Background & Context§
While the AI spotlight remains fixed on trillion-parameter large language models (LLMs) from tech giants, a quiet revolution is underway in parts of the world where internet connectivity is sporadic and data-center infrastructure is absent. Small AI—models with at most a few billion parameters that run entirely on devices like smartphones, drones, or Arduino boards—is delivering life-saving services in fields ranging from pharmaceutical authentication to mosquito detection. The July issue of IEEE Spectrum highlights efforts by researchers and entrepreneurs who argue that this ‘small AI’ trend is not just promising but essential for global equity. As World Bank President Ajay Banga noted at Davos in January, most discussions around AI assume access to massive computing power and data, yet outside developed nations—except maybe India and China—few countries have that combination.
The News: What Happened Exactly§
In 2019, Nigerian entrepreneur Adebayo Alonge faced a crisis during a product demonstration in Cape Town. His startup RxAll’s handheld spectrometer, the RxScanner, scans pills with infrared light to verify their authenticity, sending molecular data to an AI model hosted in the United States. However, with the server 14,000 kilometers away and limited bandwidth, a single scan took over five minutes. "I was shocked," Alonge recalls. He instructed his engineers to shrink the AI model to run locally on his Android phone, which they accomplished in two hours. This emergency patch evolved into a new product: a fully offline RxScanner that authenticates pills in places without broadband, computers, or even reliable electricity. The incident turned Alonge into an advocate for small AI, which he now believes will eventually touch more lives than giant models due to its sustainability and accessibility.
Separately, researchers in Brazil, Uruguay, and India have deployed small AI for critical tasks. At the Federal University of Itajubá, professor Marcelo José Rovai and his team developed TinyML models that generate electrocardiograms using an Arduino device, enabling cardiac diagnostics in remote Brazilian regions lacking complex equipment. Another of their projects detects ant infestations in a Uruguayan vineyard using a low-power device, while a third identifies breeding sites of malaria-carrying mosquitoes. In India, Bala Murugan at Vellore Institute of Technology created a drone-based system that captures images of cashew plants and identifies disease splotches—all processing occurs on the drone itself, eliminating the need for an on-site computer or cloud connection. These examples illustrate the diverse applications of small AI, which the World Bank actively promotes through grants, mentorship, and policy advice in countries like Rwanda.
The technical feasibility of small AI has grown due to hardware improvements and model compression techniques. In 2025, over a third of smartphones shipped globally can run generative AI, a figure expected to reach 45% by end of 2025 and over 50% by 2026 (Counterpoint research). Meanwhile, open-weight models like Google DeepMind’s Gemma 4 and Alibaba’s Qwen 3.5 allow developers to fine-tune them for specific industries, such as dairy farming. Rovai demonstrated this on a Zoom call, showing an Arduino UNO Q ($50 device with Qualcomm chipset) running a language model that analyzes sensor data to detect mosquito breeding pools, consuming only 3 watts. Such advancements make it possible to deploy AI where previously only simple logic sufficed.
Historical Parallels & Similar Incidents§
This migration of intelligence from the cloud to the edge echoes a pivotal shift in computing: the rise of mobile computing in late 2000s. Back then, smartphones gradually replaced feature phones as processors became powerful enough to run apps locally. Companies like Nokia initially attempted to offload processing to servers (the ill-fated N-Gage platform), but the industry quickly learned that local execution was essential for usability in areas with patchy connectivity. Apple’s iPhone, with its focus on local apps, catalyzed a paradigm where many tasks moved from servers to handsets. Similarly, today’s small AI leverages smartphone neural processing units (NPUs) to run inference offline—a strategy that parallels the earlier hardware evolution that enabled apps like offline GPS navigation.
A more direct analogy can be found in the history of machine learning in resource-constrained settings. In 2016, researchers at Google introduced TensorFlow Lite, a lightweight framework for on-device inference, explicitly targeting applications where latency and connectivity were bottlenecks. One early use case was the ‘Google Lens’ feature on Pixel phones, which performed object recognition locally. However, the breakthrough for truly small AI came with the development of knowledge distillation—a technique where a smaller ‘student’ model is trained to mimic a larger ‘teacher’ model. This method, proposed by Geoffrey Hinton et al. in 2015, allowed creating compact models that retain high accuracy for specific tasks. The current wave of small AI builds on this foundation, but the critical difference is the expansion from ‘mostly offline’ to ‘completely offline’ operation, as demonstrated by Alonge’s RxScanner running without any server connection. Whereas earlier edge AI often relied on periodic syncing, today’s small AI can operate for days or weeks independently, accepting updates only when convenient.
The lessons from these parallels are clear: just as mobile computing thrived on local apps, AI adoption in underserved regions will depend on local models. The pain point of latency and the cost of cloud connectivity force a design philosophy where the model fits the device, not vice versa. Alonge sums it up succinctly: "I think the future of AI is not like one giant model, at a center. I think it’s millions of small, precise models deployed at the edge."
Formatting§
See content above.