By Dr. Sanjay Chawla
Truly innovative AI should ultimately be thoughtful, constrained and able to identify niche areas where its application brings advantages and competitive edge, writes Dr. Sanjay Chawla. [GETTY]
The word “irrational” has already started emerging in mainstream conversations around AI. In a recent interview with the BBC, Alphabet’s CEO Sundar Pichai says that investments in AI have “elements of irrationality.” This is a bit of an understatement.
OpenAI, beacon of the AI revolution and still a start-up, is committed to investing $1.4 trillion in AI infrastructure, approximately $200 for every human alive. The company has warned that it will keep losing money until at least 2028. This is a massive risk not just for investors but the entire world economy, given that nearly 20% of the US stock market is owned by “foreigners” and investments are increasingly geared towards AI. Whether the global economy stays afloat or crashes is now hinged on the success of one US start-up.
The good, the bad…
There is no doubt that AI in general and Generative AI specifically has been a total game changer. Despite concerns about “hallucinations” and being prone to errors, AI tools have become mainstream and aggressively used for work, rest and play.
AI is fun, at least for low-risk activities. Want to make dark chocolate at home with jaggery, instead of sugar? Nearly all AI tools, including Qatar’s Fanar Arabic Large Language Model, will provide a near perfect answer to this and other questions.
Conversely, AI tools are also being used for potentially high-risk activities. Platforms have experienced a surge in requests for medical advice, a reflection that many users find these tools more empathetic than doctors. It has also been claimed that ChatGPT encouraged a user to commit suicide. While such an outcome points to a major safety flaw, AI’s march remains relentless and shows no sign of stopping.
The conventional wisdom is that building AI tools is very expensive, relies on cutting edge computing infrastructure, large amounts of data and a super-skilled workforce. This ‘version’ of AI is also considered out of reach for many countries, especially in the Global South. But that’s only partly true. While AI leans on high quality data and infrastructure, it does not have to be ‘cutting edge.’
Neither do we need an army of skilled AI engineers to make things happen; just a few folks who understand the basics and can customise code. Truly innovative AI should ultimately be thoughtful, constrained and able to identify niche areas where its application brings advantages and competitive edge.
Put another way, more frugal approaches to AI can advance innovation - as has been repeatedly demonstrated by China, most notably its now famous Deepseek model.
All about the data
There is substantial hype about AI models, super-intelligence and the singularity moment when AI will self-recursively improve itself without human intervention. This is mostly entertaining pulp. A salient point to note is that current AI models are “sentence completers” and it is unlikely that they themselves will discover something truly radical. Yes, they may help in discovering potential new molecules for pharmaceuticals but they will not usher a paradigm shifting moment in medicine.
Organisations (and countries) should aggressively hold on to their data but run small pilots to explore how internal data can be married with open source AI models running on relatively inexpensive hardware. To use a smartphone analogy, we don’t need the latest iPhone 17 to experiment with most internal tasks. Being three generations behind is more than enough.
Some academic institutions, including Qatar’s Hamad Bin Khalifa University, are using hardware acquired in 2018 to train small but effective AI models.
Consider the following less exciting but nonetheless important example. A local company is continuously plagued by corrosion problems whereby heat, humidity, and salt substantially reduces the lifespan of its infrastructure. Much of the company’s corrosion inspection activities are currently carried out manually. In response the company could start addressing problems by collecting images of corrosion surfaces and annotating them with the help of internal experts. Using the data, it could make use of a lightweight open source AI model to classify corrosion type, trained on a low-end graphics processing unit (GPU) workstation by one AI engineer.
The cost breakdown in terms of data, personnel and hardware will be of the order of 10:4:1, that is, for every riyal spent on hardware, the cost of personnel will be four riyals and data will be 10. The productivity gain could be on the order of 5x, and the long-term impact on reducing climate emissions could be substantial. All of this can now be done without hype while accumulating tangible gains.
The lessons are plain to see. Identify a task for AI to work on, collect high quality data, install a low-end GPU option, and re-train local IT staff to run AI models. It is really that simple!
* The article first published in The New Arab, in 10 December 2025. Opinions expressed in this article remain those of the author and do not necessarily represent those of The Levant Files.