Marvin Gao

Understand from the ground up. I explain AI, CS & engineering through code & experiments. Ex-IBM AI Scientist · Ex-unicorn EVP · JHU computer science alum.

[Marvin Gao @MG1DO]

WorldModel is super cool. let's build it from scratch! let's learn and present what's the magic behind WorldModel! https://t.co/c5UFgFPClS

[Marvin Gao @MG1DO]

I built an industry research system with Graph Engineering to trace Microsoft → NVIDIA → TSMC—with source evidence on every edge and time-aware queries. Here��s the architecture, working code, and what the graph can’t tell us. https://t.co/c1QljQsCeb

[Marvin Gao @MG1DO]

@StarKnight @mranti 费的是 Money 的“马”

[Claude @claudeai]

Introducing Claude Sonnet 5, our most agentic Sonnet yet.

It makes plans, uses tools like browsers and terminals, and runs autonomously at a level that just a few months ago required larger and more expensive models.

[Marvin Gao @MG1DO]

We just fully rearranged our workspace for AI-native collaborative development.

Fewer people. More leverage.
What used to take a room now fits into a system.

And yeah—those empty chairs?
They’re not a mistake.
They’re a reminder of how much has changed.

[Marvin Gao @MG1DO]

From first principles: generic AI cannot solve industrial problems reliably.

Manufacturing knowledge is either too sparse in public data — or permanently private. No model training run touches it.

That's why I built @KeploreAI — a private industrial skills library that gives AI agents the real expertise to work on the factory floor. Live with pilot customers.

[Marvin Gao @MG1DO]

We like to make concrete Agentic AI . Manufacturing — Accurate, Efficient, and Reliable. If AI is really powerful, let’s apply it to some really challenging place! We made it.

[Marvin Gao @MG1DO]

Accepted by NVIDIA Inception. Let’s help more manufacturers with Agentic AI!

[Marvin Gao @MG1DO]

I’m so proud to live in this era.

[a16z @a16z]

It's time to build.

[a16z @a16z]

It's time to build.

[Marvin Gao @MG1DO]

🌟We’ve officially closed our pre-seed round at a post-money valuation of $27M. At the same time, we’re excited to share that KeploreAI is moving to the heart of Silicon Valley — Santa Clara Square.
Surrounded by of the world’s most inspiring and influential companies, we’re stepping into an environment that pushes us to think bigger, move faster, and aim higher.
This relocation marks the beginning of a new chapter:
Building the world’s most powerful agentic computing system — empowering AI innovation and enabling real-world applications at unprecedented speed and scale.

[Marvin Gao @MG1DO]

No! The robotics is kind of different compared to previous ones, but it still faces lots of unpredictable and uncertain challenges. Most importantly, it cannot build data flywheel easily because of it’s hard to be applied to real customers with a large amount just by providing some specificity functions, especially in generalist robots, which is a hotspot recently.

[Talia Goldberg @TaliaGold]

Robotics has been VC's favorite way to lose money. ~10 cents returned for $1 invested over the past decade. Hardware is hard, they said. They were right.

At our partner offsite, I presented on Physical AI. "This time is different."

Famous last words! But converging tailwinds are rewriting the equation: powerful VLMs, edge compute, lower cost hardware, and top talent commercializing breakthrough research.

Publishing excerpts from the internal presentation w/ @AlexandraSukin @bhavikvnagda

Why this time is actually different, what we're looking for, what we're avoiding.

Founders, hit us up!

[Marvin Gao @MG1DO]

Coworkers can do much MORE than “tools”.

[Chubby♨️ @kimmonismus]

Jensen Huang on why the AI industry is significantly larger than traditional software industry. Software is “tools”, while AIs are “workers”:

[Marvin Gao @MG1DO]

Definitely, I had been thinking more than 100 names before I decided the final name is Keplore, which is easy to understand but a new word, “keep explore”. Moreover, it is mainly named after a very important scientist, Kepler, who started a new paradigm for modern science.

[Startup Archive @StartupArchive_]

Peter Thiel on why the name of your startup is predictive of success or failure

“This is a slight aesthetic thing that I believe in very strongly: the names of companies are often very predictive of future failure or success.”

PayPal and Napster are the first example Thiel gives:

“PayPal was a very friendly name — it was the friend that helps you pay. Napster was a bad name — you nap some music, you nap a kid. That sounds like a bad thing to be doing, and it’s no wonder the government then comes in and shuts the company down within a few years.”

“You want to be very careful of how you name companies,” Thiel warns founders. In the context of the sharing economy, he likes Airbnb more than Uber:

“Airbnb sounds very innocent like this virtual bread and breakfast — this very light, non-threatening sort of company. Uber sounds like a bad name from Germany sometime in the 1930s. What are you exactly above? Maybe the law? This is probably something that, again from a regulatory perspective, I think Airbnb is a vastly better name than Uber.”

And on the social networking side, Thiel likes Facebook more than MySpace:

“You can say that all these social networks involve both reading and writing… Over time, reading dominates writing. Facebook was about learning about people around you — about their real identities at Harvard. MySpace started among wannabe actors in Los Angeles, and it was about them coming up with fictional narratives around themselves and a lot of other people in LA who are generally like that. And because reading dominates writing, Facebook would ultimately dominate MySpace. There’s a certain version where the whole arc of the company and the whole product arc was implicit in the names.”

Video source: @mercatus (2015)

[MIT CSAIL @MIT_CSAIL]

Why can’t programmers tell the difference between Halloween & Christmas?

Because oct 31 = dec 25.

[a16z @a16z]

China has overtaken the US in cumulative open-source AI model downloads:

[Marvin Gao @MG1DO]

The Arrival of Home Robots Will Be Much Later Than People Think

Everyone’s waiting for the era of home robots.
But it will come much later than people imagine.
Not because AI isn’t smart enough —
but because the world they need to live in is pure chaos.

I’ve spent years researching embodied intelligence and published in top journals.
Here’s what most people — and most investors — refuse to see 👇

��

Software evolves through clear feedback loops.
You solve a tiny, concrete problem.
It fits into existing workflows.
Users respond.
You iterate.
Capital flows in.
That’s how the system compounds.

Robotics doesn’t have that privilege.

⸻

A robot that only folds laundry?
No one buys it.
But a robot that does everything?
Costs billions to build and still fails half the time.

So teams chase “general-purpose” dreams —
burning years of R&D and millions of dollars,
without ever finding a tight feedback loop.

⸻

Every home is a unique chaos system.
Different layouts, objects, pets, light, people, habits.
An infinite set of edge cases.
No standardization.
No scalable feedback.
No stable ROI.

You can’t iterate a physical product inside a million different environments.
It’s like trying to debug physics itself.

⸻

So what happens?
Huge upfront costs.
No short-term adoption.
No data, no feedback, no investors.
The loop breaks before it starts.

We overbuild prototypes for markets that don’t exist.
Then call it “the future.”

⸻

The truth:
We’re not short on intelligence.
We’re short on structure.
Software thrives on iteration;
Robotics suffocates without it.

Until we learn how to start narrow, integrate deeply,
and build inside human environments —
home robots will remain a distant horizon.

⸻

The home isn’t the next frontier.
It’s the final boss of chaos.
And that’s why the age of home robots
will arrive much later than anyone expect

[Marvin Gao @MG1DO]

Setup and configuration are the elephant in the room in AI research and engineering — time-consuming, often overlooked, but critically impactful.

[Keplore AI Inc. @KeploreAI]

How much time do you spend setting up and testing the environment for your AI/ML projects?

As a team who has been actively participating in AI field for more than 15 years, we are developing a platform to eliminate manual environment setup, resolve conflicts automatically, and significantly reduce the time spent on research development.

However, we want to ensure that we address the real challenges faced by the AI/ML research and engineering community.

We are currently seeking input from advanced AI/ML researchers and engineers to better understand their concrete pain points. You are welcome to share your experiences in the comments.

For each comment, we will personally engage with you to better understand your specific research needs and collaborate on proposing a scalable solution that could be used for your resource optimization.

#AIResearch #MLCommunity #EnvironmentSetup #ResearchChallenges #LLM #LargeModel #AIPaper #CUDA #GPU #DeepSeek #LLAMA #AAAI #NIPS #ICLR #CVPR