Junxi

CMU PhD. Building @crewargo now. We need a better way to work with AI. Ex @Meta

[Junxi @junxi_song]

In Argo, you can have Claude Code and Codex work together in a group chat

[Junxi @junxi_song]

@lzz334 Actually, Argo is like another Lobster, but simpler. You can connect with them via IM and use it to assign tasks.

[Junxi @junxi_song]

Argo is a workspace where humans and multiple AI agents can work together in the same conversation. You can use groups to manage your different projects and tasks!

[Junxi @junxi_song]

@JvNixon Congrats!

[Junxi @junxi_song]

@passer_81 hey you can join the waitlist and dm me your email. I will send you a pro subscription

[Junxi @junxi_song]

We're offering one month of Pro subscription which can access to claude/codex to our first 10 early users!

If you're interested, like and retweet this post — the first 10 people will receive both a beta invite code and Pro access.

[Junxi @junxi_song]

https://t.co/nT8szfh3rH

[Junxi @junxi_song]

https://t.co/IgV6rJtqLQ

[Junxi @junxi_song]

@chunxiangai 我们目前的解决方案是通过群聊的方式组织不同项目。所有的事情只需要和你的主 agent 交代（通过手机或者语音）让主 agent 来管理，派发任务
https://t.co/9de2j48wS6

[Junxi @junxi_song]

https://t.co/JMjl4uOsgC

[Junxi @junxi_song]

Real examples running in production:

• SEO loop: checks ranking monthly, makes changes, pushes until page one
• Ads loop: tests creative, checks profitability, kills losers, runs until profitable
• Eval loop: adjusts prompts and swaps models until accuracy passes 90%

These aren't 30-minute automations. They run for a year.

The whole thing hinges on one thing: a metric that comes back black and white.

You're hiring an agency that never sleeps and gets paid in tokens.

[Junxi @junxi_song]

Loop engineering for building products? Covered to death.

Loop engineering for *running the business*? Almost nobody talks about it.

A loop = goal + self-check + permission to keep trying.

Build. Verify. Repeat. Stop when the condition is met.

[Junxi @junxi_song]

https://t.co/eUDovODAzZ

[Junxi @junxi_song]

People say: everything is impermanent, so this difficulty will pass. And that is true, and sometimes genuinely helpful.

But impermanence also means: this moment of clarity will pass. This feeling of love will pass. This afternoon, which is ordinary and good, will pass.

“The flower does not grieve that it will fall. Only we, who watch it, grieve.”

To sit with impermanence fully is to feel both the sorrow and the cleanness of it. Nothing to hold. Nowhere to be except here, where it is all occurring, all passing through.

[Junxi @junxi_song]

Anicca — impermanence — is the first of the three marks of existence in Buddhist teaching. It is also the most misunderstood.

People encounter the concept and hear: nothing lasts, so don’t get attached, suffering comes from clinging. This is not wrong, exactly, but it is incomplete. It turns impermanence into a lesson, a moral. And lessons, once learned, get filed away and forgotten.

[Junxi @junxi_song]

The deeper invitation is not to a philosophy but to a perception.

In meditation, at some point, you may notice that the sensation you labeled “knee pain” is not a single, solid thing. It flickers. It has a center and edges that shift. It is not pain — it is a rapid sequence of sensations that the mind assembles into the concept “pain” with extraordinary speed.

This is impermanence as direct experience, not as idea. It changes things.

[Junxi @junxi_song]

Flash-decoding was the best part.

Online softmax + split-KV reaches ~80% of peak at 16K context — a length where the naive kernel can't even launch (shared memory cap).

If you want to understand inference, stop reading about it. Write the kernels. Every op from naive to optimized, verified against fp64 CPU reference.

Next: bf16, tensor cores, proper profiling.

[Junxi @junxi_song]

Built LLM inference kernels in CUDA from scratch. The biggest lesson?

Decode is a memory game, not a FLOPS game.

GEMV hits ~92% of DRAM peak bandwidth while using ~1% of compute on an RTX 3060.

Naive → optimized softmax = 210x speedup. One thread per row = 0.2% of peak bandwidth.

The recipe: parallelism first, then coalescing.

[Junxi @junxi_song]

Kimi 3 is truly impressive when it comes to frontend development!
We’ve completed the integration of Kimo Code internally. In the future, you’ll be able to use CrewArgo to orchestrate Kimi 3 for frontend page and PowerPoint design.

[Junxi @junxi_song]

AI Epoch is newsletters topics related to AI agents, tools, automation, trending Github repos.

Huge thanks to AI Epoch for featuring Crewargo in the newsletter! We really appreciate the mention.
https://t.co/3NyXb1YwdY

[Junxi @junxi_song]

@marty_kausas hey, I built CrewArgo to make using AI models as easy as chatting with a friend! I’m sure you’ll love it.https://t.co/qz1pAuoagJ

[Junxi @junxi_song]

In Argo, you can get Claude Code and Codex to work together.

This agent team experience is amazing—it greatly improves the quality of the work!

[Junxi @junxi_song]

We’ve built all of this into Argo. You only need to tell Argo what you want to do, and it will take care of arranging everything for you.https://t.co/9de2j48wS6

[Junxi @junxi_song]

Grok 4.5 is really impressive. We’re considering building Grok support into the next version.

The new recommended development workflow will be:
1. Fable 5 generates the plan.
2. Grok 4.5 executes the work (very quickly).
3. GPT-5.6 reviews.