06.02.2025
Making sense of AI wars
Has China caught up with the United States in artificial intelligence? That depends on what the race between the two countries is actually about, says Paul Demarty
The launch of R1, a new large language model (LLM) from the Chinese company, DeepSeek, caused a great conniption in the United States.
As Deepseek is apparently capable of producing roughly the same quality of output as the best LLMs available from US companies like OpenAI and Anthropic, but at a fraction of the cost, there was an immediate blow to market confidence in the big AI firms and their investors (Microsoft owns half of OpenAI, for example) with a trillion dollars being wiped off the tech-centric Nasdaq stock market index in one day. Panic spread among the AI cognoscenti.
Yassamine Mather provided a useful introduction to the technicalities of the AI sector and DeepSeek last week.1 Yet the technicalities are hardly revolutionary. Faced with certain constraints, DeepSeek’s programmers cleverly optimised their software - that is, changed it so that it does the same thing, but faster or with less resource consumption, etc - but the optimisations used are not world-changing technical discoveries. They are the same kind of things that programmers have always done when they try to squeeze a bit more juice out of a computer. The most striking effects, in the medium term, are not going to be in the far frontiers of what artificial intelligence can do, but rather in its impact on the relationship between the US and China, as they compete for economic and political dominance.
In my discussion here, I make two working assumptions: first of all, that at least some of the AI hype is true, and comparative advantage in this technical field will have a really meaningful impact on overall success in this new round of global competition. Secondly, that the appearance of total economic chaos in the early days of Trump’s new administration is false, and that its competitiveness in geopolitics will not be affected more than briefly. I have my doubts on both points, but it is still worth thinking through the implications.
Breakthrough
The DeepSeek breakthrough is, as noted, not a technological marvel, but rather the application of disciplined effort to improve the efficiency of, essentially, the same underlying technology used by the American AI models. It is easy enough to see why American firms were not able to do this. They had settled on ‘scaling’ as the main way of advancing their AI products - that is, throwing more and more chips at the problem - because it is relatively predictable (you can project quarterly spending and give more or less coherent answers to investors about how much it is all going to cost). It is also a strategy where the downside of not pursuing it is high - that is, suppose Google decided to go for optimisation instead of hoarding silicon. Suppose it does not work, and they cannot make more efficient models. By that time, OpenAI, Meta and Anthropic will have bought up all the chips. Not good!
For DeepSeek, this dilemma never arose, because successive US governments have imposed export controls on advanced silicon chips to China. Embarrassed US tech people have insisted that there must have been some evasion of sanctions here, and that DeepSeek must have had access to better chips than they claim. But in that case, they would presumably be competing over a far smaller supply of cutting-edge NVidia graphics processing units (that is, those that were successfully diverted around the export controls) than were available to the US in any case - or what? Did they all fall off the back of a lorry the size of Guangzhou? This is not serious. Plainly the constraints ensured that only major breakthroughs in performance would give any Chinese firm a chance of competing with the west.
DeepSeek’s breakthrough, indeed, may tell us more about America than China per se. The Financial Times reports that increased state involvement in Chinese tech industry has actually made it harder to take initiatives:
As state-owned funds in China have taken on a larger role funding start-ups in the past few years, the entrepreneurial ecosystem has felt pressured to guarantee returns for fear of losing the country’s assets. DeepSeek is distinctive among Chinese generative AI start-ups in that it has not raised any external financing and has therefore been free from these constraints.2
This has the whiff of FT negative spin. It cannot be denied, however, that DeepSeek is not typical of China’s tech sector. It is effectively a hobby company funded out of the disposable income of a hedge fund business called High-Flyer, run by one Liang Wenfeng. High-Flyer invested a lot in AI and high-performance computing for algorithmic trading and, like similar outfits in the US, recruitment to the firm is ferociously competitive. Quantitative trading is an extremely performance-sensitive business. Microseconds matter. It is thus no surprise that High-Flyer/DeepSeek had just the people to throw at a problem like getting a little more oomph out of old silicon. Even supposing the FT’s jaundiced view of the Chinese tech sector is correct, we should probably expect state investment to become less conservative, with such results dropping into its lap, just as this arms race heats up.
Worse is better
The major concern for the United States, under the circumstances, is that its tariffs and export controls have very obvious limitations. The usual imperialist playbook - of locking subordinate countries below you on the value chain - is visibly failing. It may well be the case that no Chinese supplier will be able to match the best chip designs coming out of the States or the state-of-the-art chip fabricators run by TSMC in Taiwan. (No US planner should be so complacent as to take this for granted, of course.) The questions posed by DeepSeek is: does that actually matter? How good is ‘good enough’?
‘Good enough’ is often a far better target, in many technical domains, than whizz-bang amazing. Departing from computers for a moment, you need only think of the AK-47 assault rifle; never especially high tech, but - precisely because of that - reliable, weather-proof, easily repaired and in continuous use for 80 years. By being a ‘worse’ gun than many flashier competitors, it makes the actual soldiers more effective, less dependent on logistical support, and so on. You do not hear it so much now, but in the 1980s and 90s the slogan, ‘worse is better’, was widely used in the software industry, to name just this paradox of very sophisticated programs somehow turning out to be less useful than relatively crude ones.
In the AI world, the big Silicon Valley firms have been improving their models by throwing silicon at the wall, brute-force training on enormous computer clusters at exorbitant expense. It was already widely known that returns were diminishing. OpenAI, for example, has faced enormous difficulties getting its GPT-5 model into a state that is worth releasing. Every attempt to train it takes months and tens - if not hundreds - of millions of dollars. If they succeed, what exactly is the point? Slightly better document summarisation - and with a real price of production that means it can never be profitable?
The silver lining to the DeepSeek humiliation is, of course, that the Chinese firm was good enough to give everything away, describing their techniques in depth and releasing the model as open-source. There is thus little comparative advantage for the Chinese state. Software optimisation, moreover, is also a world of diminishing returns: we should not necessarily expect fresh revolutions in training efficiency. Perhaps, then, the pendulum does swing back to simple brute-force scaling. The optimisations applied - a relieved OpenAI, Google, etc - can go back to building out vast data centres. The Chinese state is still faced with the daunting task of creating chip fabs that can compete technologically with TSMC and the like (or, I suppose, just annexing Taiwan …).
The two sides compete to find the first real killer app - emphasis on ‘killer’. The large language models will have their effects on the white-collar labour market and whatever else, but there is one pre-eminent axis of economic competition: the means of destruction.
In this respect, it seems fair enough to say that China has caught up. Its AI industry is ‘good enough’ to compete meaningfully with the US in the automation of warfare. We all await, with some disquiet, the result of this competition.
-
‘Coding good, politics bad’, Weekly Worker January 30: weeklyworker.co.uk/worker/1523/coding-good-politics-bad.↩︎