Small Models, Money Pits, & China (again)

TLDR;

Today, we take a look at a new rundown of AI news, focused around three themes: the growing importance of small models, how AI continues to move insane dollars, and China’s relentless catch-up on the US, as well as news on new data centers, Intel’s new savior, & more.

Enjoy!

CHINA

China responds to Google

China hasn’t taken longer than a week to respond to Google’s amazing Genie 3 model, the model that marveled the world (and me specifically) by creating interactive 3D worlds on demand.

Now, Tencent has released its own version of this idea.

→ SEE VIDEO OF THE HUNYUAN MODEL

TheWhiteBox’s takeaway:

As always, China is playing catch-up here. The impressive thing is how fast they seem to be closing the gap every time. Hunyuan is by no means as good as Genie 3, but it’s open-source and surprisingly close despite it being the first time we see this paradigm from a Chinese model (Google has released three versions of Genie).

China treats open-source as a geopolitical act, rendering the notion of Google commercializing this model a pipedream by open-sourcing direct equivalents. The reason for this is simple: China doesn’t care about profits; the US does, and they will happily constrain AI’s chances of making money as long as they need to to make the life of US Labs, in much greater need of profits and returns, harder.

Nonetheless, GPT-5 is just OpenAI’s way of saying “we need to make money,” by releasing a system where the incentives to route users to cheap models are huge, all the while paying the exact subscription cost.

It’s no surprise that they have now released ChatGPT Go in India, a $5 a month subscription. With GPT-5, they now control costs.

The truth is that Chinese Labs, while ultra-competitive amongst themselves, are not pressured to be profitable. Instead, they compete on user traction, incentivizing open-source as a means to attract users, all the while serving the CCP as a great way to make the lives of its US rivals harder and, in the meantime, increasing widespread adoption of Chinese technology all over the world.

SMALL MODELS

NVIDIA’s Fast Nemotron Model

Another day, another new small model. NVIDIA has announced a new version of their Nemotron family, which is not only strong for its size, but it’s much faster than the average model of similar size.

It shows pretty strong results on maths (AIME), PhD-level knowledge tasks (GPQA), and Coding (LiveCodeBench), with an average performance gap of around 20% compared to frontier models. For instance, the state-of-the-art in LiveCodeBench is 80%, while this model gets 71%, making it a great Pareto option: 80% of the performance of 20% of the cost (probably closer to 90/10).

But how is it so fast, even when compared to similar-sized models?

The reason the model is so fast is because it’s a hybrid model, one that combines the computation-heavy nature of the attention mechanism (the internal mechanism underpinning most models) with state-space-machine layers (usually referred…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.