Introduction: Launching into a New Frontier?
On July 20, 1969, as Neil Armstrong planted his foot on the Moon, he spoke the famous words “That's one small step for man, one giant leap for mankind.” His universal words were an interesting choice given he was on the Moon to establish the dominance of the United States in space and regain the lead in the Space Race after the Soviet Union had a series of early successes such as launching the first satellite - Sputnik, put the first man and woman in space, performed the first space walk, and launched the first probe that impacted the Moon.
Today, progress in space is still being made but NASA and SpaceX have a clear lead in the field, and there's not as much national urgency as we saw during the Space Race. The focus has now shifted towards what has been dubbed the AI Race by the media. The main players in this race are the United States and China. It had been assumed that the US had a several year lead on China in AI, by figures like former Google CEO Eric Schmidt. Less than a year after that prediction, Chinese firm DeepSeek released its R1 reasoning model that was comparable to o1 from OpenAI.
While OpenAI had been hiding o1 behind a paywall, DeepSeek released its reasoning model for free in its app, open sourced most of the code, and made it competitive with o1, but it cost a fraction of the price to train. DeepSeek gripped headlines for days, and it was even mentioned by President Trump. It had the United States spooked. If the AI Race wasn't on before, it certainly is now.
In this editorial, I want to criticize the term AI Race. The main basis for my criticism of the term is the fact that many of the AI models that are being released, are being made open source. DeepSeek is the most famous example of an open-source model, and that's from the supposed competitor, China. American companies have already taken the code, removed the censorship Chinese law requires for its companies, and have even monetized it.
While the chief American AI firm, OpenAI, is keeping its AI development a secret, Meta has made its Llama AI models open source and communities like Hugging Face are doing their best to make open source models more available and is even developing software to plug gaps, including with its Open Deep Research project.
While I think it's true it could be construed as an AI Race with regards to weapons of war, I'm a lot more skeptical about the description being applied in the context of large language models, especially when it involves the significant sharing of the source code of models.
In this editorial, I will look in-depth at the competition between the US and the USSR during the Space Race and the competition between the US and China in the AI Race. I will look at how they are different and outline a more constructive framework for understanding what's happening now.
The Space Race: A Cold War in Orbit
The Space Race got underway on July 30, 1955, when the United States declared its intent to put satellites into Earth orbit. This was reciprocated by the Russians four days later who said they would also do the same in the near future. The Space Race came as a direct result of the Cold War where the capitalist West was facing off with the communist East, which had gained ground around the world after overthrowing the Nazis and their allies in Eastern Europe during World War II and as a result of decolonization in the Global South.
Building sophisticated rockets and satellites was important for both sides. The rockets could be adapted for nuclear warheads in the case of a hot war and the satellites were useful for intelligence gathering - satellites are still widely used for this today. Feats like getting to the Moon, while less directly useful in the case of war, acted like propaganda for the respective nation's economic system.
For the Space Race to finally end, it took the complete collapse of the Soviet Union, after which, Russia collaborated with the US and its allies on space with projects like the International Space Station.
Unlike today's AI Race where many players are releasing their source code publicly, the space programs run by the US and the USSR were both secretive and nationalized. Many details about AI advancements can be found in scientific papers too, whereas in the Space Race, there was tighter control over what information was being shared to help each of the countries keep their respective advantages.
As a result of the Space Race, both countries developed their own rocket technologies, with the German V-2 rocket as the shared basis. They both developed their own satellites such as Sputnik 1 and Explorer 1, they both developed their own human spaceflight system independently, as well as their space suits and life support systems.
This is very different to the AI Race where advancements in the AI space are being shared. We are even being given roundabout information from companies like OpenAI who choose not to open up their source code.
The main consequences of developing technologies in silos were slowed scientific advancements, duplicated efforts, and increased geopolitical tensions.
The main goals of the Space Race, aside from the military applications of the satellites and rockets, was the ability to show off which country had the best political and economic system. It also acted to give the countries more prestige when they were able to achieve milestones with the technology they had achieved.
With the so-called AI Race, America is keen on maintaining its leadership for further economic benefits, and to promote its values of freedom and human rights. One significant issue we saw with DeepSeek was the embedded censorship which stopped the model from criticizing the ruling Communist Party, its leader, or its geopolitical ambitions.
Finally, it’s worth pointing out that the two races are different in that the Space Race was primarily driven by the US and Russian governments, rather than private business, with largely military ends. On the other hand, the cutting edge AI tools are being developed by private companies with a focus on productivity and efficiency. It’s true that there are military applications for artificial intelligence, which are dangerous, but primarily, the concern seems to be economic.
The “AI Race”: A Different Kind of Competition?
If the current situation around artificial intelligence was truly a race, we would expect businesses developing this software to hold all their secrets close to their chest, making it harder to catch up, but that is not what we are seeing. While some businesses can be categorized like this with their closed source models, organizations working on open models such as Meta, Mistral AI, DeepSeek, Alibaba, Stability AI, and Hugging Face. Even companies like Google, which have developed closed source models, have released development tools like TensorFlow to advance AI development. This level of openness seems to be a more collaborative phenomenon, rather than a competitive race.
Not only are companies releasing their software as open source programs, but their research is also being published in open pre-print publications such as arXiv allowing for the greater levels of collaboration. This further detracts from the idea of an AI Race.
The open publication of AI research papers means that companies from different countries are integrating the latest ideas to improve their models. Even in the case of OpenAI, which is closed source, the recent release of DeepSeek R1 impressed Sam Altman who put a special emphasis on noting what R1 delivers for the price. Now that OpenAI knows more efficient methods exist, and has R1’s code, it can make its own models more efficient and available at a lower cost to consumers.
The efficient techniques used by R1 also highlights another key difference between AI competition and the Space Race - that is, it’s mainly driven by economic and commercial factors, rather than for military ends or proving if particular countries have a better economic model.
Just days after the release of DeepSeek R1 and its unlimited thinking requests, suddenly OpenAI releases o3-mini for free users (it still has a limit), and in recent days, Google released its thinking version of Gemini Flash 2.0 for free users. This type of behavior is common in the AI race; charge an incredible amount of money from premium users until one player makes a new feature, such as reasoning, available for all and then the other players are forced to follow suit.
While it’s true that AI does have military applications, these consumer facing generative AI bots are focused on economics - such as making money for the business that makes them and adding productivity for those who consume them.
Ultimately, the AI Race narrative is an oversimplification and rather too dramatic. For the most part the development of AI and products that come as a result, are just an example of normal economic competition. In fact, the field has much more collaboration going on compared to other industries.
DeepSeek: Open Source and Cross-Border Adaptation
I heard about DeepSeek a couple of days before awareness of it exploded across international news channels. Sure, it wouldn’t talk about the Communist Party of China, its leader, or the country’s territorial disputes, but it would do pretty much everything else. Not only did it scour up to a huge 50 online news sources, but it also used reasoning, in an unlimited quantity, to give me better responses.
I could ask it all sorts of things such as for comprehensive stock exit strategies, I could ask it who it thought would win particular horse races, and I could get it to answer all sorts of speculative questions such as what month and year does it think humans will get to Mars.
Compared to the offerings at the time from Google and OpenAI, DeepSeek stood head and shoulders above. For free users, DeepSeek was the only company offering unlimited access to the web and reasoning combined. As a result, it climbed to the top of app stores and spooked the US tech industry and investors who thought it spelt the end of the AI bubble.
Unlike those inferior models from Google and OpenAI (the ones available to free users), DeepSeek R1 was an open source model. This means that the company divulges most of the inner workings of the model so that others can take and adapt it, as long as they follow any provisions set out in the open source license. While largely open source, some components remained closed including the methodologies for data collection, model training, and scaling laws.
Once DeepSeek had become widely known, and name dropped by President Trump, the company started being impacted by a distributed denial of service (DDoS) attack - the Venom DDoS Community has claimed responsibility for this attack.
While DeepSeek was being attacked, Meta started dissecting the code to improve its own Llama models and Perplexity hosted an uncensored version in the US for use on its platform, though it's not unlimited like the one hosted by DeepSeek.
The open source AI community, Hugging Face, has also hosted the DeepSeek R1 model on its HuggingChat platform, but this model doesn’t have as comprehensive web access like that found in the DeepSeek app.
The best example of DeepSeek’s technology being monetized by American companies is in the case of Perplexity which offers a handful of R1 searches to free users before asking them to upgrade to a premium tier. By monetizing reasoning models, companies like Perplexity can limit access to its compute power so that the service remains fast for all users. This added sustainability also ensures it can continue to grow in the future. Making money and ensuring access to a stable platform is in no way at odds with the open source model, open source just means that people can share their research and tools are potentially developed quicker.
Aside from the business benefits and research benefits, open sourcing models can also democratize access to AI technology. AI engineering can be a very lucrative career for anyone willing to put in the time and effort required and by opening up model source code, those who are interested, and learn more about how models are made, potentially opening up careers. It also allows tinkerers to run models on their own hardware for unlimited access and more privacy.
If the AI Race was truly a race comparable to the Space Race, China wouldn’t be allowing its models to be open sourced so that American businesses can profit on them and boost America’s economy. For the most part, it is just like any other competition we see countries engaging in.
I think it’s also generally a good thing that countries and businesses are being this open as it means we could reach artificial general intelligence and artificial superintelligence much faster. Just look at how the AI we currently have is contributing to fields like healthcare and material science - imagine what will be unlocked when we have AI that surpasses what humans are capable of. We will have lots of new breakthroughs in many fields which will lift up everyone’s living conditions. These more powerful AI systems will also require more power, this will likely see investments in renewable energy and nuclear energy, helping the shift away from fossil fuels.
Monetization and the Shifting Sands of Technological Advantage
Since the early days of the generative AI era, which kicked off in November 2022 with the arrival of ChatGPT, AI companies have sought to claw back their losses from customers by offering better language models or restricting the number of requests free users can make. To attract customers, companies need to offer the best language models and avoid being overtaken by their competitors.
Up to the present, OpenAI has largely been seen as the leader in the space. It has recently released the Operator web browser, which is AI operated and Deep Research. Both products are groundbreaking in terms of functionality. Other companies like Google and Meta are also trying to branch out from chatbots too, Google has developed NotebookLM which lets you upload sources such as reports and then get help from AI to sift through the massive loads of data. Meta has also started expanding its AI tools with the launch of AI-supported document editor.
An important consideration for AI is the hardware that it runs on. Nvidia has been a major beneficiary in this respect because its hardware has been what’s powering models like ChatGPT. Sam Altman, head of OpenAI also expressed last year that $7 trillion was needed to overcome hardware limitations affecting AI progress.
When DeepSeek R1 came along, claiming to be 27 times cheaper to run than OpenAI’s o1 reasoning model, it sent stocks in the US tumbling because there was a realization that maybe so much compute power wasn’t needed after all and also, just how the hell were AI companies planning to make money if DeepSeek was giving unlimited access to its AI and open sourcing all its code?
The release of DeepSeek put pressure on OpenAI to release its new o3-mini reasoning for all users, including those who don’t pay. While Sam Altman publicly says he is excited about the competition, there are also claims that its lead is also narrowing, which could be worrying the company internally. DeepSeek R1’s launch has not been all bad news for the American players though. The popular AI search engine Perplexity was quick to add R1 to its collection of reasoning models as part of its Pro search. Free users get a handful of Pro searches per day, but to unlock more, you have to pay. The inclusion of R1 could have made the subscription more appealing, leading to increased revenues for the company. Perplexity was able to uncensor the model and host it in the US, removing any political bias concerns that exist with the DeepSeek-hosted model.
At the time of writing, it’s fair to say that despite the benefits of open source models, they have not managed to surpass the proprietary leaders yet, although, they’re not too far behind. In terms of hosted services, these open source models contribute to a more competitive playing field. They also provide unique benefits such as being able to run them on local hardware, which is good for privacy and having unlimited access in terms of requests.
Conclusion
In this editorial, I have tried to dispel the notion that what’s happening in AI today is not an AI Race comparable to the Space Race, but rather, ordinary competition, with a whole lot of collaboration. The frequent publications of AI research papers and the release of open source models like DeepSeek’s R1 stand in contrast to the largely secretive development of rockets and satellites in the Space Race, which the media tries to draw comparisons with.
The development of new AI tools is moving quite rapidly with cutting edge features usually hidden behind a paywall until one of the market participants decides to release those features and the other players follow suit. Thanks to the greater sharing of information compared to developments during the Space Race, there is less reproduced effort going on and things are moving along quickly. As mentioned earlier, another difference between the Space Race and the AI competition is that the former was between governments, and the latter by private companies, seeking market share and profits.
Going forward, it is unclear what will happen between the closed source AI companies like OpenAI and the open source models. Monetization of AI is an important factor as it can help companies recruit and retain the best talent in the industry, which free open models may struggle to do.
According to the OpenAI Careers page, a Software Engineer (Full Stack) can earn between $160,000 and $385,000, plus offers equity. Obviously, companies like Meta and DeepSeek are able to pay these great wages despite having open models, but it does make the task easier when models are being monetized. If a company like DeepSeek with open source models does end up taking over OpenAI, it will likely happen all of a sudden with a new model landing and OpenAI not being able to respond quickly, only time will tell what happens going forward though.
Image via Depositphotos.com
1 Comment - Add comment