So far, the common perception has been that to play the AI game requires such huge investments and infrastructure that any attempt in this direction is discouraged. China has shown that this is not the case. And now it can also advance to the conquer ‘natural intelligences’ by Andrea Monti – Initially published in Italian by La Repubblica – Italian Tech
The noise stirred up by the latest release of DeepSeek, the AI model developed by the Chinese company of the same name, has covered a number of signs that provide fairly clear indications of how the AI market is changing. This is especially true of the reaction of OpenAI and other giants in the industry, which, as the BBC reports, have suffered some sort of “reputational downsizing” and potential market problems because of DeepSeek’s demonstrated ability to compete with Western equivalents at lower costs and higher levels of efficiency
AI is not (only) brute force
To understand what is happening, it is first necessary to start with the more abstract questions concerning the relationship between computing power and the ability of software to harness it.
DeepSeek shows that while having access to a relatively small number of GPUs, and moreover, not the latest generation, it is possible to compete with more capable Western, or rather, American models. One of the reasons for this result is the more efficient management of the software component, which shows how greater intelligence can compensate less “brute force.”
In this regard, it is important to remember that the problematic nature of the relationship between bloatware-or, in general, between writing bad and inefficient code-and the negative impact on the performance of the hardware that runs it has always existed, and it is not just about AI. One only has to think of the decades-old quarrels between the Windows and *NIX worlds already documented in 1999 by Neil Stephenson in his short essay In the beginning was the command line and the equally historic definition of spaghetti code published on the Free Software Foundation website.
AI, therefore, also does not shy away from the rule of thumb that characterizes the evolution of emerging technologies: at first it relies on “brute force,” crudely exploiting available resources, and then gradually optimizing its use through experience-based technical innovations and evolutions.
This process, as mentioned, is also evident in the field of artificial intelligence where, however, the rules of the game are still based on the ability to accumulate more and more powerful GPUs. However, such a belief overlooks the fact that growth in computing power is not linear with respect to energy consumption and other variables, such as precisely the ability to invent new software approaches, that define what it takes to develop an LLM. Which brings us to the issue of the actual cost-effectiveness of DeepSeek and why this is a problem for its competitors.
DeepSeek’s attack on OpenAI &C.’s marketing strategies.
Despite the emergence of approaches such as so-called “frugal AI”-which promotes more frugal and intelligent use of resources-the belief that more computing power equals more artificial intelligence remains widespread. With a nightmarish compulsion to repeat, the AI industry has adopted marketing strategies that are not at all different from those practiced in other technology sectors. Just think, for example, of the way PCs, cameras, and smartphones are promoted, all based on the “bigger” (the processor, sensor, screen, etc.) is the “better”
The reason for the choice is clear: “bigger” is “better” is a much easier concept to understand for those who hold the purse strings and need to loosen them, than unique selling propositions based on mathematical or engineering arguments-let alone those related to software-which are complex, hard to summarize in a few catch phrases, and difficult for users, investors, and decision makers to understand.
So far, as far as AI is concerned, this strategy has proven to create barriers to market entry. To date, in fact, the common perception is that to play the LLM game requires such enormous investment and infrastructure that it discourages any attempt in this direction. So, by sustaining this narrative, the giants of the sector succeed in, on the one hand, monopolizing financial resources and, on the other hand, keeping out potential newcomers.
However, thanks to DeepSeek, there is a sudden sense that it may be possible to “do more” but “spend less,” and that, as a result, it may not make much sense to invest in technologies and hardware that are unnecessarily expensive to buy and manage. This is all the more true when one considers that there are such warnings in the chip sector as well, where new ARM processors promise affordability and performance, so much so that as early as 2020 nVIDIA tried, unsuccessfully, to acquire the British chipmaker.
Was DeepSeek developed to destabilize U.S. businesses?
It is clear that the argument just made about the greater desirability of DeepSeek works if and only if the development costs are as publicly stated, that is, a fraction of those incurred by U.S. competitors. Although, in fact, the approach to the overall design of the Chinese model is clearly efficiency-oriented, it is unclear whether, and if so to what extent, there has been even indirect support from the Beijing government, for example in terms of access to the energy and computing power needed to train the model or other forms of support.
If, in fact, the lower cost of DeepSeek’s development was even partly possible because of state aid, it would be legitimate to raise some doubts about whether the project is really more sustainable than its U.S. competitors and to ask whether, instead, we are not faced with the use of economic leverage to disrupt the market by lowering the value of competitors. The latter, in fact, run the risk of being in the position of having to chase DeepSeek instead of dictating the pace and of suffering the introduction of new technologies into the market instead of controlling them. Moreover, they would lose their privileged status as suppliers of “raw material” for the rest of the supply chain that develops LLM-based products and services, since DeepSeek is more efficient, cheaper but, above all, “open source.”
The weaponization of open source
For some time now, the concept of open source -a generic term that, summarily, denotes the right to access the information necessary to understand a technology and the right to use it freely- has been moving toward losing its original role as a tool that fosters the free circulation of knowledge to become an important component of states’ geopolitical arsenal.
It is no mystery that DeepSeek was blatantly developed in compliance with the guidelines set by the Chinese authorities in relation to how to respond to issues involving socialist values and policies. In this sense, such a choice is the perfect corresponding of the “ethical constraints” embedded in proprietary and open source LLMs already available in the West.
However, the decision to release Deepseek in open source could not only reduce the value of the AI giants, it also risks taking away user market share from them. Companies, developers and researchers might, in fact, be interested in accessing sophisticated technologies without having to pay expensive licenses or suffer other limitations. We would be facing a complementary situation to the one created by the choice made by Huawei to release HarmonyOSNext (Android’s competing operating system) as a “free” version, capable of running on a wide range of devices, from wearables to terminals, potentially enabling the creation of a global technological infrastructure independent of Western technologies. And it is barely worth mentioning that Huawei also produces AI GPUs optimized for DeepSeek.
The impact on natural “intelligences”
Finally, if it spreads sufficiently, DeepSeek could become part of a strategy for spreading in ideas that do not necessarily conform to Western principles and values.
Lycurgus, writes Plutarch, in Parallel Lives, banished from Sparta all foreigners who had no good reason to stay having the fear “that they would spread something contrary to good customs. Foreigners bring foreign words; these produce new ideas; and on these are built opinions and sentiments whose discordant character destroys the harmony of the state”
Easy -not that easy- to do with people, extremely more difficult if “new”, or rather “different” ideas are conveyed by software that can be duplicated, modified and circulated without any restrictions. It is hard not to think of the TikTok issue and the reasons, real or supposed, that led the U.S. administration to order its forced sale.
Although a direct, immediate, and large-scale impact of this kind is unlikely, it is equally reasonable to think that in the long run “heterodox” ideas may more easily infiltrate mainstream thinking by passing through smartphone screens and contribute, if not to redefining it, at least to orienting it in a way that is more favorable to China or, which is the same, more critical of its own governments.
On the other hand, softpower is not only exercised through music, movies, and looks.