Post

@DeepLcom

More from @ruima

Rui Ma

@ruima

Jan 21

Almost over jetlag and finally finished my takeaways from my trip last week on new energy and robotics in China. The highlight was spending two days in Liyang, a city of 800,000 people about an hour from Shanghai that most people have never heard of. It's China's battery capital: hosts over 100 power battery companies including CATL's massive Jiangsu base, generates $14B in output and ~5% of global battery production.

We visited 6 manufacturing facilities across the energy stack: power batteries, solid-state batteries, solar, transformers, EVs. A private investor wanted their group to understand how China actually manufactures at scale, not just sit in Beijing boardrooms. So that's what we did: spent most of our time on production lines. Here's what stood out.

1. China’s Manufacturing Advantage Is a Deliberately Designed System

The biggest thing wasn't any single factory or company. It was the system that made them all possible. China's energy manufacturing is organized around place: county-level cities and the industrial parks that anchor them.

Liyang chose over a decade ago to transition from a tourist town to an industrial center. They made an early bet on batteries and secured CATL's expansion plant. But the deal wasn't just that CATL would build there. It was that Liyang would help relocate all their suppliers too. That's how you get entire supply chains within a few hundred kilometers.

Industrial parks aren't just clusters of factories. They're deliberately engineered environments: land use, utilities, logistics, housing, connections to research institutes and vocational training. Local governments compete on execution speed. Liyang's pushing a "1220" program: 1 day for business registration, 2 days for real estate transactions, 20 days for construction permits. Sometimes companies start building under conditional approvals while paperwork processes.

2. Within the System, Speed Becomes a Compounding Advantage

Once you see the system, speed becomes impossible to ignore. It's not a cultural thing. It's deliberate engineering that compounds at every layer.

One steel-intensive manufacturer sources so heavily from a single global supplier that the supplier bought and operates a dedicated ship just for their deliveries. The manufacturer built its own port to ship thousand-ton finished products directly to Shanghai. That's an advantage you can't get anywhere else in China, let alone globally.

Every company we met moved beyond standard vertical integration. They build their own manufacturing equipment, not everything, but surgically. They identify steps where off-the-shelf machinery is too slow or poorly tailored, then design custom solutions. Speed alone determined project winners more than once. Global competitors matched or beat on quality but delivered in years. Chinese suppliers delivered in months.

Read 6 tweets

Rui Ma

@ruima

Jan 11

Before DeepSeek came out of left field, many people in China would have expected Zhipu AI (智谱) to lead the country’s AI foundation model efforts. Zhipu started early, grew out of a strong academic lineage, and focused consistently on base model research rather than applications. (The company is closely associated with Tsinghua University and the Beijing Academy of Artificial Intelligence, and has been an active contributor to China’s open model ecosystem through its GLM series.)

Zhipu is also the first independent foundation model company in China, one of the “AI Tigers,” to go public. It listed on the Hong Kong Stock Exchange on January 8, 2026, with a market capitalization of over USD $7 billion.

Tang Jie (唐杰) is a professor in the Computer Science Department at Tsinghua University and a founding initiator and chief scientist of Zhipu. He is widely respected within China’s AI research community. Recently, he circulated an internal letter to employees and, around Christmas, shared a set of views on the future direction of AI.

Reading his writing, I sense growing confidence that AI will increasingly replace human labor and that progress will continue at a relentless pace. I agree with many of his points, though even as someone who is very much a techno-optimist, the speed of this trajectory is quite unsettling and I'm still not sure how to feel about his very direct assertion that the "whole purpose of AGI is to replace human labor."

In his letter to Zhipu employees on the day of the IPO, he wrote about the 2026 roadmap that emphasizing a greater focus on long-horizon task execution, with agents, reinforcement learning, and continual learning increasingly treated as core model capabilities rather than application-layer add-ons.

1/ GLM-5: An upcoming next-generation model built through further scaling and new technical improvements, with a focus on completing real, end-to-end tasks rather than demos.

2/ Post-Transformer architectures: Transformers are hitting limits in ultra-long context cost, memory, and model update mechanisms. Zhipu plans to explore new architectures, new scaling laws, and chip–algorithm co-design to improve efficiency.

3/ More general reinforcement learning: Current RLVR works for math and code but relies on manually constructed, verifiable environments. The goal is more general RL that can handle long-horizon tasks spanning hours or days.

4/ Continuous / online learning: Today’s models are static after deployment and become outdated. Enabling continual learning and autonomous evolution is seen as the hardest but most important next paradigm.

5/ Agents and domain models converge into base models: As base model capability improves, agents and domain-specific models should increasingly be absorbed into the foundation model rather than built as separate systems.

6/ X-Lab: A new internal group focused on high-risk, frontier research, including new model architectures, new learning or cognitive paradigms, and incubation of new projects across software and hardware, with an explicit mandate for disruptive rather than incremental work.

And here are his 8 thoughts on AI from Christmastime, where he suggests that AI progress will be relentless, with agents, memory, and continual learning as the decisive bottlenecks. One statement stuck out in particular:

"The first principle of AI model applications should not be to create new apps. Their essence is AGI replacing human labor, so the core of application development is building AI that can replace different types of jobs."

I feel like people may not be comfortable with that assessment.

Anyway, I've translated his thoughts below with some light editing for readability.

Read 11 tweets

Rui Ma

@ruima

Dec 16, 2025

WHAT CHINESE SCIENTISTS THINK THE WORLD COULD LOOK LIKE BY 2049

On December 6, the Tengchong Scientists Forum convened in Yunnan, where a group of leading Chinese scientists and research institutions released a report titled Tech Foresight and Future Visions 2049. The report outlines ten technology visions and corresponding future life scenarios, spanning AI, robotics, energy, health, computing, materials, transportation, and human habitation.

What stands out is not just the scope, but the framing. Advanced intelligence, including ASI, is treated directly and largely positively, as part of a broader project to expand human capability and improve quality of life over the long run. Rather than focusing on failure modes or existential risk, the report consistently emphasizes uplift, coordination, and human-centered outcomes.

The report reflects over a year of structured discussion among scientists and researchers affiliated with major research and corporate institutes, including Huawei, China Mobile, Tencent Research Institute, Shanghai AI Lab, and SPIC’s innovation center. What follows is a condensed summary of the ten visions, with their key milestones.

Vision 1: From AGI to ASI

AI evolves from task-specific agents toward AGI and eventually ASI, with the explicit goal of amplifying human cognition, creativity, and complex decision-making. Self-evolution is treated as the core requirement, enabled by embodied intelligence grounded in the physical world.

Trajectory
L1/L2 (today): Vertical AI agents drive productivity gains
L3 (~2030): Early AGI embedded in core industry workflows
L4 (~2035): Partial self-evolution reshapes labor and social organization
L5 (2049): ASI enables paradigm-level knowledge breakthroughs

Vision 2: General-purpose robots at scale

Robots transition from research tools to everyday assistants and long-term companions across homes and industries. Progress depends on advances in dexterous manipulation, tactile sensing, and data flywheels powered by real-world deployment.

Trajectory
L1/L2 (today): Research-led deployments, limited industrial pilots
L2/L3 (~2030): World models and VTLA systems enable scaling
L3/L4 (~2035): Around 30% of households consider robot purchases
L5 (2049): Consumer-level pricing and mass adoption

Read 12 tweets

Rui Ma

@ruima

Aug 11, 2025

🧵Takeaways from WAIC and Company Visits in Shanghai

The @TechBuzzChina AI Trip began as a personal mission to get a clearer view of what is really happening in China’s AI sector. When I mentioned it to a few other investors, the interest was immediate, and the group filled up so quickly that we now have enough demand for a second edition next month, which will focus more broadly on deep tech. It was my first time at the World AI Conference (WAIC), and also the first time in its eight-year history that the event had sold out. Tens of thousands of people showed up, the halls were packed, and for a few days it felt like WAIC had taken over the city. Friends in Shanghai joked that it was the only thing happening that week.

We spent about 5 hours walking the conference floor but most of the remainder of the 5 days in smaller, more in-depth meetings hosted by companies, partners, and VCs. We met with companies ranging from Baidu, Alibaba, and Tencent (BAT) to unicorns, as well as young startups founded less than a year ago. Almost all of these companies sell to enterprises, though a few were consumer-facing. Here are some of the clearest themes that stood out from the week:

1/ Robotics is a major highlight. It’s clear that in China, AI and robotics are seen as part of the same conversation. The embodied intelligence approach is not niche but central. This trip didn’t allow for a deep dive into robotics supply chains, but we will do so in the future, and cover the entire robotics ecosystem in more detail, including the growing number of component manufacturers that are both interesting from a technology perspective and increasingly available for public market investment due to growing government support.

Some of the technology seemed very impressive. One company we met was an electronic skin maker claiming 80% domestic market share for humanoid robots — still a small market, but their sensors can detect as little as one gram of weight, which is critical for making truly useful humanoids and has applications in robotic arms as well. That said, multiple VCs commented on rising valuations in robotics, calling them irrational and inflated compared to just a year or two ago.

2/ Enterprise AI is strong, but monetization is hard. Many companies we met serve enterprise customers and have technically advanced offerings. The challenge is not capability but revenue: willingness to pay in the domestic market is low, which pushes some founders to consider overseas markets early — especially Japan. This time we also met several organizations that actively facilitate such expansions. Monetization challenges can also be structural. One founder described a problem where their product was so effective that Chinese clients, being cheap, decided that they would only order every few years, making renewals difficult and forcing a rethink of the business model.

Read 11 tweets

Rui Ma

@ruima

Dec 10, 2024

One of the hottest trends in Chinese consumer internet this year …

Revenue from short video dramas 短剧 — serialized stories of 1-2 minute episodes typically totaling 60-120 minutes, designed for vertical viewing on phones—has surpassed movie box office.

Operators compare them to “mini games” rather than traditional long-form media.

Creators attribute their addictiveness to the fact that they cram “2-3 movies’ worth of drama into one series” and the “3-7-21 rule,” where every 21 seconds must evoke a new emotion to maintain engagement. I forgot what the 3 & 7 stand for though 😂

Hongguo app from ByteDance has broken out to be a category leader with many predicting it to be the next 100mm DAU app. (It was at 30mm in October and many think it’s about 50mm now). It was also the fastest growing app this year in China.

Exact breakdowns are hard to find, but it appears the 45+ age group has the highest ARPU in this category. Surprisingly, surveys, including one from JD, reveal a counterintuitive trend: younger people (under 35) are highest in health-related spending, while older adults are increasingly driving online entertainment spending.

Read 7 tweets

Rui Ma

@ruima

Nov 5, 2023

Fascinating thoughts from a Chinese-American founder who’s working on AI in global markets re: disruptive power of this new tech:

“996, labor-intense, low salary companies might be a concept of the past, when you pay only $50,000 for an engineer …

… which is very decent a price tag in most part of the world, you are not going to spend another $10,000 for training, tools, etc.”

His point is that in the US where labor is so much more expensive it’s much easier to justify buying AI training and tools for the workforce

And that these will yield much more interesting, unpredictable and scalable benefits over time.

“when i presented thses AI new trends in HR, finance, etc … they don't give a sh*t, especially in Singapore”

To be fair, it’s hard to find people who give a sh*t anywhere …

Read 4 tweets

Share this page!

Enter URL or ID to Unroll

Rui Ma

Try unrolling a thread yourself!

More from @ruima

Rui Ma

Rui Ma

Rui Ma

Rui Ma

Rui Ma

Rui Ma

Did Thread Reader help you today?

Don't want to be a Premium member but still want to support us?

Send Email!