The illusion of intelligence: What the AI race misses

There is a surrealness to the current fascination with artificial intelligence (AI). The technology is not new. 

The most famous application of AI, OpenAI’s ChatGPT, was released in November 2022, five years after its core architecture was first developed at Google. Around the same time, Google’s DeepMind made waves with AlphaGo’s victory over Go champion Lee Sedol in 2016.  

But the history of AI benchmarks stretches back much further: in 1957, Herbert Simon and Allen Newell’s General Problem Solver successfully tackled the Tower of Hanoi puzzle — 68 years before Apple researchers posed the same challenge to modern models like Claude 3.7 and DeepSeek-R1.

To be sure, certain developments did cause stirs. Some Chinese People’s Liberation Army officials saw AlphaGo as a precursor to a general strategising agent exceeding human capability. In the 1980s, 30-odd years before AlphaGo, US Defense Advanced Research Projects Agency planners saw the promise of “generic software systems” underpinned, in part, by AI.

Every development linked to the promise of general intelligence, however, has failed.

Today, as the US and China tussle over AI and their presence in Gulf states, this history is often forgotten, as though this time is undoubtedly different. Is it?

An accounting of the facts will give us clarity.

American firms traditionally set the AI development agenda. Companies including Google, OpenAI, Meta and Anthropic, among others, have led the construction of systems underpinned by artificial neural networks. These networks rely on massive training datasets and enormous amounts of computing power to learn statistical associations of data. Access to advanced chips, data centres and energy is therefore seen as critical to AI’s development.

America’s AI lead stems from a concentration of world-class talent, research facilities and access to requisite compute (including advanced chips). Though the supply of chips has diversified since the generative AI boom in late 2022 — and China has made limited progress in achieving self-reliance in semiconductor production and adoption in the face of severe US-led export controls — these necessary components in AI development are finite. The global playing field is thus unbalanced.

However, the US tech sector was roiled by the release of DeepSeek-R1 in January 2025 by Chinese startup DeepSeek, independently replicating the advancement that OpenAI had previously made in September 2024 with its first “reasoning” model “o1”. Nvidia’s stock dropped 17% in mid-January over fears that DeepSeek had constructed its model with only a fraction of the computing power thought necessary, thereby reducing the barriers to entry in advanced AI development and widening the window of opportunity for actors in China to “catch up” to American model development.

Such fears have since abated. Speculation about DeepSeek’s access to advanced chips (either through stockpiling prior to export controls or through smuggling) instead intensified existing debates about the efficacy of US-led export controls. Research firm SemiAnalysis judged that DeepSeek had access to 50,000 Nvidia H-series chips that were mostly export-compliant, but 10,000 of which were export-restricted H100s.

The Chinese high-technology ecosystem nevertheless perceives itself to be ascendant. Leadership rewarded DeepSeek with access to state-funded data centres while promoting Manus AI on state media in March.

In Hangzhou, where DeepSeek is based, a thriving tech culture exists comprising firms including MiniMax, Moonshot, Alibaba, Baidu and Huawei. Liangzhu, a suburb of Hangzhou, hosts a tech culture echoing the early days of America’s Silicon Valley.

Some Chinese firms are looking abroad. In June, OpenAI identified Chinese startup Zhipu AI as a leader in Chinese AI — previously backed by Saudi Aramco’s Prosperity7 with US$400 million — with the company offering AI infrastructure solutions in emerging markets. Huawei, for its part, is attempting to export a small amount of its indigenously produced Ascend 910B AI chips to the United Arab Emirates (UAE), Kingdom of Saudi Arabia (KSA) and Thailand.

That said, American policy shifted in May during President Trump’s visit to the Gulf region. Two AI deals with the UAE and KSA stand out.

First, the US and the UAE signed an “AI Acceleration Partnership” launching the “Stargate” AI infrastructure initiative. Stargate involves a joint US-UAE investment in a 1GW data centre cluster in Abu Dhabi, with 200MW online by 2026 and an Emirati commitment to expand US infrastructure. 

The deal permits the Emirati import of 500,000 top-of-the-line chips every year until 2027, with one-fifth reserved for Abu Dhabi-based AI conglomerate G42 — a previous flashpoint in US-UAE diplomacy given the firm’s alleged ties to Chinese intelligence services. The deal is, importantly, not complete, with national security concerns delaying its finalisation.

Second, KSA launched HUMAIN during President Trump’s visit, serving as a platform for AI model training, inference and services. It is backed by American firms, including Nvidia and Qualcomm. There are 18,000 Nvidia GB300 Blackwell chips tentatively approved for export to HUMAIN to power 500MW of data centre capacity within the country.

As the Middle East Institute’s Mohammed Soliman notes, these deals indicate that the UAE and KSA are “positioning themselves as potential backends of AI for emerging markets across Asia and Africa” through US-aligned projects.

American policy debate now centres on whether the US is “outsourcing” its AI industry to the Gulf, creating an unhealthy reliance on the latter’s access to energy for powering data centres.

But the debate suffers from amnesia. Today’s “reasoning” models from major American AI labs are the hook for AI relations, but converging research indicates they are not the pathway to human-like intelligence.

The premise behind reasoning models is that by training on not just the outputs of human reasoning (e.g. human-generated texts) but also on verbalised “thought processes” — such as “chains-of-thought” — models can learn to reason in human-like ways. Post-training, these models are further optimised to generate problem-solving trajectories resembling such chains, which are then weighted and fine-tuned.

This premise is flawed. First, it conflates language text with thought. Human thought is expressible through natural language texts that have the linear structure seen on this page. But thought itself — scaffolded by language — is hierarchically structured within the mind. “Chains-of-thought” emulate only the former. 

The premise also fails on its own terms: “chains-of-thought” are not predictive of model outputs. In one experiment, a transformer model trained on deliberately corrupted chains-of-thought outperformed those models trained on uncorrupted, sound chains-of-thought. The researchers speculate that “chains-of-thought” were not predictive of model output — suggesting that the reasoning traces do not actually guide the answers.

This aligns with other findings. Models that are prompted to “show their work” on math problems often perform worse than when they are only required to produce final answers. Even slight variations in benchmark design (e.g. how questions are phrased or structured, the GPU cluster used) can significantly shift performance, indicating the instability of these evaluations.

Reasoning models lack generality. They deteriorate on longer planning problems, hallucinate answers to unsolvable tasks, and attempt to rationalise these fabrications. New types of hallucinations have emerged as well: one study of ten models found a consistent tendency to overgeneralise clinical trial results — a high-stakes task — with such behaviour substantially increasing in newer models like ChatGPT-4o and DeepSeek.

Apple’s The Illusion of Thinking provoked controversy by highlighting sharp performance declines on more complex tasks — but growing evidence converges on this point. One recent study found that as mathematical problems require novel reasoning, frontier models exhibit “sharp performance degradation as problem complexity increases”.

The reasoning models that are the subject of geopolitical intrigue fail to exhibit the characteristics of general intelligence. Taking their actual trajectory into account clarifies three aspects of US-China-Gulf AI relations. 

First, the stakes and scope of AI relations are considerably tempered. Fears of “outsourcing” American AI leadership hinge on these models being the key to not only continued progress, but the right kinds of progress. They may not be. 

The US cannot “outsource” a set of capabilities it does not possess (just as it cannot “outsource” non-existent energy capacity). The relative increase in the US government’s willingness to permit deeper joint US-Gulf AI development projects, as implied by the May 2025 deals, may extend the reach of American technology, though an upper bound on their outcomes is expected.

To be sure, this follows China’s existing willingness to share technical expertise with Gulf states to secure foreign investment and deepen its technological influence — an often overlooked but important factor in facilitating AI implementation and pursuing AI supremacy.

If played wisely, China could leverage US delays in finalising chip and data centre deals to its advantage by combining compute-efficient AI developments with the promise of being a more reliable partner compared to American dysfunction.

Yet the scope of such engagement — however skillfully executed — is constrained by the persistent limitations of current-generation AI models. The value of reasoning models, to the extent that they excel in certain areas over others, will ultimately be shaped by traditional market dynamics.

Second, the importance of diversified basic research and public funding is re-affirmed.

Private US over-investment in transformer-based techniques could hinder progress toward the broader capabilities needed for developing generally intelligent systems — such as performance guarantees in sensitive domains, robust abstraction and generalisation, explainability, continuous learning, and compute efficiency. This risk is heightened as the US shifts away from public funding for basic (high-risk, high-reward) research, potentially leading to a loss of scientific talent.

This provides an interesting window of opportunity for China. Some Chinese actors recognise the need for foundational research, including DeepSeek CEO Liang Wenfeng, who has chosen to reinvest the firm’s revenue into such efforts as it works on the delayed DeepSeek-R2. State-backed efforts critical of generative AI have also arisen.

Still, even having favoured tailored models before general-purpose language models like DeepSeek appeared, the bent of China’s AI ecosystem has been towards models that adhere to the principles of current-generation AI. It remains to be seen what balance China will find here. The constraints it faces, on access to chips, talent, and even energy, while improving, are real and pervasive. AI research problems are just as — if not more — overwhelming, with barriers that apply universally to all states.

Finally, compute and basic R&D are tightly coupled in the development of AI. Should the US, China or Gulf states separate them — particularly by emphasising compute over basic research — this period in AI will not be different, after all.

Related: How China’s all-in-one platforms are reshaping the AI battlefield | Why US sanctions can’t stop China’s AI rise