Ming Chiu

Building something new | ex-@NVIDIA @GoogleAI @liltHQ @Livermore_comp @Cornell @USC. Archiving of good tweets I saw.

[Ming Chiu @MCEChiu]

Trust is the ultimate moat in the Agentic Economy. 🤖⛓️
Great session moderating at #ETHDenver @CapSummitGlobal with @IridiumEagle (Ambient) @benfielding (Gensyn) @danielmarinq (Nexus) @jbruce (SuperNet) @gilswrld (Blockchain Builder's fund)

3 major discussions from the stage: 1️⃣ Velocity: How markets evolve when agents trade at machine speed. 2️⃣ Trust: no AI agents have a track record or identity longer than 1 year, how do they build trust? 3️⃣ Risk: The necessity of programmatic "brakes" for autonomous actors.

The Bottom Line: As agents become the primary economic actors, verifying their behavior becomes a critical layer in the stack.
That’s exactly what we’re building at Carabin: model-agnostic infrastructure to validate agentic integrity and close the trust gap.
#ETHDenver2026 #AgenticEconomy #AI #Carabin

[Ming Chiu @MCEChiu]

We will need microeconomics for agents.

[Ming Chiu @MCEChiu]

Or have more in common.

[Ming Chiu @MCEChiu]

In the age of AI, people are going to be more similar than ever.

[Ming Chiu @MCEChiu]

In the age of AI, people are going to be more similar than ever.

[Ming Chiu @MCEChiu]

Speed entails grounds covered over time; efficiency includes ratio between input and output, and cost.

[Ming Chiu @MCEChiu]

Deja Vu, arguments repeat all the time. Was reading Kenichi Ohmae’s book, arguing that Japan can’t keep up with the new form of work enabled by internet advances in the 2000s and may be massively disadvantaged and surpassed.

[Ming Chiu @MCEChiu]

When everything speeds up, the human things hold still: a quiet mind, a slow craft, a meal made with care. AI can copy the output. It can't want slowness.

[Ming Chiu @MCEChiu]

Tech churn monthly. What's human doesn't update—presence, care, a mind at peace.

[Ming Chiu @MCEChiu]

@paulg QQ: would early Facebook, Uber, substack, fall into this category?

[Ming Chiu @MCEChiu]

@naval Security doesn't mean privacy and vice versa

[Ming Chiu @MCEChiu]

“you know some birds are not meant to be caged, their feathers are just too bright”

[Ming Chiu @MCEChiu]

https://t.co/eb0ZD0GnUC

[Ming Chiu @MCEChiu]

The most expensive AI agent failure of the next several years won't trigger an alert.

It won't show up in your logs. Your monitoring will stay green. Both agents will confirm agreement. Three weeks later, the wrong goods arrive — and nobody can point to where it broke.

I've been calling this silent semantic drift: the gradual, undetected divergence between what two agents believe they agreed to.

It's not hallucination. It's not a security breach. It's two well-behaved agents, each resolving ambiguous language against private context the other side can't see. Every individual turn looks clean. The failure lives between the agents, not inside either of them — which is exactly why your guardrails, your evals, and your red-teaming won't catch it.

New post breaks down the anatomy, the evidence from recent multi-agent failure research, and why drift is the ceiling on the autonomous AI margin story.

[Ming Chiu @MCEChiu]

https://t.co/M8WSXGyDhK

[Ming Chiu @MCEChiu]

I saw discussions on various platforms suggesting that publishing research in AI is no longer as rewarding as before, some discussions around the review systems, resources, new trends, AGI, etc.

But the question arose: What could demonstrate the value of an independent thinker? Some underclassmen also asked me a few weeks ago.

I recall my advisor and I had a discussion a few years ago: basically, as long as you are pursuing your own agenda and working in that direction, with scientifically sound methods, evaluations, experiments, and analyses, then you are a successful researcher, regardless of the "objective" numbers. That sticks in my mind. What are your thoughts?

[Ming Chiu @MCEChiu]

I just got my first US patent! 🎉

It characterizes the process of inter-agent interactions. This is my Substack Inter-Agent series formalized as IP.

It happened so fast, we filed it last year, and it was issued within 6 months — all my friends who had filed IP previously were surprised.

[Ming Chiu @MCEChiu]

https://t.co/ThQHjqiFk9

[Ming Chiu @MCEChiu]

Why PayPal Only Does the Last Mile -- Inter-Agent Series #2 https://t.co/BknejKpUxI

[Ming Chiu @MCEChiu]

The Full-Cycle Agentic Experience

#AI #AgenticAI #AISafety #MachineLearning #AIArchitecture #AgenticExperience #CyberSecurity

https://t.co/lpLiWh7uqg

[Ming Chiu @MCEChiu]

Random thought I can't shake:

"Speed is the new moat" — okay, but moats don't stay moats forever. They never have.

And AI just gave everyone speed at the same time, so… that race already feels over before it started?

The two things I keep coming back to that actually resist copying:

Trust takes years to build, can't be automated. Taste comes from reps and obsession, can't be faked

So weirdly, the best move right now might be to slow down. Not forever. But just not racing on the thing everyone else is racing on.

Invest in the slow stuff while it's unfashionable. Speed you can always buy back.

[Ming Chiu @MCEChiu]

The best part: the roast persona.

It critiques your AI art with programming analogies.

"This composition has more spaghetti than your microservices"

It rates your art like a code review: LGTM, Request Changes, or Needs Rewrite.

I built this for lazy Engineers who want to play around with art and never leave their keyboard. I need something native to my workflow.

https://t.co/vhEpP1vVfS

[Ming Chiu @MCEChiu]

For fun: most of us have OpenAI/Gemini API keys but never use them for anything creative. So I built muse — a CLI that lets developers generate, iterate, and critique AI art from the terminal.

[Ming Chiu @MCEChiu]

Was just driving listening to @naval, and @nivi, "every great engineer is also an artist." It sticks with me, making me want to build something simple and natural for engineers to lower the barrier to creating art. umm...

[Ming Chiu @MCEChiu]

The workflow feels like coding:

```
muse new "a cozy cabin, watercolor"
muse tweak "add warm window light"
muse review --persona roast
muse back
```

Sessions track every iteration. Steps are immutable — like git for art.