The Municipal Corporation as an AI Observatory

New intelligent Micro, Small, and Medium Enterprises (MSMEs) built on Small Language Models (SLMs) are emerging during the infosphere revolution. Large Language Models (LLMs) have lit the slow fuse of imaginations. Training LLMs require arrays of Graphics Processing Units. These GPUs provide the parallel processing power needed to handle the huge datasets these models learn from.

As we build more efficient digital architectures, develop advanced hardware, and co-design better training techniques, the gap between SLMs and LLMs is narrowing swiftly. This narrowing opens vast opportunities for countries in the Global South. The disruptions of AI assemblages will be immense and need to be scrutinized and interrogated among the poorer strata of society worldwide.

SLMs offer fresh and exciting applications that further democratize AI and its potential to impact lives and livelihoods. The AI path across Latin America and the Caribbean (LAC) may involve small and medium language models, operating in controlled environments such as domain-specific SLMs. SLM-powered AI may become critical for MSMEs across LAC. So what exactly are SLMs?

Simply put, they are language models trained on specific data sets to produce tailored outputs. All the data in the data lake is kept within the firewall domain. This means external SLMs are not trained on data that can cause harm or adverse consequences if disclosed, misused, or accessed without authorization.

SLMs scale computing power and energy consumption to the project’s objectives. This reduces environmental impacts and helps to lower ongoing expenses. SLMs have fewer parameters from millions upwards. LLMs on the other hand have trillions. This difference in size converts to several gains. The advantages can be arranged into three clusters.

Effectiveness: SLMs need less computational power and memory. This makes them ideal for deployment on smaller devices or even edge computing situations, creating opportunities for on-device chatbots and personalized mobile aides.

Openness: With lower resource requirements, SLMs are more accessible to a diverse cadre of developers and government agencies. This democratizes AI, allowing niche teams and entrepreneurs in the Global South to explore the models without significant infrastructure investment or state subsidies.

Fine Tuning: SLMs are not intricate and easier to fine-tune for specific domains and tasks. This empowers developers in LAC to build models tailored to niche applications, leading to higher performance, specificity, and accuracy. In LAC the AI discourse is thin. The imposed Eurocentric methods of using theoretical categories have barricaded any recognition of the historical specificity of distinct social phenomena, and their determinations in the region. The new capabilities of all-pervading LLMs to code our lives are not innocent. They transport many of the prejudgements and prejudices of the European Enlightenment.

SLMs are trainable with relatively small datasets. Their basic designs heighten interpretability, and their compressed size simplifies deployment on mobile devices. A noteworthy benefit of SLMs is their fitness to process data locally. This makes them useful for Internet of Things (IoT) edge devices and enterprises bound by stringent security protocols.

However, deploying small language models involves a trade-off. Due to their training on discrete datasets, SLMs have narrower knowledge bases than LLMs. Furthermore, their understanding of language and context can be restricted. The result is less accurate and nuanced responses compared to LLMs.

Despite this limitation, SLMs are now core to smart city initiatives like Connected Arima. During the second session of the United Nations Habitat Assembly in June 2023 in Kenya, one hundred and ninety-three countries requested UN-Habitat to develop international guidelines on people-centred smart cities through Resolution HSP/HA.2/Res.1.

The recommendations will be a non-binding framework to guide the drafting of regional, national, and domestic smart city protocols, conventions, plans, and approaches. This will ensure that digital peri-urban infrastructure can contribute to making cities open, welcoming, prosperous, and respectful of human rights. The guidelines are expected to be finalized in 2025.

In ground-breaking projects like Connected Arima, Municipal Corporations become a hive of AI activities with a portfolio of fascinating projects around SLMs. Early-stage activities may involve understanding SLMs and the role of AI in urban revitalization and development as SLMs become part of the tool kit for intelligent city innovation. SMLs can be used for peri-urban data refining and combining to create new data for data-driven urban planning.

At the level of local government, this can support citizen-facing digital service delivery using SLM insights. The integration of SLMs into traffic and transportation management, and its use for environmental monitoring are fresh frontiers in intelligent city design. These SLM projects increase public engagement and buoy a participatory urban planning model. Learning analytics provide feedback mechanisms for continuous improvement.

All of this must be within a legal framework that embraces data privacy, data protection, and the ethical use of AI in urban revitalization. AI-driven urban planners must understand the scalability and maintenance of SLM systems by making the Borough Corporation an AI Observatory and an AI Hacker Space. In these laboratories, hands-on design and testing will allow development teams to make presentations on projects for group critique ahead of launching a minimum viable product that will evolve and grow with public participation. To do nothing is to never fail. We can learn from Intelligent Failure. We must fail small and recover fast.