Things are heating up in the AI industry since the release of Chinese model Deepseek
DDoS attacks, new western models copying Deepseek... and it's only been a week
If you like this article, please don’t hesitate to click like♡ and restack⟳ ! It’s a quick way to make sure more people will get to enjoy it too.
And don’t hesitate to subscribe to my newsletter for free to receive my new articles right in your inbox!
Just last week, a new AI model called Deepseek (which is also the name behind the company that produced it) was released on the worldwide internet.
Immediately, the AI caused a fuss. Not just because of its capabilities (it is 30-50x more effective than the leading western models in terms of energy consumption while providing results that are just as accurate), but also because Deepseek was made a Chinese company that was previously completely unnoticed in the West — but not unknown.
Mind you, it’s far from the first time the “Chinese newcomer” suddenly takes the world by storm, and the piece linked above goes into some of that as well. We didn’t see Deepseek coming (despite the company publishing research papers since mid-2024). We didn’t see Chinese electric vehicles coming. We didn’t see China lifting 800 million out of poverty coming. We didn’t see China reaching the Paris Climate Accords 5 years ahead of time (while the West has all but abandoned them). We didn’t see Huawei recovering from the embargo imposed on them. We didn’t see…
It seems that at every turn, the West is surprised that China is achieving great things while we still bicker over whether minimum wage should pay a living wage or not. It is not the 1800s anymore, and this refusal to look earnestly at China’s achievement, and work and learn with them is only going to dig the gap deeper much like Deepseek did.
In the West, we refuse to understand how China actually works — to take a long honest look at their system and understand that it’s completely different from ours in all ways. Instead, we prefer to rely on buzzwords such as “authoritarian”, “police state”, “one-party” so that we can comfort ourselves as still being “superior” to China while we are being left behind. And sure, it works… until the day it won’t.
The U.S. Navy, meanwhile, banned the use of Deepseek. It is possible that in time, the entire model (even its open-source version) will be banned in the U.S. In the meantime, a newly-introduced bill by U.S. Senator Josh Hawley proposes that anyone who downloads a Chinese AI model could face up to 20 years in prison or a one million dollar fine. It’s unlikely that this bill will pass as is, of course, but it tells us out loud what the ruling class is thinking and where their priorities are at nowadays. China is such a problem for them that they would go as far as considering decades of imprisonment for downloading — not even using — an AI model.
Deepseek’s tour-de-force wasn’t just that it’s more effective, it’s also that it’s completely free. There are no $20/month plan like Claude or ChatGPT have. One can use DeepSeek right now on chat.deepseek.com, all for free — although the web search has been disabled temporarily following a huge DDoS attack.
A DDoS attack works by sending lots of dummy data (or requesting lots of dummy data) to a web server every second (or even faster). In doing so, the web server is overwhelmed and cannot serve its web pages or services to any customers.
The attacks started as soon as the Deepseek interface on chat.deepseek was released, but peaked during January 27 and 28 at which point, experts say, the amount of data that was sent to the Deepseek servers exceeded the equivalent of three days of Europe’s internet consumption all concentrated on the Deepseek servers.
The attack was rightly investigated, and found that most of the infrastructure originated from the US, the UK, and Australia. The attack was also very complex, able to find new vulnerabilities in the Deepseek web servers even as the company fixed the issues. And, of course, the sheer amount of data that was sent to the Deepseek servers points to a very likely originator: the U.S. government. While we are not sure at this time who exactly ordered and facilitated this DDoS attack, we can certainly see that the scale, complexity and length is not something just anyone could have done.
At this time, it seems that the attack has been mitigated (thanks to Chinese cloud companies offering some of their infrastructure to Deepseek to increase their web server capabilities), but the story doesn’t end there.
Immediately after Deepseek released, other AI providers copied Deepseek’s reasoning capabilities in their own product.
The reasoning capabilities work by telling the AI to think about the problem first in a certain structured way, enclosed within <thinking> tags before the actual answer. See this (joking) example:
This reasoning system is one of the main reasons Deepseek is so efficient despite running far fewer parameters than other leading commercial models.
Of course, it was then only a matter of time until it became standard — especially now that U.S. tech companies can’t justify a $500 billion investment in their industry, seeing that Deepseek only cost 5.5 million dollars to train.
At this time, Perplexity AI (an AI search engine) is allowing 5 free uses of the reasoning model per day, and more if one pays for their $20/month plan. OpenAI CEO Sam Altman publicly praised Deepseek-r1 on Twitter, though adding “we will obviously deliver much better models” (emphasis mine), as if it was a given that OpenAI could outcompete Deepseek and not even a question, when the record shows the inverse happening.
However, there are still issues with these new models. Perplexity, like we mentioned, only offers 5 free searches with the reasoning model (which they self-host, as Deepseek-r1 is an open source model). Deepseek’s web search engine however is completely free and can find up to 50 sources per query, while Perplexity often stops around 20.
On OpenAI’s side of things, this is where it gets even more interesting. Just yesterday, ChatGPT (OpenAI’s interface to chat with the GPT models) released their o3-mini model to all users. This model also boasts reasoning capabilities, but some Twitter users, when trying it out, found something very interesting:
Some users have quickly noted that when asking questions to the new o3 model, OpenAI’s model sometimes switches to reasoning in Chinese for no apparent reason and without being prompted to.
It seems that, indeed, openAI copied Deepseek-r1 when training their model — when this was an accusation levied at Deepseek, that they used GPT models to train their AI on. OpenAI has yet to respond at the time of writing. And, like Perplexity, OpenAI is still charging $20/month for their models that Deepseek easily rivals with for free; though chatGPT does have some integrated services such as running Python code from the server or generating images that Deepseek doesn’t do yet, but will probably soon add.
Things are only heating up faster since Deepseek disrupted the industry, and it’s only been a week.
Great post! Txs a lot.
I can conclude that those countries native english speaking never grew up, some have the same mentality as in the Far West, same attitudes, same criminal behaviors, same society and philosophy. Others are adolescents that do not want to take responsibilities, vicious, spoiled and arrogant, completely ego centered, unable to see the other as an incentive to grow, to get better.
And the other third think they are God blessed so that give them the right to life and death even in the business.
And all of this is so sick, or better, DEEPly S(EE)icK!