Picture Loan: Nightcafe Studio: Link
Why do US LLMs show bad return for investment compared to China
Darbagaz Truth: Problem Data Lack!
GPT 3.5-GPT-4-Turbo was a quantum jump.
GPT-4.5: Very tremendous information processing sources (as proven at large prices), but not a serious entry into the competition.
Gok 3: 200,000 GPUs in the world’s largest AI set – but there is no quantum splash.
Increased improvements.
From where?
Get ready to see why Big Teh might have hit himself on his leg with the GPU calculation and drowned himself with tariffs!
And why China is not only the winner, but on all the fronts.
The model is just as good as the data you give!
It was never about calculation – beyond a certain point.
All internet – copyright, non -copy, intellectual property, free use, consumed:
There was nothing else for the models to learn.
Synthetic data is not as good as the data produced by the Internet.
More calculating and throwing a scale to the problem – and you will be the wrong propaganda for more chipset.
Yes, GPU calculation is required.
However, algorithms calculate Trump. (for cinas)
Evidence?
GPT 4.5 and Deepseek R1.
Astronomical prices, vast performance improvement
The GPT 4.5 has a price (famous astronomical) for 75 $ 75/1 million input coins and $ 150/1 million output markers.
Deepseek (emotional intelligence exception) competing with GPT 4.5 in almost every criteria Free for use for chat.
The average cost of API is 82x times cheaper.
And open source for booting!
Yes, maybe GPT 4.5 is a pioneer for GPT -5 –
However, this still does not justify the astronomical price tag.
This is not pricing: this, especially in third world countries, a pudding for the money in your wallet.
Yes, Deepseek was overly loaded with demands in the past, but now less.
And at free price?
I’m ready to wait for the last performance!
This is the leadership table in the LM Arena as of the date of writing this article:
GPT 4.5 with pricing should be head and shoulders above the others.
Instead, a MIT licensed model (Free for open source and commercial use) Less than 30 points in LMSYS Arena.
And why do Google’s LLMs manage the package?
IMHO:
It is the main reason for Google’s large amount of data dominance in Juggernaut Enterprise (7 models in the top 20, more than all other companies) worldwide!
Meanwhile in China:
China did what Openai would do.
Create LLMS that move towards AGI and make it free and open for the whole world.
The Chinese government poured billions of money for AI research and financing and the latest research in universities for private companies.
In the case of research, China and China have neck and neck. In 2024.
And now Trump has reduced the financing because the US has applied tariffs that can destroy the US’s chances in research and the US’s artificial intelligence race –
China is preparing to take the polar position under the domination of the world and the world.
I am eagerly waiting for Deepseek’s R2 (like the rest of the world, some in fear).
China has been prevented access to expensive GPU chips, but it could have been the best thing to come to China!
China, forced to innovate, found fantastic algorithmic and productivity innovations.
This is The opposite of what the ban must succeed.
China became AI efficiency for Deepseek R1 and LLMs and the world leader in the economy.
Algorithms are key, not calculation
Calculate only as long as it meets high quality data and meets the high demand for existing LLM products.
To create a real network, almost all AI experts agree that better algorithms are the key.
Reinforcement learning Perhaps he is the most promising candidate in this field.
However, the latest studies focuses on the need to be careful in AGI.
Google Deepmind Recently, it has published a projection to the future for the emergence of AGI, available in the following link:
In fact, they reflect AGI with autonomous agents coming up to 2026-2027!
I agree with them.
Naval Ravikant is both correct and wrong about AGI (in the interview).
(I understand that going against Naval Ravikant may be problematic, that the man is a strange legend.)
In order not to produce network, it is right about more calculation.
I agree with him there.
But it’s wrong about ASI and the emergence of AGI.
AGI will improve itself from existing autonomous AI agents.
People will not build network.
AI Will.
Similarly, people will not create ace.
A AI Will reaching AGI.
If it were left to people and the network, perhaps 100 years is a realistic expectation.
With all respect for Deniz Ravikant (again, the man is a big, great legend! Legend!):
AI will build itself and considering sufficient basic and autonomy, people will turn into AGI and ASA in a way that does not understand!
And now, with the latest actions of the Almighty Potus:
China is first positioned as a country to reach AGI (and therefore ace).
This is my definite guess.
Solution
The US major technology companies have been sold on more AI Cips, more intelligence ideas with the intertwined advertising and HYPE combination.
Now, they pay the price.
Tariffs do not help in every country created with a formula from Chatgpt.
Eliminating research and development in more than one sector in the USA is a disaster on all fronts.
We all know who is responsible for all this, so I will avoid calling names.
However, algorithms are the key to developing AGI.
Regardless of the natural evolution of AI Ai agents, financing is required for the police.
Unfortunately, if things continue in this direction, China will be the dominant superpower of the world.
The reason I say this is that AI will make breakthroughs in multiple technology sectors that we have never seen before.
Stayings that cannot come to mind by people.
AI race is the Manhattan project of the 21st century.
And in this existing scenario – China certain winner.
Even though I don’t make money directly from the spelling, your support helps me to continue to create articles like this without a payment wall or paid subscription.
If you want articles that appear in your business and appear on your keywords for your platforms, you can get them (not in Hackernoon, yes on other platforms)!
Contact me:
For your article! (Prices can be negotiated and I offer parity pricing throughout the country.)
If you want to support my writing, consider a contribution to Patreon in this link:
https://patreon.com/c/thomascherical/Membership
Alternatively, you can get me a coffee from this link:
https://ko-fi.com/thomascherickal
Cheers!
In this article, AI was not used except for images.
All links in the article itself, so this time there is no reference list.
