Apply Now 
As Hamad Bin Khalifa University’s Dr. Sanjay Chawla sees it, despite hours of research and countless academic papers, there’s still no such thing as a “Science of AI.” If there was, the emergence of DeepSeek would not have been as controversial as policymakers and media would have us believe.
One of Donald Trump’s first actions as US President was to announce the $500 billion Stargate project. Launched January 21 in the presence of industry heavyweights including Oracle’s Larry Ellison, Softbank’s Masayahi Son, and OpenAI’s Sam Altman, the initiative seeks to maintain the United States’ supremacy in AI by investing in a massive computing and energy infrastructure. Within a week DeepSeek-R1, a Chinese Large Language Model (LLM) developed at potentially a fraction of the cost of US counterparts, was also launched. DeepSeek-R1 wiped over a trillion dollars off the US stock markets and prompted concerns that society had witnessed a 21st-century Sputnik moment.
The above series of events should not be regarded as country A trying to demonstrate that it is better than country B (or vice-versa) but an indictment of “experts” who fail to realize that AI technology is still at a very early stage in its development. Government leaders instead need a panel of true experts who are not afraid to speak truth to power and can cut through the unnecessary hype surrounding AI.
Here is the “bitter truth” about AI. Despite hundreds of thousands of research papers written every year, there is no established “Science of AI” and everyday use of AI is primarily based on empirical findings. Moreover, if the “Science of AI” had already existed then people would not be shocked by DeepSeek-R1’s performance - it would have been predictable.
In 2017, Google researchers released a new form of AI model called Transformers which provides the basis of all current modern LLMs including ChatGPT and DeepSeek. Transformers were able to correctly disambiguate sentences like the following from a well established benchmark:
1. The trophy doesn’t fit in the brown suitcase because it is too large (Question: Which is larger, the suitcase or trophy?).
2. The trophy doesn’t fit in the brown suitcase because it is too small (Question: Which is smaller, the suitcase or trophy?).
By conducting extensive experiments costing millions of dollars, an empirical finding emerged (called scaling laws) which led to the conclusion that big models and big data can be used to create AI machines that are extremely fluent in conversation on almost any topic. There was opposition within Google to releasing these models (sometimes pejoratively dubbed as “stochastic parrots”) and that is where OpenAI stepped in and released the model to the public. And the rest, as they say, is history.
However, as far as the “Science of AI” is concerned, no new major progress has been made in the last eight years. Yes, progress has been made in creating better ways to tease out information from AI models and making them more efficient but that would not qualify as scientific progress. The biggest open question in this area is whether AI models “know anything” or are just proficient “sentence completers.” The problem of hallucination, where AI models make things up, is primarily because we don't know how they really work. We don’t even know whether a “hallucination” is a feature or bug of AI systems.
There is also a deep urge to draw parallels between Stargate and the Manhattan Project or Apollo Space Program. Comparisons with AI can nevertheless be regarded as what are known as “category errors.” These are regarded as a “type of logical fallacy where things that belong to one kind are treated as if they belong to another.”
Before commencement of the Manhattan Project the science of nuclear fission was already well established. Between 1932 and 1939, a series of scientific studies confirmed that a nuclear chain reaction is possible. What was left was a major undertaking of creating enriched uranium which was primarily an engineering exercise. Similarly, the science behind a “soft landing” on the moon was well known many years before the Apollo Mission and even now remains very hard to achieve in terms of engineering. In both cases science preceded engineering (it is never a straight line though).
In December 2024, HBKU released “Fanar,” an Arabic-centric Generative AI Platform. Fanar is possibly the first 7B sized model that was designed and implemented from scratch by a university entity. More university-led efforts are needed to bring the “Science of AI” to the center stage.
Dr. Sanjay Chawla is a chief scientist at Hamad Bin Khalifa University’s Qatar Computing Research Institute.
This piece has been submitted by HBKU’s Communications Directorate on behalf of its author. The thoughts and views expressed are the author’s own and do not necessarily reflect an official University stance.
This article was originally published by Al-Fanar Media and is reproduced with full permission.
