It’s quite the interesting experience to come here and see Zai in one corner, Unigroup building in another corner, the school for automation (using unreal new technology), Galbot robot booth, Infinigence, Vidu AI, Robotera and many other AI and semi startups.

Pay attention to the tips they have about the transitioning toward Ascend from Nvidia chips by researchers. Also, check the part about researchers in China finally finding Kimi K3 to be good enough for vibe coding compared to GPT and Claude. There are still some concerns about inference sped for Kimi. I think they need to get more compute.

Things have changed a little bit since the podcast, but CXMT remains very important, K3 remains the top Chinese model and lithography is still a big story.

I did some testing today on VS Code using cline plugin. DSv4-Flash is essentially free, so I’m going to keep it as a backup to GLM & Kimi’s latest models, since those coding plans are going to get used up quickly.

Going forward, I see 2 paths for open src/weight models:

1. Smaller models that can run locally with some souped up hardware.

2. Larger models that can need SuperNode to run and can compete with closed source flagship models.

All the developers are going to want to run these smaller LLMs locally. You don’t have to worry about AI company getting your data and can fine tune them anyway you want. This is going to drive an acceleration in “souped up” local hardware. Very bullish in my opinion for memory chips.

For the larger models, I think there really isn’t a lot of room for middle ground. You need to make them larger to compete fully with the Anthropic and OpenAI models.

So the first thing I did here is Kimi K3 and GLM-5.2 and what this tells us about China’s AI development. I think the overall story I want to put out there is the self improvement part of it and also models likely expanding over time and improved training data and flywheel with the Chinese labs.

By this point next year, we could very well (imo) see 5T+ parameter models coming out of China.

I just want again to point out something that I think is very under reported out there. Chinese models are getting larger and they will keep getting larger. Biren Tech reported this week that Chinese models will get to 5T+ range next year and it looks like we will reach 10T the year after as larger super node becomes available that can pool together all the high bandwidth memory and do efficient inference of these large MoE models. All of the major Chinese AI players like Moonshot, Zai, DeepSeek, Qwen and Minimax are going into new and larger sized models this year. Larger models require more compute to train, more data and more longer term optimization for inference. All of this has been made easier by the increased automation these firms can use to improve and speed up their own model development and training time. We are now seeing these very fast iterative improvement from Kimi and GLM in recent times. These are the new standards. The public is no longer happy with a new big improvement every 3 to 4 months. This needs to be every 1 to 2 months and eventually every month by the end of this year. Teams will need to have both flagship LLM and smaller and cheaper distilled models.

Finally, I did Q&A from the audience. A lot of AI related questions, especially with AGI stuff.

I’ve had very good experience with GLM-5.2 so far and will probably move up a level in coding plan if their next release makes a similar level of jump. I’ve also had good experience with Kimi 2.6 & 2.7. Although in terms of Software engineering, GLM-5.2 is clearly better from my personal experience.

I would just recommend everyone to get off Opus and Claude Code. Try the cheaper GLM through ZCode or Kimi through Kimi Code. Or you can get both. Sometimes, one model would be better and other times, the other one will be better.

We talked about everything for quite a while. I was really interesting in how someone with multipolar view from Europe would look at the Russian/Ukraine conflict as well as the Iran conflict. I think you will find that we had a lot of common view point. It seems like European leaders are waiting for Trump to fall apart and lose elections so that the democrats can get back in power and fix everything. I’m not sure that’s a good approach, but there is definitely the European pitch of “If you continue to support us with Ukraine, we will help you squeeze China”.

As two people that think this world has profoundly changed to multipolar, we both think this is not an approach that works well with the current power dynamics globally. In fact, a dramatically weakened US from Iran conflict would not only hurt Trump, but hurt the position of the collective West.

And we also finished up agreeing on something I’ve personally been thinking about. What happens to UK as it goes on this decline? Do we have Scottish independence or a unification of Ireland. These are actually real problems that few people seem to be talking about publicly.

I remember watching some of the earlier content by Will and just seeing his driving experience on various NIO and Li Auto models. This was before Li Auto was really known by anyone outside of China. You could see back then just how good the L1 and L9 were and why people in China were buying them instead of BMW.

So, I managed to get hold of Will this week and bring him on the show to talk about how he moved to China, got into doing videos of Chinese EVs, his experience with NIO, Li Auto and experiences with China’s EV sector as a whole. We ended the show talking about his experiences from the recent Beijing Auto Show and also what vehicles Will would buy. It was a really great chat with Will and I would love to have him back on the show in the future. Please follow him on his Youtube channel and X/Twitter.

I was pleasantly surprised that he mentioned Tuya as the under the radar company that he thought very highly of. I actually only found out about them a few months ago as I was getting really into embodied AI supply chain in China.

So, I recommend everyone to check it out and let me know what you think. And I recommend everyone to visit Shenzhen.

I didn’t break ground on anything new, but I did voice my opinion regarding China’s housing market slowdown at the end and how I think people are completely misreading the story.

Also since was right before Trump’s visit, we actually did talk about Thucydides Trap but not in those war exactly. How do you resolve issues between reigning hegemon and upcoming power? I think these are all hard questions that need to be resolved.

Please feel free to ask any questions you may have regarding to this interview. I’m glad to answer them.

I think it was great to hear from him on just how this China/USA tech competition started and his experience through all of this as someone who first spent time in China in the late 80s! Ever since, he has always taken a very fair and thoughtful role in being vocal about the importance of China/US ties in the media.

But I guess the focus of our talk today was on how the October surprise sanctions came about and the rationales behind these decisions. And I think it’s important for people to remember that while Biden started all these sanctions, the brunt of Chinese retaliations did not hit until 2025 after Trump started his tariff war. The reason is that export control regimes take time to develop. While America is the master of this art, China hasn’t been that involved in tech export controls (especially on supply chain side & extra-territorial) until the wake-up call from Biden sanctions.

The vast majority of the DC crowd simply did not ever think there would be consequences to their actions, so there was apparently very little cost-benefit analysis when these decisions were made in the first place.

We didn’t really get into the semi industry part of this which will be fun to talk about next time. We did talk a lot about critical minerals and Paul has actually worked with American customers of Chinese Rare Earth magnet customers as well as Indium. We talked about why Coherent visited China with Trump this time.

We also talked about why the Chinese Rare Earth retaliations only came around when the trade war started. Simply put, China just had not developed the same type of export control back in 2022. It only ramped up building these export control policies after it suffered from them. So, I think Trump was just unlucky that he was the one that got hit hard by them. Those controls were actually in retaliations for 3 years of US export controls prior to Trump admin.

I get the feeling of just incredible hard work that Zhipu is going through as part of their effort to produce the best models and products out there. You can see that with GLM-5.1, which I can attest to as being very good for programming. It can do a lot of thinking of solve problems. I get the sense that GLM models are now able to do some form of RSI (Recursive Self Improvement) where the models help improve or create new more powerful models.

One thing that stand out to me about GLM models is their size. They are in the 750B param range which can still fit as Quantized models and run locally on boxes with really large VRAM. However, they are also large enough that they are competitive with larger models. I think there are probably some sacrifices that you have to make in a 750B param model vs 1.6T or 3T param models.

It will be very interesting to see where Z.ai goes from here since they have now received a lot of funding since IPO which gives them a larger war chest to do the work that they need to do.

I would recommend everyone to read and listen to Kyle’s substack. I’m sure I will be doing more stuff with Kyle in the future. There are a lot of interesting topics and trend that we are both seeing. AI is just the start and we can probably discuss EVs, renewables, batteries and other industries and supply chains in the future. You can see Kyle’s podcast of this episode here.

If you would like to know how Chinese entrepreneurs can start new AI robotics/embodied AI product, this would be a good episode to listen to. We also talked about World Model vs VLA, educational institutions involved in AI industry and some of the projects that we have seen recently like the Honor and Xiaomi humanoid robots.

I talked about the difference between reinforcement learning in world model vs LLMs. How can you have recursive self improvement in real world models?

And finally, we also talked about the recent Sugon rack innovation. How Chinese players leveraged domestic supply chain to create a high voltage rack that allows superior cooling to existing solutions.

We tried to look at it from the view of how the conflict started, how it’s progressing and where it is going. We covered what Iran is likely getting from Russia & China in the form of intelligence and supply chain. We looked at the possible land invasion plans that’s been rumored.

Since Snek is very familiar with both this conflict and the Russia/Ukraine conflict, we also talked about the parallels and similarities we saw in the two conflict. Finally, we went into the geopolitics a little bit based on what we think will play out in the region.

The CRC Substack

Gao Wen: Xi's vision for Chinese AI's development

Read more

6 months ago · 7 likes · ChinaResearchCollective

It is very good to have him on the show, because he is in Tsinghua University right now and have access to many of the people involved in China’s LLM and embodied AI industry. Of course, I have my own experience with China’s embodied AI industry which I’ve talked about before. So, I think this is a good thing to explore as I discussed Galbot investment, Xiaomi’s usage of robots in its factories, the entire supply chain, D-Robotics AI chips and services like Tuya (which help customers find suitable model, hardware module & ODM). It is amazing to have access to this supply chain, which allows prototyping in just 4-6 weeks. That’s something not possible if you are working on these projects in America.