward hacking compounds. But it’s still amazing to see someone can glue the whole RL pipeline like this behind OpenClaw.
For multi-agent with role based task sharing, I have mixed feeling based on my own usage.
The issue is the task cards on that board are summaries, and summaries lose accuracy and context. One agent (or a human) writes "Build user profile endpoint with auth." Another agent picks it up and works from that string. The original intent, the implicit assumptions, the tradeoffs already considered, all compressed into one sentence.
Hermes has a shared repo as a fallback. Agents can read the actual code. Good. But the default path still leads from the card, not the source. You're relying on the tunnel vision to know when to ignore the spec and go read the raw files.
What actually works, based on what we do in production: main agent across sessions, writes its own progress.md, reads its own progress.md. Can fan out, but no handoff. The writer and the reader are the same. Information compounds instead of degrading.
Kanban is a great UI for watching multi-agent happen. It's not a fix for why it breaks.
phulin.me/blog/simtower This was a good read, when coding is cheap as today: 1. the model is not the bottleneck anymore, the verification loop is, 2. long-tail software was never demand-starved. It was supply-starved, now the expensive end just collapsed.
Training a model that's good at both math and code sounds like a data mixing problem. It's not. RL gradient updates are far more aggressive than SFT, optimizing for code actively degrades math. The domains fight each other. Kudus Nemotron-Cascade 2 finds a different bet!
🚀 Introducing Nemotron-Cascade 2 🚀
Just 3 months after Nemotron-Cascade 1, we’re releasing Nemotron-Cascade 2: an open 30B MoE with 3B active parameters, delivering best-in-class reasoning and strong agentic capabilities.
🥇 Gold Medal-level performance on IMO 2025, IOI 2025, and ICPC World Finals 2025:
• Capabilities once thought achievable only by frontier proprietary models (e.g. Gemini Deep Think) or frontier-scale open models (i.e. DeepSeek-V3.2-Speciale-671B-A37B).
• Remarkably high intelligence density with 20× fewer parameters.
🏆 Best-in-class across math, code reasoning, alignment, and instruction following:
• Outperforms the latest Qwen3.5-35B-A3B (2026-02-24) and even larger Qwen3.5-122B-A10B (2026-03-11).
🧠 Powered by Cascade RL + multi-domain on-policy distillation:
• Significantly expand Cascade RL across a much broader range of reasoning and agentic domains than Nemotron-Cascade 1, while distilling from the strongest intermediate teacher models throughout training to recover regressions and sustain gains.
🤗 Model + SFT + RL data:
👉 huggingface.co/collections/n…
📄 Technical report:
👉 research.nvidia.com/labs/nem…
Just read SWE-CI arxiv.org/abs/2603.03823, finally a benchmark measuring AI coding maintainability over time. Score weights later iterations more heavily so technical debt shows up. Though: what if Architect generates weak requirements tank the Programmer's score? Ablation Study?
This is interesting and we've been practicing on-demand image pulling for a while (cc: github.com/alibaba/accelerat…), curious if it's similar. We go a bit further in serverless offering where image pre-heat w/ in-place rolling is possible.
Just launched: image streaming for GKE cloud.google.com/blog/produc…. If you use autoscaling, have large container images, or just want your applications to start faster, you will want to use this.
Just curious can this minimal apiserver (github.com/kcp-dev/kcp) help Virtual Cluster (github.com/kubernetes-sigs/c…) by eliminating the full apiserver for my tenant? /cc @smarterclayton
Sharding etcd, load balancing apiservers, enabling bookmark and incremental node heartbeats, adding rate limiter, enforcing controller best practices ... 💪💪 treat your apiserver as pets 😉
This quoted post is unavailable.
In last week's CM@WS (Cluster Mgmt @ Web Scale) meetup, we shared a bit on building abstractions at scale, and this slide raised many discussion/feedback. So I put the slides open (sensitive data removed): docs.google.com/presentation… /cc @bgrant0607
Lei Zhang (Harry) retweeted
Today we are very excited to have 5 new TOC members! Please welcome @erinaboyd, @cdavisafc, @resouer, @dzolotusky, @ahcorporto! 👋
cncf.io/blog/2021/02/02/welc…
💪The good news today is we finally fixed most issues on the legal side to move #openyurt project to @CloudNativeFdn : github.com/cncf/toc/issues/5…, if you are not familiar with OpenYurt, it's probably the coolest #edge plugin for upstream #kubernetes! github.com/alibaba/openyurt
Lei Zhang (Harry) retweeted
#TechNotes, September 2020: No, it is the unit tests that are wrong; Cases for and against the 10x Programmer; Google and user interfaces; imperative pipelines and multiple management models; and why apps, not machines, are the center of the universe. joabj.com/Writing/Tech/Notes…
#Kubernetes is neither a PaaS nor a infrastructure under PaaS. It seems many projects/products made wrong assumptions. Kubernetes is a unique kid, I guess it already broke the "Three-Tier Model" of cloud.
Lei Zhang (Harry) retweeted
We are excited to share the #KubeCon + #CloudNativeCon NA 2020 conference transparency report!
Our second virtual event saw record-breaking registration and attendance, with 67% being first-time attendees 🤗
Find more great stats in the full report 👇
cncf.io/blog/2020/12/18/kube…
It sees AWS is keeping building its new portfolio around developers & applications, will this be the new trend of 2021?
Copilot is designed to help you deploy a production ready containerized service with only a few commands. You provide your Dockerfile, and let Copilot build, push, and launch your container on AWS.
aws.amazon.com/blogs/contain…
🤩🤩 @alibaba_cloud is proud to be part of #CloudNativeKitchen, can’t wait to see what the community is cooking on Dec 16! Come and join!
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
This experience seems super dope, not sure if it will integrate with copolit cli 🧐
@clare_liguori and team went to work. We spun up a new team in NYC under @nrh. The result is AWS Proton aws.amazon.com/blogs/aws/pre…