Zhipu AI - can it challenge OpenAI ?

If you're a CXO, founder or investor - follow me on LinkedIn & Twitter, or join my newsletter on my website. I share latest simplified AI research and tactical advice on building AI products.

Zhipu AI is China's most promising answer to America's OpenAI. This startup has raised $340 million from Chinese investors very recently. Investors include Alibaba, Tencent, Meituan, Qiming Venture Partners, Legend Capital and Tsinghua Holdings.

But does it have what it takes to beat OpenAI at it's game ? Let's see.

Let's use P.A.C.E.R framework for AI products to analyse their competitive power against each other.

P (Platform effect)

OpenAI - 2/10

Zhipu AI - 8/10

AI products (actually any product) are people game. More living people use it, more is the opportunity to collect golden feedback and accelerate improvement in learning.

This is where nature and how company operates play a crucial role. These two companies are fundamentally different in their mission. Zhipu AI is what OpenAI wanted to be. That is being a open platform for AI innovation.

Zhipu AI is what OpenAI wanted to be.

Instead OpenAI has transformed itself into a closed company and all of it's research and products are commercial now, nothing being free or open source.

Whereas Zhipu AI (as we can see above) has open-source products like ChatGLM (rival to ChatGPT) and CodeGeeX (coding assistant rival to Github Copilot).

Imagine ChatGLM used by millions of chinese people whereas ChatGPT and Copilot being limited to few selected organizations.

This seriously limits improvement opportunities for OpenAI.

Zhipu also has collaborative research platform Aminer where researchers collaborate with company to produce Intellectual Property.

Aminer also connects with research talent pool which gives unique leverage to company in terms to talent access as well as network effects.

A (Approach innovation)

OpenAI - 2/10

Zhipu AI - 6/10

In research, a innovative new approach can overnight change the game and dramatically boost accuracy of models with reduced time and cost.

OpenAI has been a pioneer in Reinforcement based learning approaches and combined with Human Feedback has evolved it's approach into RLHF which has given tremendous boost in accuracy as we see now in ChatGPT.

However it has not ventured much into other approaches like Graph based learning.

Zhipu on the other hand has focussed more on Graph based approaches. However, no great innovation is seen and their research is still only an incremental improvement in traditional graph neural learning approach.

Given Graph learning is superior to RL since it captures higher level structures which RL can not, Zhipu has a brighter future here.

C (Compute & Monetization)

OpenAI - 7/10

Zhipu AI - 4/10

Compute & Monetization are closely linked as you need compute power for experiments and training models, and you need money for building compute power.

VC money is limited and you can only depend so much on that. However, if a company has monetized it's models well and has great revenue then it can possibly built proprietary compute capacity. In long term (obviously) it gives leverage like nothing else.

OpenAI has partnered with Microsoft which has it's own cloud Azure so it has some advantage here. Microsoft has even built a supercomputer (one of top 5) exclusively for OpenAI.

OpenAI also has commercial APIs for it's models which are very popular and has many enterprise customers. However, this cash is still not sufficient enough compared to training costs.

Zhipu, on the other hand, will have to struggle after effects of US ban on export of AI chips to China. NVidia H100, processor of choice for OpenAI, was already banned earlier and Chinese companies had to settle for slower H800 chips. However, recently even their exports was also banned. Similar restrictions also apply to Intel and AMD chips.

This will slower progress of Zhipu AI but silver lining for them is having Alibaba (which has it's own cloud computing) and Tencent on board.

While processing power is certainly disadvantage for them, access to cloud computing power is not limited.

Also not to forget, here we have X factor of Chinese government which can provide unlimited power to Zhipu anytime. This is highly possible if Zhipu struggles, given status of Zhipu in Chinese AI ecosystem.

E (Ecosystem)

OpenAI - 4/10

Zhipu AI - 8/10

Any product which builds an ecosystem of products around it's platform automatically becomes a winner. Unity is strength is an old saying afterall.

Think Apple Appstore, which made iPhone a winner. Or Microsoft tying up with hardware partners which beat Apple initially, of course without Steve Jobs.

Zhipu has obvious advantage here due to it's free and open-source products.

Plus it has Aminer collaborative research platform and knowledge graph platform which are ecosystems in themselves.

But importantly Zhipu has taken one step ahead and built consumer applications on top of it's models, like Cognitive Digital Human platform, which takes all of it's progress in language, vision models and knowledge graphs and apply in real world applications. This has tremendous potential in industries like advertising, media and corporate operations.

OpenAI, on the other hand, is compeltely isolated. There are no collaborations except Microsoft. Even Elon Musk who provided first cheque for the company is not happy with their u-turn in mission, and has formed X.ai to achieve OpenAI's original purpose.

R (Research velocity)

OpenAI - 4/10

Zhipu AI - 7/10

Any AI company can not remain static and has to focus on progressing it's research to remain competitive. Otherwise it will tend to lose to another innovative startup with new better approach.

OpenAI is relatively slower in this aspect to Zhipu mostly because Zhipu has built a research platform. OpenAI is a company with limited researchers while Zhipu has collaboration with most of chinese researchers.

Zhipu was founded in 2019 and has reached very close in accuracy of models to OpenAI which was founded in 2015.

Being a later entrant is certainly an advantage for Zhipu but it also shows power of collaborative research.

Conclusion

Zhipu AI is a real threat to OpenAI and it's dominance in AI world.

If OpenAI does not use it's computational advantage for creating ecosytem around its products - Zhipu may dominate AI world.