Jeff Tang

Founder of LibAI, the creator of and

[Jeff Tang @JeffereyTang]

The biggest lesson we’ve learned building Ideal House 🏠:

The pretty AI room image isn't the product. It's just the hook.

The real product is the invisible engine behind it: 📦 Syncing real merchant SKUs 🔍 Previewing exact rugs, tiles, or furniture 📏 Respecting true size, material & variants 📊 Tracking what shoppers actually care about 🛒 Integrating seamlessly into PDP workflows
We aren’t building just another image generator. We’re building a catalog-to-visual-commerce workflow.
Pretty images win attention. Reliable workflows win adoption. 🚀

[Jeff Tang @JeffereyTang]

The popular post. How many of them did you follow?

[Jeff Tang @JeffereyTang]

Here it is the testing site:
https://t.co/uhN5H3btc5

[Jeff Tang @JeffereyTang]

The absolute biggest "dark horse" in the AI video space right now is HappyHorse 🐎! It just parachuted straight to the top of the video arena leaderboards, completely disrupting the space with its native audio-video capabilities. 🤯

A quick look under the hood:
🔹 15B parameter single-stream Transformer
🔹 Lightning-fast inference: 1080p in just 38s on a single H100
🔹 Native lip-sync support for 7 languages

🔥 The ultimate flex: Fully open-source & commercially permissive! We've been deep in the trenches building multi-modal video generation pipelines, and an open-source model of this caliber sends a massive shockwave. The traditional bottlenecks for producing AI short dramas—high rendering costs and strict audio-visual alignment—are practically being leveled overnight.

What’s your take on this open-source surge? Is HappyHorse the "iPhone moment" for AI filmmaking and short dramas? How do the closed-source giants defend their moats from here?
Drop your benchmark thoughts and hands-on experiences below! 👇💬
#AIVideo #HappyHorse #OpenSource #GenerativeAI #AIFilmmaking #TechTwitter

[Jeff Tang @JeffereyTang]

The rumored Manus acquisition — and reports that it may be blocked — raises some interesting questions.

If a country chooses to restrict such a deal, is it primarily about protecting strategic technologies and data?

If the deal is allowed, is it about enabling founders to access global capital and scale innovation faster?

As AI companies become more foundational, where should the line be drawn between:
national interest, open markets, and technological progress?
And more broadly —
are we entering a phase where every significant AI company is no longer just a startup, but something closer to infrastructure?

Curious how others are thinking about this.

[Jeff Tang @JeffereyTang]

will be testing its coding ability

[Jeff Tang @JeffereyTang]

DeepSeek V4 might be the most polarizing release in AI this year.

On one side:
“Open-source SOTA.”
“Blazing fast reasoning.”
“Insane cost efficiency.”
“Engineering masterpiece.”
On the other:
“Underwhelming in real use.”
“Hallucinations still a problem.”
“Hype > reality.”
So what’s actually going on?
We might be missing the bigger picture:
DeepSeek V4 isn’t trying to win benchmarks —
it’s trying to win the deployment layer.

Speed + cost + decent reasoning
perfect reasoning that no one can afford to run
That’s a different axis of competition.
But here’s the uncomfortable question:

If a model is fast, cheap, and “good enough” —
does it matter if it’s truly SOTA?

Or are we entering the era where
engineering beats intelligence?

Curious where people land on this 👇

[Jeff Tang @JeffereyTang]

@IdealH_official Home renovation approach is renovated!

[DeepSeek @deepseek_ai]

🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.

🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.
🔹 DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice.

Try it now at https://t.co/GCdiMzk1Dl via Expert Mode / Instant Mode. API is updated & available today!

📄 Tech Report: https://t.co/drlDrxkYtp
🤗 Open Weights: https://t.co/T13Y8i7SDM

1/n

[Jeff Tang @JeffereyTang]

@JamilKhabir396 HI

[Jeff Tang @JeffereyTang]

World Models are quietly becoming the most important frontier in AI.

While everyone is amazed by Seedance Pro 2.0 and GPT Image 2.0 pushing the limits of generation, the real shift is happening underneath: models that don’t just generate — they understand, simulate, and predict the world.

Text → Images → Videos
Now: Perception → Dynamics → Imagination
A true World Model doesn’t just render pixels.
It learns how the world works — physics, causality, agents, consequences.

That’s the path to AGI:
Not bigger outputs, but deeper internal models of reality.
The question isn’t “can it generate?”
It’s “does it know what will happen next?”

We’re moving from:
📸 Snapshot intelligence → 🌍 Simulation intelligence
And once machines can simulate the world…
they can reason, plan, and act within it.

That’s when things get real.

[Jeff Tang @JeffereyTang]

Current AI paradigm: human gives prompt → agent executes → human reviews → repeat. That's not symbiosis. That's a vending machine.

The Symbiosis Engine flips it:
— Agents hold context across time
— Agents propose, not just respond
— Humans and agents co-evolve

One direction leads to tools. The other leads to civilization. The next leap isn't smarter agents. It's agents that make humans smarter — and humans that make agents evolve.

Symbiosis Engine. Agent Republic. Self-evolving mesh.

This is what we are building at LibAI Lab. Any suggestions?

[Jeff Tang @JeffereyTang]

Anthropic built "Undercover Mode" to prevent internal leaks.
Then shipped the entire source in a .map file — probably generated by Claude Code itself.
Security is a process, not a feature. And Claude Code is not an API wrapper. It's the real engineering bar for AI
Agents in 2026.
🔚 (18/18)

[Jeff Tang @JeffereyTang]

🧵 Anthropic accidentally leaked the ENTIRE source code of Claude Code via a single npm source map file.
512,000+ lines. 1,900 files. Everything.
Here's what's inside — and what every AI engineer should learn from it.
(1/18)

[Jeff Tang @JeffereyTang]

Key takeaways for AI builders:
1. Layered memory beats one big context
2. Agents should be skeptical of their own memory
3. Background tasks must be isolated from main reasoning
4. Tool permissions = your security first line (least privilege)
5. Build complete first, ship with feature flags second
(17/18)