Image Credit: NightCafe Studio: Link
Why the US LLMs Are Showing Poor Returns For Investment Compared to China
The bottleneck is real: the problem is the lack of data!
GPT 3.5 to GPT-4-Turbo was a quantum leap.
GPT-4.5: hugely enormous computing resources (as evidenced by the massive prices) but not a serious entry in the competition.
Grok 3: 200,000 GPUs in the world’s largest AI cluster - but no quantum leap.
Incremental improvements.
Why?
Get ready to see why Big Tech may have shot itself in the leg with GPU compute and Trumped itself with tariffs!
And why China is not just the winner - they are racing ahead on all fronts.
The Model Is Only As Good As the Data You Give It!
It was never about the compute - beyond a certain point.
Once the entire Internet - copyright, non-copyright, intellectual property, free use, had been consumed:
There was nothing more for the models to learn from.
Synthetic data is nowhere near as good as the human-generated data of the Internet.
Throwing more compute and scale at the problem - and you will have AGI - is false propaganda for more chips.
Yes, GPU compute is necessary.
But algorithms Trump compute. (pun intended)
The proof?
GPT 4.5 and DeepSeek R1.
Astronomical Prices, Abysmal Performance Improvement
GPT 4.5 has a price (famously astronomical) for its API of $75/1M input tokens and $150/1M output tokens.
DeepSeek, which is competitive with GPT 4.5 on nearly every benchmark (the exception being emotional intelligence) is free for use for chat.
Its API cost is an average of 82x times cheaper.
And open-source to boot!
Yes, perhaps GPT 4.5 is a precursor for GPT-5 -
But that still does not justify the astronomical price tag.
That’s not pricing: that’s a sinkhole for the money in your wallet, especially in third-world countries.
Yes, DeepSeek has been overloaded by requests in the past, but less so now.
And with the free price?
I’m willing to wait for cutting-edge performance!
This is the Leaderboard on LM Arena as of the date of writing this article:
With its pricing, GPT 4.5 should be heads and shoulders above the rest.
Instead, an MIT-licensed model (open-source and free for commercial use) is less than 30 points below it on LMSYS Arena.
And why are Google’s LLMs leading the pack?
IMHO:
The large amount of data Google has within its worldwide juggernaut enterprise (7 models in the top 20, more than any other company) is the main reason for its dominance!
Meanwhile, in China:
China has done what OpenAI set out to do.
Create LLMs moving towards AGI and make it free and open for the entire world.
The Chinese government has poured billions into AI research and funding for private companies and cutting-edge research in universities.
China was neck-and-neck with the US as far as research is concerned - in 2024.
And now that Trump has cut funding for US research and imposed tariffs that could destroy any chance the US has in the AI race -
China is poised to take the pole position in the AI and LLM world domination.
I’m eagerly waiting for DeepSeek’s R2 (as is the rest of the world, some more in fear).
China has been blocked access to expensive GPU chips, but that may have been the best thing to happen to the Chinese AI effort!
Forced to innovate, China came up with some fantastic algorithmic and efficiency innovations.
That resulted in the exact opposite of what the ban was supposed to accomplish.
China became the world leader in AI efficiency and economy for LLMs with DeepSeek R1.
Algorithms Are the Key, Not Compute
Compute matters only as long as high-quality data is available and for meeting the high demand for existing LLM products.
For creating genuine AGI, nearly all AI experts concur that better algorithms are the key.
Reinforcement Learning is perhaps the most promising candidate in this area.
However, the most recent studies and projections for AI are mind blowing.
Some ‘independent’ researchers came up with a prediction for AI in 2027.
It's a fascinating read!
They actually project AGI with autonomous agents arriving by 2026-2027!
I agree with them (except on the compute factor).
Naval Ravikant is both right and wrong about AGI (in his interview).
(I realize going against Naval Ravikant may be problematic, the guy is a freakin’ legend.)
He is right about more compute not producing AGI.
I agree with him there.
But wrong about the emergence of ASI and AGI.
AGI will self-evolve from existing autonomous AI Agents.
Human beings will not build AGI.
AI will.
Similarly, humans will not create ASI.
An AI that has reached AGI will.
If ASI and AGI were left to humans, then maybe 100 years is a realistic expectation.
With all due respect to Naval Ravikant (again, the guy is a huge, huge legend! Legend!):
AI will build itself and evolve into AGI and ASI in ways that human beings will not even understand, given enough foundation and autonomy!
And now, with the recent actions of the Supreme POTUS:
China is well-positioned to be the country to reach AGI (and thus ASI) first.
ChatGPT-inspired Trump Tariffs and the defunding of all education grants will cost the US the AI race.
That is my certain prediction.
Conclusion
The US Big Tech companies were sold on the idea of more AI chips, more intelligence by a combination of interwoven advertising and hype.
Now, they are paying the price.
Tariffs on every country that were created with a formula from ChatGPT do not help either.
Defunding research and development in multiple sectors in the US is a disaster on all fronts.
We all know who is responsible for all this, so I will avoid naming names.
But algorithms are the key to evolving AGI.
And funding is required to police whatever AI results from the natural evolution of AI agents.
Sadly, if things continue in this direction, China is set to become the world’s dominant superpower.
The reason I say that is that AI will enable breakthroughs in multiple sectors of technology that we have never seen before.
Breakthroughs that were not even imaginable by human beings.
The AI race is the Manhattan Project of the 21st century.
And in this current scenario - China is the certain winner.
While I do not monetize my writing directly, your support helps me continue creating articles like this one without a paywall or a paid subscription.
If you want articles like this one appearing on your business and your keywords for your platforms online, you can get them (NOT on HackerNoon, but yes on other platforms)!
Contact me at:
<https://linkedin.com/in/thomascherickal ](https://linkedin.com/in/thomascherickal)
For your article! (Prices are negotiable, and I offer country-wise parity pricing.)
If you want to support my writing, consider a contribution at Patreon on this link:
https://patreon.com/c/thomascherickal/membership
Alternatively, you could buy me a coffee on this link:
https://ko-fi.com/thomascherickal
Cheers!
All Images are AI-generated by the wonderful users of NightCafe Studio, on this link: https://creator.nightcafe.studio/
No AI was used in this article except for the images.
All links are in the article itself, so there is no list of references this time around.
