China Has Gyms?

DeepSeek and how the West is captured in a loop of sleepwalking and (rude) awakenings about China

Happy Year of the Snake to all my Chinese friends.

This morning, I made a series of phone calls to my dearest ones when a friend answered my WeChat call from the gym. Still spellbound by the latest Western collective awakening regarding Chinese advancements, I imagined that had our call been a LinkedIn post, one of the comments would have asked, dumbfounded: ‘China has Gyms?’.

The general public's reactions to the DeepSeek revelation in the West – of course, the company and its solution have been around for some time, and it was simply an update to its main language that triggered such a wave of emotion here – fall into two categories.

There are the ‘Quick-fingers' who immediately tap on their Apple Store to download the app because it’s so cool to have and they can chat with their friends about it. Once signed up – by the way, the flurry of new registrations is creating issues for DeepSeek – they will ask the machine the same silly questions they had typed into ChatGPT when it first became ‘a thing’. They will soon tire of their new toy because LLMs are far more advanced than any question those users could conceive and will return to swiping through their Instagram feeds.

Nearly all the others, whom I’d refer to as the ‘Haters’, spend their time attacking the technology – mainly because it originates from China – and predominantly using language and reasoning that are hardly specific to what they are discussing. “Try asking about Tiananmen Square,” is their favourite line this time. ‘To which my response would be: get yourself a bloody book. You know, that rectangular object with a spine, covers, and a stack of thin, inked slices of tree inside?’ Why do people believe LLMs are the best way to learn history? Why do people still think that Western LLMs are uncensored, even after the recent statements by Mark Zuckerberg?

But no harm done, for both types tend to subside after a while and return to their sleepwalking, ready to spring into action the next time something emerges from China. These cycles of sleepwalking awakenings have become a fundamental characteristic of the West. It is eerily akin to the infamous ‘Panem et Circenses’ (simple Latin for ‘Bread and Circus’, no need to ask the AI), an expression coined by the Roman poet Juvenal to describe the strategy of the elites in Ancient Rome, who kept the populace content with free grain and entertainment, distracting them from the true political actions – the wars, corruption, and malfeasance that led to the fall of the Republic and the rise of the Empire.

They make President Trump sound like a genius – and perhaps he is, after all, ‘nobody knows’ – when he declared to the stunned press that DeepSeek was a good thing, a positive development so that US scientists “won’t be spending this much, and you would get the same result… hopefully.”

I retained four teachings from this DeepSeek story.

First, you call it stealing what one should call progress. Perhaps Trump’s quote is simplistic, but it accurately reflects how the technological race for the definitive AI model (if the concept of definitive makes sense here) is conducted; or any race for progress, for that matter. There tends to be an innovator who enjoys exclusive advantages for a time. In the modern era, it is usually only a short period before new market entrants use the original product (or parts of its modules) as the basis for an enhancement, which will soon enough be overshadowed by the next competitor.

Nothing is ‘stolen’ in the process, by the way. It’s just how progress works in an interconnected world. And hear this: Americans do it too. A Fortune’s piece asserts that Meta has established four war rooms where two teams will try to decipher how High-Flyer [the venture that launched DeepSeek] lowered the costs of training and running DeepSeek, with the goal of using these tactics for Llama. Is Zuckerberg ‘stealing’ here, or ‘plundering,’ as DeepSeek is being accused of doing in 99% of the LinkedIn posts and comments I encounter? Stop complaining and get to work, using less money and energy: remember Moore’s law?

Not that everyone appeared to be able to compete in this leapfrogging game. With unfortunate timing (so typical of Europe), the DeepSeek success craze has coincided with the very brief release of what French authorities have termed the French response to ChatGPT. The chatbot – named Lucie and developed by the French government-backed company Linagora – was taken offline three days after its launch when it reportedly began to provide nonsensical answers to the simplest questions (my personal favourite was when it claimed, with the characteristic French grandiosity and self-importance, that the solution to the expression 5 x (3+2) was 17 – a prime number, by the way! – and that, in any case, "the square root of a goat is one.” There are too many articles, all hilarious, to choose one to link here; just google it ‘Lucie’). It appears that while Americans are ensnared in a sleepwalking-awakening cycle, Europeans never truly awaken, save for reaching a consensus on sustainable plastic bottle tops.

Second, it’s all about size. One of the most significant advancements of DeepSeek relates less to its energy consumption or response accuracy and appears to focus more on its size. I have listened to a very interesting BBC radio show during which the host was desperately trying to elicit expressions of concern from his three guests, all AI experts, who instead were enthusiastic about the progress that DeepSeek represented. One of the experts explained that he was overjoyed because DeepSeek was “a smaller module that could be integrated into someone’s computer“ and that the true AI revolution was not ChatGPT, nor even AGI, but the fact that ordinary people could access AI in their pockets without needing the internet connection.

I am not an expert at all, but this sounds like a significant departure from the theory that for AI to truly revolutionise our lives, full and uninterrupted connectivity is the necessary condition. Such a ‘closed-end' version of AI could also address privacy issues that, although they might have some merit, are slowing progress in the industry.

Third, use trade policy with care. Since September 2022, the U.S. government has imposed bans on the export of Nvidia’s most advanced AI chips to China. President Biden has enacted this through multiple rounds, the latest of which, heavily criticised, occurred just a few weeks before Trump’s inauguration.

If this action had been aimed at preserving US dominance in AI, then it failed spectacularly. Not only did it not stop China from developing a highly competitive alternative to all Western LLMs. But also, with fewer and cheaper chips, DeepSeek has been built at a fraction of the cost and with less chips than the others. It uses 2,000 NVIDIA chips in its architecture (and probably not of the highest available quality) compared to the 16,000 used by ChatGPT. Never underestimate the ingenuity of humankind when faced with difficulties.

Fourth, the debate surrounding the electricity AI needs to be reframed. For many months, I have read studies and interviews about the high stakes that the development of AI imposes on the world of energy and electricity. If DeepSeek is indeed viable and offers superior or even similar performance with significantly less energy, the debate on the electricity needs for AI development would increasingly resemble those concerning AI wiping out all our jobs or the infamous urge interspersed in the Club of Rome’s Limits to Growth. Furthermore, the substantial recent investment in energy infrastructure, particularly nuclear, by technology companies will need to be re-examined and may end up partially written off in their balance sheets. The risk here is that the renewed interest in nuclear technology – a significant push towards a more balanced energy production between fossil and non-fossil technologies – could diminish along with it.

Indeed, they have gyms in China, and guess what? They are open on New Year’s Day. Go figure.