The Barcelona Supercomputing Center - Centro Nacional de Supercomputación (BSC-CNS), a public research center located within the Polytechnic University of Catalonia (UPC) in the western part of the city, occupies the land within the beautifully preserved Torre Girona estate, established in the 19th century, surrounded by lush green spaces and a pond. This setting creates a juxtaposition with the site’s history as a thriving hub for trade, industry, and banking.

Since 2005, the BSC-CNS has housed the MareNostrum (Latin for “Our Sea”) supercomputer. In tune with the city’s reputation for stunning architecture, the MareNostrum 4 supercomputer housed within the chapel with its neo-classical architecture built in the 1940s was chosen as the most beautiful data centre in the world in 2017.

MareNostrum 5: A Supercomputing Powerhouse

Now in its fifth iteration, MareNostrum 5 was inaugurated in December 2023 and represents the pinnacle of high-performance computing in Spain. As part of the European High Performance Computing Joint Undertaking’s (EuroHPC JU) family of supercomputers, it plays a crucial role in advancing scientific research and achieving technological sovereignty for EU’s member states.

Occupying a space roughly the size of three tennis courts, it boasts a computing capacity equivalent to roughly 380,000 high-end laptops - this translates to an impressive 314,000 billion calculations per second. Its unique architecture combines classical computing with an accelerated partition specifically designed for artificial intelligence.

As of June 2024, MareNostrum 5 ranks among the world’s top supercomputers, holding the 8th position on the TOP500 list. And in addition to raw power, it’s also one of the greenest supercomputers - the BSC-CNS has implemented efficient cooling systems, renewable energy sources, and energy optimization strategies to minimize its environmental impact that ranks it #15 on the complementary GREEN500 list.

MareNostrum 5 plays a crucial role in Spain's national AI strategy, presented in 2020, driving advancements in fields like AI surrogates, digital twins, foundational models and LLMs, and image processing. These advancements have far-reaching implications for everything from drug development and medical research to climate modelling and urban planning, helping researchers tackle some of the world's most complex challenges.

Enabling Urban Innovation

The BSC-CNS plays a vital role in Barcelona’s city advancement, collaborating closely with the Barcelona Digital City initiative and its urban innovation arm.

Here are some examples of how the city has been helping shape the future of urban living:

• Digital Twins of the City: Barcelona’s virtual replica, the first of its kind in Europe, marks an urban planning revolution. It allows simulations of various scenarios, modeling the impact of potential changes, like the installation of a new metro line or the effects of a heatwave on vulnerable populations. These digital twins are further enhanced by AI, powered by the BSC-CNS, enabling real-time tracking of air and noise pollution, distribution of goods, traffic management, and building energy use to better understand an incredibly complex urban system.

• CALIOPE-Urban: This innovative air quality modelling tool, powered by the BSC-CNS's supercomputing capabilities, is designed to improve air quality management in urban areas by obtaining hourly maps of NO₂ concentrations at the street level. By analyzing vast amounts of data, CALIOPE can pinpoint pollution hotspots and help policymakers implement targeted measures to improve air quality, which can significantly reduce respiratory illnesses and enhance citizens' overall health and well-being.

• Sentilo: An open-source software that utilizes thousands of sensors installed in public spaces throughout the city to capture real-time data on various aspects, such as the flow of people, noise levels, temperature, and air quality in each neighbourhood. This data provides valuable insights into the city's dynamics, supplementing the information used to power AI-accelerated digital twins, and enabling more effective urban planning and resource management.

• Mercè: This citizen science experiment conducted from May to November 2020, demonstrates the potential of combining human insights with AI. By involving citizens in training a machine learning algorithm to analyze evaluations of over 3000 streets, Mercè helps identify urban design patterns that contribute to a more livable city, such as the presence of trees, wide sidewalks, or urban furniture.

Evolving Digital Awareness in Urban Environments

As urban environments grow more complex, the demand for intelligent solutions is greater than ever. Urban AI is an emerging concept that’s an extension of the Smart City paradigm, helping cities navigate the challenges of the 21st century, and Barcelona is at the forefront of this revolution.

On July 20th, 2022, Barcelona, along with Amsterdam and London, presented the Global Observatory of Urban AI (GOUAI) at the Barcelona Urban Innovation Center. The GOUAI's Atlas of Urban AI serves as a global platform, providing a visual representation of how cities worldwide are using AI to address a wide range of challenges.

This initiative builds upon the Cities Coalition for Digital Rights (CCDR), also launched by Barcelona, along with Amsterdam and New York City in November 2018. The CCDR recognizes the growing importance of data in city operations and emphasizes the need for digital rights, ethical considerations, and secure infrastructure in the development and implementation of urban AI solutions.

The Emergence of Urban AI

The GOUAI’s first report, released in February 2024, provides valuable insights and highlights key challenges facing the development of Urban AI:

• Urban AI has distinct political influence: Cities are complex ecosystems and prime testing grounds where AI interacts directly with millions of people’s lives, making them inherently political spaces. This reflects the origin of the word “politics” itself, which derives from the Greek word “politika,” meaning “affairs of the cities.” The abundance of data from concentrated human activity and advanced infrastructure in urban environments fuels machine learning initiatives, such as autonomous vehicles, robotics, city brains, and urban software agents.
  
  However, to fully address the societal and environmental challenges faced by urban populations, Urban AI must prioritize a human-centric approach that goes beyond technological optimization and innovation. This means grounding its development in social and ethical considerations that recognize the broader impact on quality of life and inclusivity. Currently, the majority of Urban AI initiatives focus on governance and improving urban services.

• Ethical Considerations and Cultural Differences: The current map reveals an imbalance in Urban AI initiatives, with 80% concentrated in North America and Europe. This disparity likely reflects varying levels of political and fiscal autonomy among municipal governments, influencing their ability to promote open data and AI projects. Geographical barriers, language differences, and cultural disparities can create challenges for global collaboration and the development of inclusive ethical AI frameworks.
  
  Addressing these disparities will be key to fostering international collaboration and inclusivity in AI-driven urban development. Developing adaptable ethical frameworks that respect regional diversity can help Urban AI initiatives bridge divides, ensuring equitable and effective deployment of these technologies.

• Balancing Security with Individual Data Privacy: While cities strive for security and digital resilience prompted by increased data accessibility, ethical considerations around data privacy make responsible deployment challenging - the GOUAI deliberately excludes the many controversial uses of facial recognition technologies worldwide.
  
  For example, Latin America’s use of facial recognition technology, while seen as a solution to rising crime rates, raises concerns about its potential for abuse and discrimination. Implementing robust safeguards that find the right balance between AI-enhanced utility and individual rights is crucial.

• Sustainability needs more attention: While many implementations are thinking about transparency, privacy protection, fairness, digital safety and accountability, only 6% of Urban AI projects prioritize sustainability, highlighting a gap between the potential of AI to address environmental challenges and the recognition of its own environmental footprint.
  
  Cities must recognize that AI can be both a contributor to and a solution for environmental problems, considering the energy consumption, e-waste generation, and resource extraction associated with AI technologies. A commitment to sustainability could also set a valuable precedent for the broader digital transformation of cities.

• Implementation is outpacing regulation: Urban AI solutions are rapidly emerging and encouragingly, implementation is not exclusive to large cities, with 66% of initiatives occurring in small to medium-sized cities. Cross-city collaborations, like Eurocities and C40 Cities, are instrumental for knowledge sharing and resource pooling.
  
  However, there’s little due-diligence follow up on policies (12%) and strategies (6%) - only 11% of cities have specific AI strategies or action plans (Barcelona, Amsterdam, NYC, Buenos Aires, and Vienna), and just 21% have implemented local policies and regulations. This regulatory gap indicates that as Urban AI solutions are deployed at a rapid pace, cities may struggle to keep up with necessary legal frameworks and safeguards. Closing this governance gap will be crucial to achieving the full potential of Urban AI while managing risks.
  

  Subscribe now

These are all part of the Superblock model, which is redefining urban planning in the city. This people-first approach supports sustainable public transportation and reorganizes public space, resulting in reduced noise pollution, improved air quality, and increased social interaction.

How exactly do they work? Essentially, a set of streets are closed off around a set of blocks (typically 3 by 3) from the existing infrastructure, such that car traffic is diverted around it, or in some cases reduced to speeds as low as 10km per hour. The inner streets are then reclaimed and renewed, transformed into an island of pedestrian-priority zones consisting of elements like greenery, parks, street furniture, playgrounds, and room for various local community use.

Barcelona's Urban Legacy

Barcelona is renowned for its innovative approach to urban planning. The Superblock concept is the latest evolution in this tradition, building on a foundation laid in the 19th century by Ildefons Cerdà. Cerdà, a pioneering Spanish urban planner, coined the term "urbanization" in his 1867 "General Theory of Urbanization," which detailed the science of making cities and strived to balance the best parts of rural and urban living.

His design for the Eixample district (Catalan for 'Expansion') served as a model for future urban development; over a century later, Salvador Rueda, head of Barcelona’s Urban Ecology Agency, expanded on the blueprint of the city 'block' to create the Superblock model.

Although Superblock planning and small pilot projects started in the 1980’s, the project began in earnest in 2016 in the neighbourhood of Poblenou. This initiative was motivated by the need to address several interconnected urban challenges, driven by the city’s overreliance on cars.

This car dependency contributed to an overcrowded city centre, excessive air pollution exceeding EU limits, high noise levels, and a lack of green spaces that exacerbated urban heat island effects. These environmental and health concerns are especially urgent now, as recent years have seen back-to-back record-breaking summer temperatures (2022, 2023, 2024).

Balancing Progress and Concerns

Implementing Superblocks has not been without its difficulties. The initiative had faced pushback from community members worried about a range of potential consequences, including negative impacts on local businesses due to reduced car traffic, disruptions to public transit routes, and the possibility of increased social inequalities. This resistance is especially pertinent to Barcelona with its strong anti-tourist sentiment and gentrification concerns more recently.

In the political arena, the implementation had been framed by some as a radical left-wing policy championed by the former mayor, Ada Colau, who faced a series of legal challenges from prominent retailers and stakeholders in the tourism industry seeking to reverse pedestrianization efforts. The transition to the new mayor, Jaume Collboni, who took over in 2023, could mark a period of uncertainty. A more reserved approach to Superblock expansion could be taken for the sake of political posturing, potentially affecting the pace and extent of future implementation.

Despite these hurdles, Barcelona remains committed to its ambitious goal of installing 503 Superblocks by 2030 as significant benefits are realized by the community, where a 24% reduction in air pollution levels is estimated to prevent 700 premature deaths and generate economic savings of 1.7 billion euros each year.

A Global Inspiration: Superblocks Reshaping Cities Worldwide

Barcelona's Superblocks have become a source of inspiration for cities worldwide seeking to create more sustainable and resilient urban environments to address current and future climate change issues. In 2022, over 250 cities sought information about implementing Superblocks, demonstrating a growing universal interest in this innovative urban solution for a common vision of the future - greener, less polluted, and more livable cities.

Architects and city planners are praising the model for its flexibility, noting its potential to be integrated into both grid-oriented cities and those with more unique layouts. Some notable examples around the world include:

• Vitoria-Gasteiz, Spain: The capital city of the Basque Country in northern Spain, not too far from Barcelona, is one of many Spanish cities implementing “Super-manzanas” (Spanish for “Super Apples”), complemented by a comprehensive Green Urban Infrastructure strategy to prepare for increased precipitation and flood risks.

• Berlin, Germany: Directly inspired by Superblocks, the city has over 50 ‘Kiezblocks’ (German for “Neighbourhood Blocks”) initiatives, each customised in their implementation to address increasing amounts of traffic that are contributing to high noise and air pollution, along with the dangers for children and the elderly.

• Los Angeles, USA: One of the originators of car-dependent urban design in the 20th century, is now embracing Superblock principles, a particularly noteworthy shift. The “Park Block” program was introduced in 2023 as a way to tackle poor air quality and particularly high pedestrian casualties represented by over 312 people killed in traffic collisions in 2022.

• Buenos Aires, Argentina: This capital city has been aiming to reduce its carbon footprint and increasing its citizen quality of life through prioritizing public transportation, cycling, and pedestrian walking - today, a majority of its downtown core is car-free consisting of “Macromanzanas”, inspired by Superblocks.

• Full YouTube Recording (3.5 hours)

• List of Resources

Attracting a global audience of over 600 people from over 70 countries, the event highlighted a universal desire to harness the potential of AI agents - autonomous systems that can perceive, reason, and act towards goals creatively and flexibly within defined boundaries for various applications across work and life.

The session dove deep into the key issues facing AI agent development today, highlighting six main areas:

1. Infrastructure

2. Cost

3. Reliability

4. Privacy

5. Safety/Security

6. UI

I learned a lot from the session and ensuing conversations; here’s my summary of takeaways for navigating these six challenges, which are essential for unlocking the full potential of AI agents and ensuring their responsible development and deployment:

1 - Infrastructure

Agent Architecture: Defining and Programming: Given the complexity of understanding even a single LLM's output, configuring Agents demands systems thinking and extensive programming. Multiple LLM calls are involved, along with prompts, compute optimizations, and tools organized in a graph framework, which can quickly get complicated, especially for any non-trivial tasks. Much like guiding a competent human, clear communication and feedback are vital - the more of the Agent that we can explicitly define, like incorporating expected queries and answers into examples, the better we can maximize its output potential.

Evaluating Agent Performance: Evaluations have become a greater focus in LLM development, and are even more elusive for agent performance that chains LLM outputs to autonomously make rational decisions and actions. The history of evaluations spans from text-based games in 2015, 2020 to practical applications like WebShop in 2022, which captured basic agent actions in making online purchases. Recent and impressive benchmarks like SWE-bench and TAU-bench offer more thorough assessments of agent capabilities for software development and practical applications that use tools.

The Importance of Benchmarks: Benchmarks provide standardized measurements for LLM output and Agent quality across various domains. However, their realism and applicability to subjective topics and business-critical applications (like law and medicine) demand constant improvement as they become increasingly capable. Model developers should be also aware of possible “overfiting” with highly publicized benchmarks that can reduce their effectiveness in real-world scenarios at the expense of vanity metrics.

2 - Cost

Cost Optimization: Beyond Accuracy: Knowing how computationally expensive it is to run frontier models, the industry is shifting focus beyond accuracy and output quality towards the costs for inference, especially for complex tasks and large-scale deployments. Smaller models can be surprisingly effective and may be the best solution for resource-constrained users.

The Power of Simpler Models: Smaller models, with careful tuning and repeated sampling, can sometimes outperform even state-of-the-art models (when zero-shot prompted). Techniques like shorter prompts, varied model temperatures, and better few-shot examples can help ‘warm up’ a model to reach its full potential.

Repeated Inference for Higher Quality: OpenAI’s new ‘Strawberry (o1)’ model is making waves in the industry for taking on this new approach towards inference to extract higher quality outputs, touted as “advanced reasoning abilities”. By baking in established prompt engineering methods like ReAct, reinforcement learning, and Chain of Thought (CoT), the model can self-reflect and think step by step, in alignment with a safe set of human values, goals, and expectations. 

Inference is getting much cheaper: Encouragingly, LLM inference costs continue to rapidly decrease, similar to the data costs of wireless telecoms, coming down by 100x over 2 years. This is essential for enabling effective agentic workflows, as inference calls can drastically increase to support systematic reflection, reasononing, and iteration of outputs.

3 - Reliability

Ensuring Reliable Agent Behaviour: The inherent variability of LLM outputs makes it challenging to ensure consistent and reliable agent behaviour. Reinforcing the points on Infrastructure, programming more of the architecture instead of relying solely on natural language prompting can improve consistency and reproducibility, vital for any business.

100% Accuracy Is Not Realistic: Maximizing accuracy can lead to unbounded cost. It’s practically impossible to achieve 100% accuracy, and the cost of additional inference calls rises exponentially the closer you try to get to perfect (ex. 80% —> 90% could cost the same as 95% —> 96%). Acknowledging this limitation, a future where humans remain involved in key-decision making processes, supported by powerful agents that offload much of the work, seems to be the way.

From ML to LLM Reliability: ML reliability relies on deterministic and rules-based measurement that follows explicit instructions for any kind of input. In contrast, LLMs are more nuanced, requiring a deeper understanding of human intent and preferences, often involving subjective judgments or “vibe checks”. For example, there’s no perfect answer for “Write a creative email for an outbound client lead.”

4 - Privacy

Protecting User Privacy: As agents become more helpful and handle increasingly sensitive data, it’s imperative to maintain user trust and prioritize ethical AI development. Well-designed agents must incorporate state-of-the-art privacy-preserving technologies that can make use of sensitive data with minimal compromise.

Navigating Regulatory Compliance: Adhering to evolving data privacy regulations like the EU AI Act and California Consumer Privacy Act (CCPA) presents obstacles and important guardrails for development. At the same time, AI agents can also be instrumental in the regulatory compliance process itself by analyzing data to enhance risk assessment and management processes.

5 - Safety/Security

The AI Safety Institute’s Role: Established by the UK government, the institute functions like a start-up, focusing on developing safety evaluations, addressing misuse, societal impacts, and risks associated with autonomous systems.

The AI Risk Repository: A Comprehensive Resource: While not explicitly mentioned during the sessions, this comprehensive database introduced in August by MIT researchers catalogs over 700 AI risks, providing an important, common frame of reference for understanding and addressing potential threats from AI. Google’s DeepMind has been actively involved in AI risk research, releasing reports like “The Ethics of Advanced AI Assistants (2024)”.

Ensuring AI Safety: Just like accuracy, achieving 100% safety in critical scenarios is near impossible, necessitating continuous due diligence and risk mitigation efforts for the one of the most pressing world problems. In a future with potentially millions or even billions of agents, evaluation systems must evolve to keep pace with progressively greater ranges of freedom and limited supervision of these systems. They can easily be misused by malicious actors seeking to spread misinformation or conduct cyberattacks.

6 - User Interface (UI)

Human-Centred Design: Recognizing the impracticality of achieving 100% autonomous and reliable agents, more focus is needed on increased transparency of an agent’s capabilities, limitations and decision-making processes to foster trust and understanding. Human input will be essential to verify and address those final percentages for accuracy and safety that make applications in high-stakes scenarios viable.

Creating Intuitive Interactions: Creating user experiences that allow for natural communication and interaction with AI agents has lots of potential. Advancements in natural language processing and innovative use of multi-modal models that can process voice audio and video will bring us closer to realizing the personal assistants we see in science fiction.

Continuous Learning and User Feedback: Extending the established concept of “Human in the Loop”, humans will play a crucial role in teaching agents to improve by providing valuable feedback, especially as they’ll inevitably run into unforeseen situations. We should also be mindful of when to incorporate human feedback - do we always know the best action to take? When should agents delegate to humans?

Emerging AI Agents in the Market

We’re seeing a surge in the launch of AI assistants and agent-like products that go beyond simple Q&A chatbots, performing increasingly sophisticated and useful tasks.

Examples from major tech companies are enabling versatile and more personalized functions on top of their commercial LLM offerings like scheduling appointments, data analysis, and generating creative content. These include:

• OpenAI’s Assistants (November 2023)

• Anthropic Claude’s Artifacts (August 2024)

• Google Gemini’s Gems (August 2024)

We’re also seeing a push toward novel applications that push the limits of AI capabilities in groundbreaking ways, starting to take on the whole workload of software engineers:

• Cognition AI‘s ‘Devin’ (March 2024): This full-fledged developer agent goes beyond coding to replicate the complete work scope of a traditional software engineer, including designing, testing, and deploying software.

• Cosine’s ‘Genie’ (August 2024): This innovative AI coding assistant is trained on data that mimics the cognitive processes, logic, and workflow of human engineers, making it a truly collaborative coding partner.

In Toronto, the Gardiner Expressway, constructed from 1958 to 1966, exemplifies this trend, echoing similar developments in major cities like Seattle, San Francisco, Seoul, and Montreal that prioritized car-centric mobility and efficient industry throughout the mid-20th century.

The city’s waterfront has historically thrived as an industrial hub, with various factories and shipping facilities leveraging their proximity to water for transport - this offered practical economic benefits, fueling rapid urban growth and establishing transportation networks that efficiently connected suburbs to city centres. They also captured the spirit of modernity and freedom associated with cruising along expansive motorways.

The Gardiner Expressway Strategic Rehabilitation Plan

Recognizing the challenges posed by aging urban highways for rapidly growing populations, some cities have opted to tear down or repurpose them.

In contrast, Toronto has committed to a major rehabilitation of the Gardiner Expressway since 2016. After over 50 years of service, this core piece of the city’s transportation infrastructure has weathered progressive deterioration, necessitating extensive repairs. It currently handles approximately 140,000 vehicles daily.

Faced with mounting upkeep expenses and failed attempts to implement tolls, the city sought relief through a "New Deal" forged with the provincial government at the end of 2023, which will ensure the Expressway’s long-term existence. Due diligence will be conducted by end of this year to facilitate the transfer, including an assessment of rights-of-way, airspace, and underground ownership, along with legal and corridor management controls.

"By uploading the Gardiner and DVP (Don Valley Parkway), the city will be able to spend billions more on affordable housing, fixing transit, and building communities," Mayor Olivia Chow stated.

However, the deal comes with trade-offs. The city has agreed to relinquish some control over the province’s redevelopment plan for Ontario Place, a move that has sparked controversy, particularly regarding the proposed Therme spa on prime waterfront land originally intended for a public park.

And as the Ministry of Transportation will now have jurisdiction over the spaces surrounding the highways, concerns have arisen for Toronto’s autonomy in shaping its urban landscape and pursuing sustainability goals, particularly given the province's prioritization of car mobility.

Toronto's World-Leading Traffic

The rehabilitation project, currently in its second of six phases and coupled with increasing return-to-office mandates, has exacerbated Toronto's already notorious traffic congestion:

• Worst congestion in North America: A recent report confirmed that the city has the worst traffic congestion in North America, trailing only behind London and Dublin in travel time.

• Morning rush hour travel times increased by 250%: Commercial drivers are spending more time on the Gardiner, greatly impacting emissions and business productivity, with lengthy commutes costing an estimated $11 billion in lost productivity in 2022.

• Public buses caught in the traffic jam: GO bus users are experiencing immense frustration, stuck in traffic for up to 45 minutes within walking distance of Union Station Bus Terminal.

• Traffic will continue to slow down, below 20km: Based on the provincial government’s estimates, average driving speeds on local highways across the Greater Toronto Area are projected to drop below 20km/hr, even if new highways like Highway 413 are built.

While a $73 million stimulus funding in response aims to accelerate project completion by a full year, from April 2027 to April 2026, it remains to be seen if these efforts will alleviate the growing congestion and its cascading effects.

Creating a Public Realm Under the Highway

Acknowledging the complexity of urban mobility challenges, city planners are not only focusing on the expressway itself but are also turning their attention to the space beneath it.

Vehicles are also efficiently connected via Lakeshore Boulevard, yet increasing gridlock on this additional surface artery presents a unique and urgent opportunity. The underutilized space beneath the Gardiner can be transformed into inspiring community spaces, fostering a more connected and cohesive public realm.

The city’s official Under Gardiner Public Realm Plan envisions converting the 7-kilometre stretch beneath the elevated expressway into a vibrant, multimodal corridor that values pedestrian and cyclist safety, accessible amenities, cohesive identity, and productive natural components to offset the Gardiner’s increased emissions and environmental impact, while also improving air quality and reducing the urban heat island effect.

This ambitious revitalization aligns with the objectives of the "Shoreline Stitch", a broader city initiative striving to reconnect downtown Toronto to its waterfront by addressing the physical and psychological barriers created by rail lines, the Gardiner Expressway, and Lakeshore Boulevard. The waterfront itself is also undergoing a significant transformation.

Realizing the Under Gardiner Public Realm Plan will be a multi-decade endeavour, demanding collaboration and thoughtful planning to align with existing City plans like Vision Zero, Complete Streets, and the 2040 net-zero emissions target.

By encouraging alternative modes of transportation like walking and cycling, this project can help achieve manageable traffic levels for a densifying population, crucial for unlocking the potential of innovative, hybrid infrastructure that can distinguish the city.

Today, Neural machine translation (NMT) is the industry standard, utilizing artificial neural networks to translate text, leveraging vast bilingual corpora and glossaries (i.e., text pairings) to learn patterns and produce accurate translations based on established rules.

Popular services like Google Translate, DeepL, and Microsoft Translator showcase the power of NMT.

# Illustrative bilingual text pairs (English to French) used in NMT training - x thousands, x millions, x billions
{"The cat is sleeping on the couch." : "Le chat dort sur le canapé."},
{"I love to eat sushi on weekends." : "J'adore manger des sushis le week-end."},
{"The Eiffel Tower is a famous landmark in Paris." : "La Tour Eiffel est un monument célèbre à Paris."}
{"Climate change is a global concern." : "Le changement climatique est une préoccupation mondiale."}
{"She plays the piano beautifully." : "Elle joue magnifiquement du piano."}
...

The Power of Context: LLMs Evolve Translation

Large Language Models (LLMs) are transforming the language processing landscape. They take a fundamentally different approach to translation, developing a broad understanding of language through unsupervised learning on massive, diverse datasets. LLMs identify patterns and relationships within the data itself, eliminating the need for explicit multilingual training or fine-tuning.

This contextual understanding allows them to interpret and 'learn' languages without explicit instruction or labeled data, and recent research suggests that the full potential of LLMs for translation remains to be fully explored and understood.

LLM Translation in Context: "He's in hot water" to French.

Context 1: A man is enjoying a hot spring.
LLM Output: "Il est dans l'eau chaude." (LLM interprets the phrase literally based on the context of relaxation)
(Literal translation: He is in hot water.)

Context 2: The employee made a mistake at work.
LLM Output: "Il est dans le pétrin." (LLM recognizes the idiomatic meaning of the phrase based on the context of a workplace error)
(Idiomatic translation: He's in trouble.)

However, the immense scale of data required to train foundation model LLMs, coupled with the computational demands of generating outputs (inference), can result in substantially higher operational costs. While their probabilistic nature enables them to produce more natural and relevant translations, it also increases the risk of generating hallucinations or inaccuracies.

The need for additional scrutiny and evaluation of these outputs adds further overhead, making LLMs a critical consideration in domains demanding precision, such as law, medicine, and other high-stakes content generation.

The future of AI-powered language conversion and the localization industry likely hinges on a synergy between NMT and LLMs, both poised for continued rapid progress. NMT's specialized strength lies in accurately translating factual text, providing a reliable foundation. Meanwhile, LLMs, with their nuanced understanding of context and language, offer flexibility and adaptability, particularly valuable for creative and culturally sensitive adaptations.

Navigating Cultural Nuance: Balancing Progress and Preservation

Many foundational models now boast impressive and ever-improving multilingual capabilities, as evidenced by established benchmarks like MMLU (Massive Multitask Language Understanding), but it's crucial to be mindful of their inherent biases.

The quality of translation is intrinsically linked to the volume and diversity of pre-training data. This often leads to a performance gap where high-resource languages like English, French, and Mandarin Chinese typically outperform low-resource languages, like Korean or various African languages.

Another recent research paper has illuminated how cultural stereotypes embedded in societal narratives can inadvertently seep into the rendering of text from one language to another, pooling together in a sense as we bridge the gap between high- and low-resource languages.

As the dominant global language, English often exerts undue influence on translations into low-resource languages like Hindi, potentially amplifying or even creating new stereotypes that extend beyond the original culture. These biases can perpetuate global power imbalances, harmful misconceptions and impede genuine cross-cultural understanding.

Localization in the Age of AI: Opportunities and Responsibilities

The responsibility to acknowledge and address potential biases in AI falls heavily on the shoulders of the scientific and scholarly publishing communities, among others. The optimal application of AI in this domain lies in augmenting human capabilities, positioning AI as a valuable co-pilot for translators and educators.

This collaborative approach ideally empowers professionals to dedicate their expertise to the nuanced, creative, and culturally sensitive aspects of translation that demand human intuition, emotional intelligence, and the more nuanced context that AI/LLMs can’t grasp for the foreseeable future.

Developing trustworthy AI, while navigating the intricate landscape of AI governance, is crucial for its successful global integration. These technologies hold tremendous promise for enriching global connections and fostering mutual understanding, playing a pivotal role in education and international collaboration - their implementation necessitates a thoughtful and balanced approach, ensuring that the pursuit of efficiency doesn't inadvertently erode cultural integrity or linguistic diversity.

Multilingual, Multimodal Realities

Our globally connected world strives to be even more seamless and culturally aware. We're already witnessing a glimpse of this future, with innovations like real-time visual language conversion through smartphone cameras and the emergence of AI-powered language learning apps.

Looking ahead, we can anticipate further advancements that will redefine how we interact with language:

• Real-time audio translation: The rise of the Translator Earbuds market and the prospect of Google’s AR Translate Glasses hint at a future of wearable AI where instant spoken translation during conversations is possible, dissolving language barriers and fostering effortless communication across cultures.

• Personalized language learning chatbots: AI-powered chatbots and established language learning apps like Duolingo and Rosetta Stone are already transforming how we learn languages, offering a variety of tailored experiences that adapt to individual needs and preferences.

• Run efficient translation models locally: The development of compact and efficient translation models, like Microsoft's phi-3-mini, enables offline and privacy-focused translation on personal devices, offering cost-effective solutions for individuals and small to medium-sized businesses.

• Interpret sign language and gestures: Projects like SignAll are developing AI that can translate sign language in real-time. Future systems will likely interpret cultural-specific gestures and enhance video call interpretations by capturing non-verbal cues, helping bridge communication gaps for the deaf and hard-of-hearing communities.

• Translate multimodal documents: Advanced AI tools like Smartling are translating complex documents while preserving formatting and adapting content for cultural relevance, a boon for global businesses.

Digital twins have traditionally been used in manufacturing and industrial processes to simulate complex physical systems, like airplane engines and factory floors. Now, advancements in remote sensing, IoT devices, cloud computing, and AI are enabling us to collect and analyze vast amounts of urban data, creating increasingly detailed and dynamic virtual replicas of entire cities.

What was once imagined by early simulation games like SimCity and Cities: Skylines, is now empowering urban planners, architects, and policymakers; digital twins are transforming how we understand and interact with our urban environments with a powerful perspective to test build scenarios, optimize infrastructure resilience, and improve overall citizen quality of life.

Digital Twins in Action: Barcelona & Ottawa

Barcelona is leading the charge in Europe with its pioneering digital twin. The Barcelona Supercomputing Center (BSC-CNS) is using this virtual replica in its first phase to experiment with hospital and metro station placements, as well as traffic strategies to improve air quality. The project has ambitious plans to ultimately optimize the design of public policies that can be applied to any city in the future.

In Canada, Ottawa showcases how digital twins are reshaping urban planning and management into a dynamic, data-driven process. By using enhanced 3D GIS visualizations, the city is striving to make complex spatial data more accessible and relatable to both the public and decision-makers.

This increased transparency offers the predictability and clarity needed to navigate large-scale infrastructure projects with greater efficiency and confidence. Additionally, routine open communication with the public about design proposals, zoning regulations, and potential construction impacts paves the way for smoother project execution and community buy-in.

Converging Physical and Digital Realities

The potential of digital twins extends far beyond urban planning. Enhanced real-time modeling and simulations, fueled by AI’s progress, promise to revolutionize various sectors, from optimizing traffic flow and energy systems to improving disaster preparedness and fostering citizen engagement. The AI for Energy Report 2024 from the US Department of Energy highlights AI’s crucial role in modeling complex systems like power grids, essential for our transition to renewable energy sources.

Digital twins are part of a broader shift towards immersive technologies like Virtual Reality (VR), Augmented Reality (AR), and the metaverse, collectively known under ‘Extended Reality (XR)’—these technologies are changing how we experience and understand the world around us, blurring the lines between reality and the digital realm.

A further emerging concept of ‘virtual twins’ is marked by progressively advancing computational power to simulate and interact with hypothetical future scenarios, allowing us to more readily anticipate, plan and respond to potential challenges and opportunities faced by growing cities.

The Metaverse Meets Urban Planning in Saudi Arabia

Saudi Arabia's futuristic city, Neom, exemplifies this convergence. It's investing in a digital twin envisioned as a metaverse, growing in tandem with the physical city. Virtual visitors can participate in crowdsourcing the city's design, like customizing apartment colours, creating an interesting "human-centric" digital approach to urban development.

Collaborative virtual spaces offer transformative potential for public participation and city planning on a global scale. However, it also underscores the critical need to address concerns around trust and data privacy in large-scale data sharing. Robust safeguards, cohesive policies, and ongoing public dialogue are essential to ensure responsible and ethical use of individual and ever more diverse sets of public data.

Expanding my perspective to encompass urban planning and environmental sustainability, I'm inspired by the power of geospatial analysis and the advanced, rich visualizations that better describe the interconnected challenges and opportunities facing our communities and our planet.

Since the development of Geographic Information Systems (GIS) in 1969, remote sensing technologies like satellites have revolutionized our ability to observe and understand Earth, providing crucial data for confronting the urgent challenge of climate change.

Foundational Satellite Programs

Programs like ESA Sentinel-2, NASA Landsat, and NASA MODIS have been instrumental in offering open public data for comprehensive Earth observation for decades. This data covers most of the globe, enabling researchers, decision-makers, and the general public to access valuable insights about Earth.

Satellite imagery is characterized by three key dimensions:

• Spatial: The location of objects on Earth's surface using latitude, longitude, and altitude. The spatial resolution, typically measured in metres per pixel, determines the level of detail captured in the image. For example, Landsat 8 and Sentinel-2 have impressive resolution ranges from 10 to 60 metres. In contrast, commercial options can achieve even higher resolutions, down to 15 to 50 centimetres per pixel.

• Spectral: Captures the unique interaction of materials with different wavelengths of light, whether reflected or emitted. By using multiple spectral bands beyond the visible spectrum, we can 'see' the Earth in novel ways, revealing detailed information about the composition and properties of different surface materials.

• Temporal: Tracks how frequently a satellite revisits a specific point on Earth, allowing us to monitor changes over time, uncovering meaningful patterns in the progression of weather patterns, landscapes, and human activity. Sentinel-2 and Landsat 9 have a revisit time of about 5 and 8 days respectively, while MODIS, along with private, specialized satellites with wider swaths, can revisit most locations daily (and multiple times at that).

Embeddings for Geospatial Intelligence

AI is transforming how the geospatial industry processes, visualizes, and extracts predictive insights from satellite imagery. At the heart of this transformation are embeddings—mathematical representations of input data, like text, images and audio, that enable more efficient and context-aware processing by machine learning models.

This marks a paradigm shift from traditional pixel-based analysis, which focuses on individual pixels in isolation, towards a more holistic understanding of satellite imagery. Object-based image analysis (OBIA) was an important intermediate step in this direction, facilitating a grouping of pixels into meaningful objects that laid the foundation for understanding spatial context.

Embeddings further advance this approach, encoding data into hundreds or even thousands of dimensions to capture complex patterns and relationships far beyond the 2-3 dimensions that humans can intuitively grasp.

By leveraging the self-attention mechanism used in language models, AI models can understand context in Earth observation, just as they do in sentences:

• Words: The word "bank" in “I’m going to the bank after work” refers to a financial institution, whereas in “It’s nice to hang out by the river bank,” it describes the side of a river.

• Satellite Image: A tree in a city park might be recognized as part of an urban green space, while the same tree in a dense forest is interpreted as a component of a natural ecosystem. Improved classification through embeddings allows us to more accurately assess the rich dynamics of our built and natural environments.

Advancing Earth Observation

The contextual analysis of Earth observation data by AI models enables more accurate and faster object detection, change detection, land cover classification, and crop yield estimation.

In turn, enhanced geospatial analysis powered by open-source tools and data deepens our understanding of Earth and opens up exciting new possibilities for data-informed, sustainable solutions.

Clay, an emerging non-profit open-source project, is significantly contributing to this evolution. Their AI model provides a robust foundation for advanced satellite imagery analysis and integrates it with other data sources. Through their AI for Earth Challenge, they encourage further customization and fine-tuning for targeted applications.

Clay’s Executive Director envisions a future where Earth observation is as accessible as a Google search—imagine searching for "muddy waters", "flood prone areas", or "docked container ships" just as easily as we find restaurants or hotels today.

Discover Your World

The open availability of satellite geospatial data makes this vision both achievable and inclusive, inviting us all to explore our planet like never before.

Platforms such as NASA Worldview, Sentinel Hub EO Browser, and Microsoft's Planetary Computer, among others, offer essential tools for accessing and analyzing this data, supporting a wide range of innovative applications and research.

July 2024 was also officially recorded as Toronto’s rainiest month ever, which highlights the pressing need for more innovative solutions.

China's Pioneering Water Resilience

Faced with similar challenges, China’s unprecedented urbanization over the past few decades, which recently used as much concrete in two years as the US did in the entire 20th century, reached a breaking point in 2012 - a catastrophic rainstorm in Beijing killed 79 people and caused extensive damage, exposing the limitations of concrete-heavy urban development and inadequate drainage systems being built across the country.

The Chinese government responded to this crisis by adopting sponge city planning at a national level. This innovative approach prioritizes nature-based solutions such as permeable pavements, rain gardens, green roofs, bioswales, constructed wetlands, and urban forests.

By mimicking natural hydrological processes, these elements absorb, retain, filter, and slowly release rainwater, offering a more sustainable and resilient alternative to conventional “grey” infrastructure (networks of pipes, concrete channels, and pumps) that are proving insufficient to handle increased amounts of storm runoff.

While grey infrastructure has been vital for economic growth and urbanization, its reliance on fossil fuel-intensive construction and operation has exacerbated climate-related risks. Sponge cities offer a nature-based alternative that seeks to restore ecological balance and address these challenges more sustainably.

Kongjian Yu, the Chinese landscape architect who pioneered the sponge city concept, has been championing this transformative approach since the late 1990s. In a CBC interview, he described it as "a fundamental change... an evolutionary change of the relationship between man and nature.” China aims to have 80% of its urban areas absorb and utilize 70% of annual rainfall by 2030.

Sponge Cities & Integrated Water Management

Incorporating a ‘sponginess’ component into urban water management is necessary to mitigate the impacts of future floods and droughts, which are expected to intensify as the average global temperature continues to rise. This approach is being adopted as a proactive and practical measure for new urban developments around the globe, recognizing that retrofitting existing grey systems is often more challenging.

Every city is different, and tailoring local-specific solutions that engage local communities, utilize native plant species that promote biodiversity, and respond to the unique climate conditions of the region will further enhance the effectiveness of sponge city projects.

Equally important, they should be complemented with robust planning and disaster risk management strategies, coordinated across multiple levels of government, to minimize the loss of life, property damage, and disruptions to critical services.

Addressing Water Scarcity

The growing global freshwater crisis is urgently prompting a broader shift towards integrated urban water management (IUWM), a holistic approach that seeks to better align urban development across various sectors — land use, housing, energy, transportation, etc. — to ensure sustainable economic, social, and environmental outcomes, with equitable and safe drinking water being a significant challenge for developing urban areas with limited infrastructure.

Absorbent sponge city planning provides scalable, natural-based solutions for managing increasingly scarce freshwater resources, helping reimagine our urban landscapes as resilient, adaptive, and regenerative ecosystems.

Given this escalating energy demand, I’m motivated to become more carbon aware and energy efficient in my AI development. The "Carbon Aware Computing for Generative AI Developers" course offered by deeplearning.ai, in partnership with Google Cloud, provided valuable insights into the carbon emissions produced throughout the machine learning (ML) model lifecycle, outlining the main factors to consider:

• Hardware: Everything from GPUs, TPUs, and servers to RAM contributes to what’s called "embodied carbon". The complexity of global supply chains makes it difficult to accurately calculate the total emissions across Scope 1, 2 and 3 levels associated with the production and shipping of these components.

• Models: A wealth of specialized, smaller models offer greater energy efficiency. The Hugging Face LLM-Perf and ML.Energy leaderboards are valuable resources for comparing energy costs alongside a model’s performance metrics.

• Data Centre Location + Time of Day: Where and when you train models greatly impacts carbon emissions - this is measured by the grid’s carbon intensity (gCO₂eq/kWh), which represents the carbon emissions per kilowatt-hour of electricity and is determined by the local energy mix. Low-carbon energy sources like wind, solar, hydro, and nuclear are ideal.

Google is helping the industry adopt dynamic strategies like shifting computing workloads to different servers throughout the day to "follow the sun and wind," taking advantage of the times and locations where low-carbon energy is most abundant. This approach is part of their commitment to ‘boldly accelerating climate action with AI’, but their overall carbon footprint remains a significant challenge.

As well, Google Cloud's region picker offers an overview of the different grid carbon intensities for each of its 40 data centres around the world, using data from Electricity Maps.

Focusing on Inference: The Carbon Cost of Thinking

A recent paper, "Power Hungry Processing: Watts Driving the Cost of AI Deployment?", by developers at Hugging Face and Carnegie Mellon University reveals a wide variation in the carbon footprint of AI model "inference" - the process of using a trained model to make predictions or generate responses. This ongoing energy use is comparatively harder to measure than the significant, one-time energy cost of training.

The researchers tested 88 different models across 30 datasets and 10 task types, highlighting the stark differences in their computational and energy demands:

The authors emphasize that using “multi-purpose” models (i.e., LLMs) for discriminative tasks like classification is far more carbon-intensive than using task-specific models. While convenient for users and easy to deploy, their widespread use in digital products like web search and email can often be overkill, contributing to a substantial and unnecessary carbon footprint. Developers can significantly reduce energy consumption by using smaller, more specialized models when appropriate — from 3x less for content generation to 30x less for discriminative tasks.

Sasha Luccioni, one of the paper's authors and a member of the Hugging Face team, is advocating for Energy Star Ratings for AI models. Inspired by the familiar ratings system for appliances and devices, this would establish a standardized methodology for quantifying energy consumption and carbon emissions of ML models.

Prioritizing energy efficiency in AI development and incorporating carbon emissions as an essential metric in conjunction with performance, speed, and security is pragmatic, and I look forward to the industry adopting more sustainable practices and innovating in this critical area.

Canada’s north encompasses a vast 40% of the country’s landmass, yet it’s home to only approximately 114,000 people across the Northwest Territories, Yukon, and Nunavut. This translates to a remarkably low population density of roughly one person per 30 square kilometers, a testament to the region's extreme climate and limited daylight outside of summertime.

Built Environments in the Cold

Indigenous and Inuit populations have flourished in the region for millennia, adapting to its frigid, harsh conditions. However, the 20th century brought a wave of global industrialization and modern urban planning that sought to reshape these long-established communities, viewing the North as a testing ground for ambitious experiments. Often naively and negligently conducted, these experiments disregarded traditional knowledge and practices in an effort to conquer the harsh winter climate.

This approach coincided with a broader colonial government policy of assimilation, aimed at integrating Indigenous and Inuit populations into a modern, Western way of life. As a prominent example, the Canadian government commissioned British architect Ralph Erskine to design a model Arctic town at Resolute Bay, featuring novel cold-climate urban design.

The Concept of “Nordicity”

In the 1960s and 70s, Québec geographer Louis-Edmond Hamelin introduced the concept of "nordicity", which highlights the distinct polar values and conditions of northern communities. The extremes of the midnight sun and long, cold winter nights foster a unique appreciation for the functional value of place, shaped by shared community and cultural experiences across vast distances and changing seasons. The conventional ideas of 'city’ and ‘urban environment’ aren’t quite as relevant here.

Nunavut, established in 1999 as Canada’s largest territory, adopted a decentralized governance model due to its vast landmass, challenging Arctic and Sub-arctic conditions, and the desire for a less hierarchical, more accessible government for its dispersed populations.

However, this structure has presented practical and environmental challenges. Fossil fuels remain heavily relied upon for everyday power needs, making energy expensive; sustainable alternatives, like solar and wind power, remain difficult to implement due to reduced sunlight and extreme cold that strains infrastructure. Aviation is crucial for long-distance transportation, while informal snowmobile and ATV paths, independent of the formal road network limited by pervasive ground permafrost, create a vital seasonal transportation network essential for maintaining social connections and cultural traditions.

Defining Canada’s North

The Arctic regions are among the most vulnerable to rising global temperatures. The thawing of permafrost is reshaping the landscape and deteriorating existing infrastructure, such as roads, airstrips, and buildings, disrupting mobility between northern communities. While melting sea ice opens up new waterways for shipping and uncovers resources like oil, diamonds, and rare earth minerals, promising significant economic growth, these changes also raise critical questions about national security and environmental protection.

Canada has an opportunity to navigate these complex challenges by supporting a vision for northern urbanization that respects and incorporates the diverse cultures and traditions of its communities, acknowledging that Indigenous peoples make up a significant proportion of the population in each territory: 86 percent in Nunavut, 51 percent in the Northwest Territories, and 23 percent in the Yukon. A recent investment of over $1.4 million in projects across Nunavut, the Northwest Territories, and Yukon from the Canadian Northern Economic Development Agency (CanNor) is an encouraging step, aimed at upgrading infrastructure, adopting digital platforms, and training and preparing the workforce for a sustainable and prosperous future.

It’ll be interesting to see how a distinct Canadian Arctic urbanism takes shape, balancing economic development with environmental protection and cultural preservation.

Democratizing access to powerful tools has empowered users and companies with systems like Linux, Firefox, Wikipedia, and countless others. Open source components are found in virtually all industry codebases, improving time-to-market, cost savings, and development resources.

Models: The Building Blocks of LLMs

The history of machine learning (ML) and natural language processing (NLP) is built on increasingly sophisticated models, from early text and voice recognition to image recognition and recommendation systems. In 2021, Stanford researchers coined the term “foundation models” to describe a new category of models trained on broad data and adaptable to a wide range of tasks. This laid the groundwork for the Large Language Models (LLMs) that have rapidly emerged and become commonplace today.

The Center for Research on Foundation Models (CRFM) maintains a comprehensive leaderboard of top-performing models, offering a detailed overview of standardized benchmarks across various domains:

Advantages of Open-Source Foundation Models

While closed-source models, particularly those at the frontier of AI research, will likely always occupy the top spots on these leaderboards, the open-source LLM landscape is quickly growing, offering numerous benefits:

• Enhanced customization and flexibility: A wider range of model sizes, indicated by their parameter count in billions (ex. 7B, 70B), allows for fine-tuning to specific needs and computational resources, ensuring greater accuracy and adaptability, especially for niche applications.

• More efficient computational costs: Often have lower computational requirements and provide a superior performance-to-cost ratio, as demonstrated by the CRFM leaderboard for 'Efficiency.' This efficiency translates to lower energy consumption and a smaller carbon footprint, making them more sustainable.

• Transparency: The ability to assess an LLM's training data and architecture allows for easier bias detection and mitigation, ensuring more accurate and fair outcomes. This openness fosters scrutiny and improvement of the technology, building trust and accountability.

Platforms like Hugging Face, a popular open-source AI platform, serve as an invaluable hub for the vibrant machine learning ecosystem by hosting hundreds of thousands of models (over 761,000 at the time of writing), datasets, and applications. Leveraging these resources and experimenting with effective prompting techniques has been instrumental in my AI journey.

Building with Llama 3

Meta's Llama model series has quickly become one of the most popular open-source language models, with Llama 3 released earlier this year and already demonstrating performance comparable to commercial top benchmarked performers. Available in 8B and 70B sizes, I used the 70B model hosted by NVIDIA for my project for the Generative AI Agents Developer Contest by NVIDIA and Langchain.

Inspired to streamline my blog writing process, I developed an application that creates high-quality blog drafts using a specialized team of five agents. Each agent has a unique role, taking a series of inputs and iteratively refining a draft up to 5 times. They’re equipped with web search tools to gather relevant information and integrate it into the draft, showcasing the potential of basic AI agent systems for creative content creation. Check out this 1-minute demo video to see it in action:

From Open Source to Open

A larger 400B parameter version of Llama 3 is currently in training and slated for release later this year, promising state-of-the-art language capabilities, multimodality, multilingual support, and longer context windows (memory). However, with development costs estimated at about $10 billion, it's unclear whether this version will be released under the same community license agreement for research and commercial use. In a recent interview, Mark Zuckerberg's response when asked about open-sourcing such an expensive model was pragmatic: "As long as it's helping us then yeah."

It’s worth noting that Llama models are not truly open-source, but rather openly available with certain restrictions. A definition for open source AI is in development, but would at least entail sharing the source pretraining data and the full disclosure of its underlying architecture — something that Meta and large tech companies are unlikely to do. Although these open models can serve as strategic tools for large players to establish market dominance through “commoditizing their complements," their contributions are undeniably accelerating and democratizing AI adoption. This trend is evident in the open models offered by Microsoft, Google, and Apple.

An example of a truly open-source AI is OLMo from The Allen Institute for AI, a non-profit research institute, with an LLM framework designed to provide access to data, training code, models and evaluation code to empower academics and researchers.

How Open Will The Future Be?

The rise of open-source AI also confronts existing challenges, as highlighted in the 2024 Open Source Security and Risk Analysis Report. It stresses the need to improve open-source management, and highlights the importance of keeping components up-to-date to minimize high-risk vulnerabilities and protect against data breaches and malicious attacks.

For example, hackers have been abusing open-source code libraries by mimicking the names of legitimate Python packages. Awareness and due diligence is needed at all levels - from developers to government agencies, efforts in policy-making and regulation must prioritize these emerging challenges in open-source AI security.

As AI continues to evolve, we can expect a more decentralized landscape beyond cloud-based LLMs, with a diversity of models offering various capabilities and functions that are more energy-efficient and cost-effective. Open-source language models are poised to play a significant role in everyday devices and appliances, powering voice assistants, translation tools, and other utilities, ultimately making AI more accessible and affordable for everyone. This internet of models will be a critical component of responsible innovation, along with the decarbonization of AI.

Can a city densify without losing its connection to nature?

That’s an intriguing question posed on a construction board for King Toronto, an ambitious new urban development project in Toronto’s King West neighbourhood.

It plans to build with a mix of heritage and modern building materials, characterized by hundreds of modular terraces weaved with natural elements and accessible communal spaces. Complementing the existing parks, amenities, and transit options in the surrounding area, especially with the upcoming King-Bathurst subway station, it’s a bold attempt to evolve urban living, harmonizing density with the natural environment in the heart of the city.

Biophilia: A Love of Life

Biophilic urban planning is a global movement that’s reshaping our cities. By incorporating natural elements into built environments, biophilic design aims to improve citizens’ well-being, generate economic benefits, and add resilience against the increasing threat of extreme weather events.

The term 'Biophilia', coined by psychoanalyst Erich Fromm in the 1960’s, stems from the Greek meaning of “love of life or living things”, and later popularized by biologist Edward O. Wilson in 1984 with theories from his book of the same title. It's rooted in the belief that humans have an innate need to connect with the natural world.

The concept gained traction through the work of visionaries like Stephen R. Kellert, a Yale professor who scientifically demonstrated nature's positive impact on human health. And starting from the 1960s, Danish architect Jan Gehl advocated for human-centred urban design, prioritizing pedestrians and public spaces over cars. Their work, along with contributions from others, laid the foundation for today's biophilic design principles.

The COVID-19 pandemic highlighted our need for nature in urban spaces, as many of us found comfort and connected with the outdoors during lockdowns, whether through gardening, adding plants to our homes, or simply spending more time outside. This collective experience has reinforced the value of integrating green spaces into our cities for their crucial role in promoting physical and mental health, mitigating climate change impacts, and fostering social connection.

Revising the Concrete Jungle

The concept of biophilia seems straightforward now, but 20th-century urban development patterns, especially in North America, often undermined this principle. The post-World War II era saw a surge in consumerism, fueled by rapid industrialization and emerging technologies that promised an enviable lifestyle. This, coupled with urban sprawl due to the rise of the automobile, culminated in the creation of car-centric, industrialized cities marked by concrete, steel, and lots of oil, reflecting a desire to dominate nature rather than coexist with it.

Dense networks of highways were designed to facilitate long-distance travel by car, which has disrupted social cohesion within city neighbourhoods. Early brutalist architecture, with its emphasis on exposed concrete, and modular, geometric elements, reflected a sense of urban detachment from the natural world and mirrored the anxieties of modern life. For some time, cities prioritized machines and efficient infrastructure driven by endless growth over human-scale interactions and long-term sustainability.

The Biophilic Street: Reimagining Urban Spaces

While parks have traditionally been the focal point of urban greening, attention is now shifting to where people spend most of their time: the streets. This shift in focus aims to integrate nature into the everyday fabric of city life, recognizing that streets are not just for mobility but vital public spaces.

A recent framework from Sustainable Earth Reviews, a community-focused journal, proposes six categories for the ‘Biophilic Street’:

1. Traffic planning: Trees and bushes can psychologically encourage drivers to slow down, reducing accidents and improving safety.

2. Energy management: Trees and vegetation can reduce building emissions by providing shade and insulation, thus lowering energy consumption for cooling.

3. Stormwater management: Green infrastructure like rain gardens and permeable pavements can help manage stormwater runoff and mitigate flooding by absorbing and filtering rainwater.

4. Biodiversity management: Urban green spaces can support a variety of plant and animal species, contributing to a healthier ecosystem and enhancing biodiversity.

5. Street furniture: Benches, planters, and other elements can be designed to incorporate natural materials and patterns, enhancing the aesthetic appeal of streets and creating a more inviting environment.

6. Education: Informative signage and educational programs can help raise awareness about the benefits of biophilic planning and encourage people to engage with nature in their urban environments, fostering a sense of stewardship and connection.

Investing in Nature for Economic Returns

Beyond its aesthetic appeal and established benefits for human productivity, preferences, and health, many studies highlight the tangible financial returns of exposure to nature, underscoring the need for further investment in biophilic design.

Examples include:

• Healthcare savings through reduced chronic diseases in green communities

• Crime reduction correlated with increased tree canopy density

• Property value increases from street trees, boosting tax revenue

• Reduced energy costs from green roofs and walls that lower heating and cooling expenses for buildings

• Increased worker productivity from offices with natural elements

Additionally, a recent report from the Canadian Urban Institute (CUI) and Natural Assets Initiative (NAI) recognizes the untapped economic potential of natural assets in Canada, such as wetlands, forests, parks, lakes, rivers, and fields, which can play crucial roles in federal priorities for infrastructure resilience, net-zero targets, and 2030 biodiversity targets. Given a rapidly growing population, critical housing shortages, and a changing climate, it strongly advocates for evolving federal housing policies with sustainability at its core.

These initiatives draw inspiration from places like Paris and Singapore, along with the traditional ecological knowledge of Indigenous communities, who have a deep understanding of sustainable land management practices and have been stewards of the land, water, soil, and air for millennia.

Greener, Smarter Cities

In the face of the escalating climate crisis, and as we approach two degrees of global warming, biophilic design and natural asset management must be essential components of urban planning.

For thousands of years, humans lived outside among fields and trees, with our senses attuned to sunlight, open air, and the sounds of running water. However, centuries of civilization and city building have distanced us from these natural rhythms. Thoughtfully merging our built and natural environments, in ways that resonate with our biology, can improve overall physical and mental well-being while also benefitting businesses.

Recognizing that cities are on the front line of climate action, there has been growing enthusiasm for technology to optimize resource use and enable utopian visions of smart cities. This involves AI-powered sensors and other smart technologies to optimize building energy usage, traffic flow, waste management, and more. However, they’ve also faced challenges like data privacy concerns and criticisms that these new developments can feel cold and homogenous, lacking in character and soul.

Biophilic urban design offers a way to humanize these essential technological advancements into our existing spaces, creating a realistic vision of thriving urban centres that integrate nature and technology in a more balanced way.

This approach is invaluable for anyone seeking to extract insights from private data. For businesses, this can be a game-changer, offering a competitive edge with proprietary or confidential business data - and with the rising threat of data breaches and cyberattacks, safeguarding sensitive information while leveraging AI is crucial. Consider this detailed example illustrating a customer chatbot at a food delivery company:

RAG Analogy: The AI Librarian

To extend the example: think of an AI assistant as your personal librarian, accessing a secure collection of your chosen documents. The assistant can answer specific questions about your data much more effectively when they have your documents figuratively in front of them, rather than relying solely on their general knowledge. RAG inherently promotes transparency by providing a clear trace of the information used, reducing the risk of fabricated responses or hallucinations. It can also leverage external databases and various tools to ensure comprehensive and up-to-date knowledge retrieval.

Retrieval-Augmented Everything?

At the recent Collision tech conference held here in Toronto, industry leaders recognized the immense potential of enterprise knowledge management systems built with RAG, along with the challenges for its adoption. They emphasized how these systems can enhance user experiences and cut costs through smarter, more personalized customer chatbots, search engines and content generation. Despite the hype surrounding AI, most businesses haven’t yet realized its full potential.

Poor data quality, a longstanding issue for many businesses, negatively affects LLMs’ ability to extract valuable insights, which can lead to inaccurate or misleading outcomes, diminishing user trust and possibly being counterproductive. Additionally, ensuring robust data privacy and overall trustworthiness remain key challenges. Research highlights the importance of preventing systems from being manipulated, whether consciously or unconsciously, to produce unreliable decisions that could be harmful, especially in safety-critical scenarios.

While RAG still currently faces issues with complex 'reasoning' and 'mathematical' tasks, the emergence of multimodal models like ChatGPT-4o, which can process audio, images, and video, signals a shift towards more natural AI interactions that were once just science fiction. These models aim to build a more comprehensive understanding of the real world, leading to unprecedented accuracy and insight in their outputs. This opens up a world of new possibilities, but it also underscores the critical need to address and prioritize data privacy concerns.

It's an exciting, daunting, and potentially transformative prospect, depending on one's perspective on how AI will play out.

Try out RAG

Curious to experience RAG yourself? Here are some ways you can get started:

• Upload to regular LLM Chats: Most, if not all, accessible LLMs now have features that resemble RAG, allowing you to upload a few documents at a time and ask questions about them. While this isn't fully-fledged RAG, it provides a practical introduction to the concept. Always remember to check the privacy policy for the model you’re using and ensure it meets your comfort level.

• Cohere’s Coral Agent: Cohere’s chat web UI includes RAG-like features, such as specific web search capabilities and file analyzing functions that demonstrate its fetching capabilities from external sources. They also offer great foundational training to learn more about RAG.

• You.com: Founded in 2020, You.com started off as a search engine that became the first to integrate consumer LLMs like ChatGPT with real-time Internet access for up-to-date and accurate responses with citations. Without using your own data, it’s a notable example of RAG’s transparency.

We’ve become deeply attached to our smartphones, computers, and countless other accessories that shape our identities and enable our modern lives. Yet, when we’re done with our indispensable gadgets, we often move on to the next with little regard. A 2015 UNEP report estimated that 60-90% of the world’s electronic waste, worth nearly 19 billion USD, is illegally traded, incinerated, or dumped each year, affecting local communities through the toxic chemicals that get released from poor disposal, contaminating the air, soil, dust, and water to ultimately harm our health and well-being.

From Earth to Body: Toxic Consequences

Massive waste from industrialization and consumer society, composed of hazardous chemicals, precious metals, and other materials improperly dumped, has unfortunately already damaged generations of populations. E-waste adds further to that. Here are three prominent examples:

• Lead Poisoning: Historically valued for its durability, malleability, and corrosion resistance, lead became an escalating problem in the mid-20th century. Increased use in consumer products like paint and gasoline led to a surge in lead exposure, affecting a generation of Americans with diminished IQs, altered personality traits, and increased crime rates.

• PFAS (Per- and Polyfluoroalkyl Substances): Used in various consumer products like non-stick cookware and food packaging, these chemicals are leaching into our drinking water and soil when disposed. Current scientific studies indicate adverse effects for pregnant women, developmental effects or delays in children, and increased cancer risks.

• Microplastics: These tiny particles from plastics are ingested unknowingly from what we eat, drink and breathe. They are embedded in our all parts of our bodies, including our heart and brain. While the long-term effects are largely unknown, early scientific insights suggest increased vulnerability to heart disease, cancer and other illnesses.

Arcadia Earth - An E-Waste Exhibit

Museums have always been sources of knowledge and inspiration, and they’re evolving to address contemporary issues like climate change. During a recent visit to Arcadia Earth in Toronto, I was inspired by one striking installation called the 'E-Waste Generator', featuring creatively repurposed e-waste and a throne made entirely of discarded keyboards.

Embracing The Electronics Circular Economy

Inspired by the visit and the spirit of taking action, I undertook extensive spring cleaning, selling and donating old electronics accumulated over years.

However, tackling e-waste requires systematic changes. In 2019, the United Nations set a global framework for a circular economy approach to e-waste management called the 'New Circular Vision for Electronics'. It calls for transforming our current systems of ‘take, make and waste’ to a circular economy that benefits all:

Reduce (Take)

• Dematerialization: The easiest solution for all e-waste issues is still not to generate any e-waste in the first place. Companies are increasingly prioritizing recyclability and circularity at the manufacturing stage, using minimal resources and ethically sourced materials and considering how metals and minerals can be more easily recovered from devices in a financially viable way. The trend of reducing physical hardware and material use is underway by shifting to more virtual, service-based models like cloud computing and streaming media.

• Sustainable Design at CES 2024: At the Consumer Electronics Show, big tech companies showcased commitments to sustainable product designs, with particular focus on ‘closed loop recycled plastics’. For example, LG aims to use recycled plastics across 19 product lines with a 95% recycling rate by 2030.

Re-Use (Make)

• Planned Obsolescence: Too many products are replaced once a newer, more sophisticated version comes out, or disposed of when they show signs of malfunction. It’s a strategy used by companies to increase sales by encouraging this wasteful behaviour among consumers, subtly slowing down devices or causing them to wear out prematurely to artificially reduce their lifespan.

• Right to Repair: This legal right for device and equipment owners to freely modify and repair products is gaining great momentum in 2024. Google has proposed policies in a new white paper to empower people to prolong the life of their Pixel smartphone devices. More legislation is being passed in many US states, and the EU’s updated legislation, passed in April, advances requirements such as modular design, user labels for repairability and durability, more easily replaceable batteries, and upcoming regulations for repairing laptops.

Recycle (Waste)

• Consumer Technology Circularity Initiative (CTCI): Announced on the first day of CES 2024, this voluntary industry initiative to reduce waste, enhance recycling, and lower climate impact. Founding members like Lenovo, LG Electronics, Panasonic, Samsung, and Sony Electronics Inc. are committed to developing innovative solutions that help the circular economy become a reality.

• Urban Mining: This promising and lucrative practice reduces our carbon footprint by extracting valuable metals and minerals from billions of smartphones and other consumer devices that would otherwise end up in landfill or be incinerated. Recovering and circulating materials into new products not only has significant economic potential, but also reduces the strain on local ecosystems and pollution from mining at the source, especially as the world’s consumption of raw materials is set to double by 2060.

The Informal Sector and E-Waste Challenges

While progress is being made, standardizing recycling processes remains inherently difficult due to the sheer scale and variety of components in different electronics. Increasing consumer options and the growth of lithium-ion EV batteries, IoT devices, cryptocurrency, and AI data centre hardware further compound this challenge.

With much of the world's e-waste going unrecycled and unaccounted for, a shocking amount ends up being illegally traded or dumped in lower-income nations. This has given rise to informal "backyard" sectors comprised of resourceful entrepreneurs and low-wage workers who collect, disassemble, and try to extract value from e-waste through crude techniques like open-air burning and acid baths.

It’s a common source of income for marginalized communities and often supports formal waste agencies like municipalities; however, there are virtually no safety measures, environmental safeguards, or oversight. The ILO and WHO estimate millions, including many women and children, work in these unregulated conditions, exposing themselves and their surroundings to hazardous toxins from improper handling and processing of e-waste components.

Innovative E-Waste Solutions

Recognizing the economic benefits delivered by a circular model for electronics, there’s ample opportunity to create more sustainable, green jobs for repairing used electronics, along with collecting and pre-treating e-waste. Formalizing informal sectors around the world will increase the rates of valuable recovered materials and greatly improve working conditions.

AI-powered robotics can also play a pivotal role; they give us a glimpse into the future and can significantly assist with e-waste management by automating sorting, disassembly, and recycling processes. This minimizes human exposure to harmful substances and increases the efficiency of extracting valuable metals. For example, Apple's Daisy robot disassembles iPhones to recover valuable components. These technologies can enhance recycling efficiency, reduce labour costs, and ensure safe handling of toxic materials at scale.

To bridge the digital divide and empower communities, using and sharing existing essential resources are both crucial and economically promising. As well, joint global agencies and organizations have been raising the alarm bells for awhile now. Prioritizing legislation will help with uneven global progress on e-waste management - the EU is at the forefront of e-waste management leading with a 42.5% collection and recycling rate, but regions like the Middle East, India, and Brazil have recycling rates close to 0%. We have a significant task ahead to steer us in a positive trajectory; it starts with better appreciating the circular value of our electronics and thoughtfully managing its waste.

Unraveling the Black Box

The 'black box' nature of technology has often fueled innovation and adoption, but modern LLMs have taken this opacity to unprecedented levels. Even the developers who created these complex models face increasing challenges in fully understanding how they arrive at their outputs, as the complexity of these models has grown exponentially in recent years.

This has raised alarms across multiple sectors, including technology, ethics, and governance, prompting widespread calls for increased caution and regulation. In March 2023, many tech leaders like Elon Musk and Steve Wozniak signed an open letter calling for a six-month pause on ‘Giant AI’ more powerful than GPT-4. This pause aimed to provide time to set safety standards for AI design, mitigating the risks of an unrestricted AI that forms an artificial mind of its own, posing potentially dangerous and unpredictable threats to humanity.

From Data to Knowledge: How AI Models are Trained

The training data and processes at the core of an LLM’s language abilities are crucial. This involves two main functions that ‘configure’ the knowledge:

Pre-Training

This initial phase, typically conducted by major tech companies like OpenAI and Google, involves processing massive datasets in a computationally intensive process at large data centres (producing a significant carbon footprint). Two key types of data are used:

• The 'Digital Commons': A vast collection of publicly available sources like websites, books, and Wikipedia. This foundational open-source data provides the model with broad general knowledge.

• Proprietary Data: Companies often supplement the commons with their own exclusive data assets like user-generated content, product analytics, and code repositories. This enhances accuracy, relevance, and differentiates their models from competitors.

Fine-Tuning

Fine-tuning is a crucial step in tailoring a pre-trained LLM for specific applications. By training the model on specialized datasets, its expertise can be enhanced across various domains while still preserving its general language capabilities.

In highly regulated industries such as healthcare, finance, and law, fine-tuning must adhere to strict data compliance standards. Proprietary or sensitive data is carefully managed to ensure security and confidentiality, allowing organizations to leverage the power of LLMs while maintaining compliance.

The Ethical Considerations of AI

As AI models become increasingly capable, their potential to amplify societal biases and cause negative political and economic consequences cannot be ignored. Careful consideration is required in three key areas:

The Importance of Data Provenance

Transparency regarding the origin, processing, and access to training data is crucial for understanding implicit biases that could manifest in generated outputs. Ensuring fairness in a global context is complex, requiring a delicate balance to avoid perpetuating societal biases and inequalities across cultures and populations.

The rise of AI intensifies concerns about copyright, intellectual property, and privacy violations in the digital world. As more of our online activity is captured, the ongoing contention over ownership of “unique data” from social media companies like Facebook, Twitter, and TikTok further complicates the debate over data rights and usage.

And with AI systems generating exponentially more content, diligent filtering for misinformation, hate speech, deepfakes and hallucinations is essential to maintain data integrity. Failing to do so could contaminate the digital commons, erode public trust, and exacerbate societal harms.

Data Privacy

Handling and protecting sensitive information, such as personal data and intellectual property, within the training data is critical for LLMs. As cyber threats constantly evolve, it's crucial to ensure models are robust and protected from misuse or data breaches.

As users, we should be aware that conversation histories with LLMs may raise similar privacy concerns as our web browsing history. Protecting sensitive information is particularly important in specialized enterprise domains like medicine and law, where confidentiality is paramount.

Model Plasticity: How AI Changes and Learns

Beyond the data they are trained on, LLMs themselves are also adaptable and can change based on interactions and further training. Their probabilistic nature as next-word prediction models makes them quite malleable. Much like the neuroplasticity of the human brain, they can change their response characteristics based on different prompting strategies and fine-tuning data inputs.

• AI Alignment: Extensive efforts aim to make AI helpful, safe, and reliable by encoding human values and goals. Despite safeguards, established LLMs often include warnings about offensive or biased content and explicitly caution against using them for legal, financial, or medical advice due to the potential for harm. Concerns also exist about malicious manipulation by bad actors, such as generating disinformation, deepfakes, or propaganda.

• Interpretability: This concept seeks to shed light on the model’s predictions and decision-making process. Preventing hallucinations and ensuring reliable, reproducible outputs are essential for businesses and critical industries (ex. medical diagnoses, fraud detection). Interestingly, researchers have found that LLMs can interpret and audit themselves, a promising development to enhance their reliability and trustworthiness at scale.

Building a Responsible Future for AGI

As AI systems rapidly advance, major tech companies have designated a new class called 'frontier models' - highly capable foundation models that could push the boundaries of what’s possible, and also potentially possess dangerous capabilities posing severe public safety risks.

Since the open letter calling for a pause on ‘Giant AI’ development, global coordination efforts have begun, with the second AI Seoul Summit in May 2024 marking a significant step towards establishing AI standards and governance. At the summit, major tech companies from the United States, Europe, the Middle East, and Asia signed the Frontier AI Safety Commitments, pledging to create safety frameworks for measuring and mitigating the risks of their respective AI models. They’ll reconvene for the next “AI Action Summit” in France in early 2025.

However, the recent departure of key Safety team members from OpenAI is particularly concerning, serving as a reminder that commitments alone are insufficient, and must be followed by dedicated action. Jan Laike, one of the co-leaders of the ‘superalignment’ team, stated bluntly: “I believe much more of our bandwidth should be spent getting ready for the next generations of models, on security, monitoring, preparedness, safety, adversarial robustness, (super)alignment, confidentiality, societal impact, and related topics. These problems are quite hard to get right, and I am concerned we aren’t on a trajectory to get there.”

The runaway success of LLMs and the booming AI industry inevitably propels us towards artificial general intelligence (AGI)—where models will eventually equal or surprass human intelligence. Understanding and guiding these sophisticated artificial 'minds’ is imperative, particularly given the inherent black box nature of their decision-making processes. Implementing safeguards aligned with human ethics and values to ensure AI benefits society while minimizing harm is a critical global challenge.

How we navigate and commute through our urban spaces shapes our well-being and connection to our communities.

Living in downtown Toronto since my early 20s, owning a car was both unnecessary and unaffordable. I embraced walking and biking as my main modes of transportation, staying fit while immersing myself in this uniquely Canadian concrete jungle.

Exploring a Cultural Mosaic

Toronto’s downtown core is a vibrant tapestry of cultures and cuisines, representing one of the most multicultural cities in the world. The city’s relative flatness, with a gradual slope rising from the lake northward, makes for pleasant strolls through diverse neighbourhoods with distinct flavours, bustling street festivals, and a wide variety of languages.

However, Toronto is lacking a signature pedestrian walkway, a hallmark of many world-class cities. Iconic spaces like Barcelona’s La Ramblas, New York City’s High Line, and Copenhagen’s Strøget are inspiring urban gems that enhance walkability and cultivate memorable experiences, solidifying them as cultural landmarks over time. And crucially, these spaces prioritize pedestrian safety—an urgent need here in our city, as underscored by a recent study: between 2016 and 2021, a staggering 9,700 pedestrians sought emergency care due to injuries sustained on the city's streets.

The PATH: Underground ‘Metropolis’

In the absence of a dedicated pedestrian walkway above ground, Toronto does boast an impressive alternative: the PATH, North America’s largest underground pedestrian network. Its 30 km of walkways connect retail, services, and entertainment across 75 buildings concentrated around the financial district, linking civic landmarks like City Hall and Metro Hall with popular attractions such as the Hockey Hall of Fame, Roy Thomson Hall, and Scotiabank Arena.

The PATH, although currently experiencing a slower recovery from the pandemic, remains a vital artery of Toronto’s urban life. As “one of Toronto’s great land-use triumphs”, it showcases the effective integration of transportation, work, and leisure in a dense urban environment. Additionally, it’s especially helpful for commuters sheltering from cold winters and increasingly hot summers.

Future Green Neighbourhoods

A new model of pedestrian-friendly urban planning is emerging on the eastern waterfront; the Villiers Island neighbourhood offers an ambitious vision of an innovative design that incorporates nature and prioritizes pedestrians. Plans for car-free streets and abundant green space, engineered to mitigate flooding events from the Don Valley, demonstrate a commitment to both sustainability and community well-being. This model of infrastructure not only builds resilience to extreme weather but also fosters a stronger sense of community, better equipping Toronto to face the challenges of the 21st century.

Two-Wheeled Commute

An essential approach to urban mobility, cycling has become one of the fastest growing commuting options in Toronto. It’s been my primary way of getting around since I moved here, as it’s often the quickest way to get around the downtown core (when it’s not raining or snowing too much).

The city’s network of bike lanes is quickly expanding, yet safety remains a common concern due to car congestion and ongoing construction. Some routes lack full continuity, forcing cyclists to navigate a pathwork of dedicated lanes and shared roads, often marked with potholes. I've grown accustomed to this over time (and experienced a few close calls—always wear a helmet), but cycling should feel safer and less stressful.

Despite the challenges, the growing popularity of Bike Share, a convenient and accessible option offered by the city, reflects a global trend towards urban mobility. Ridership continues to break records every year, with the system offsetting 12.1 million pounds of CO2 emissions in 2023 alone. This highlights the collective impact of daily sustainable transportation decisions, and I’m excited to see further developments that make cycling here even safer and more accessible.

Elements of a Livable City: Connection, Resilience, Safety

Engaging more intimately with our built environment can nurture deeper connections within our communities. However, established car-centric infrastructure presents a significant obstacle for Toronto—it pollutes our air, endangers lives, and fragments our urban fabric. To build a healthier, inclusive, and more connected city, we must shift our urban planning and infrastructure development focus towards people.

To address the crisis of pedestrian traffic deaths, Toronto’s Vision Zero Plan introduced in 2016 aims to eliminate traffic-related fatalities and serious injuries. Working in coordination with various advocacy groups, the plan is making strides towards a safer and more walkable city through redesigned intersections, red-light cameras, improved signaling systems, and expanded bike lanes.

Shifting the Focus: Prioritizing People Over Cars

While the TTC faces ongoing challenges, promising public transit developments like the Ontario Line construction are much needed and will help fill widening transit gaps caused by increasing car congestion. In addition, the rise of micromobility, particularly low-carbon options like e-bikes and e-scooters, offers an effective and sustainable solution for navigating the vast distances inherent in the Greater Toronto Area’s infrastructure.

Enhanced mobility is key to revitalizing the downtown core, which continues to recover from the pandemic. A thoughtful, human-centred approach to urban planning, backed by strong political will, is essential for achieving a more livable and affordable city. This means prioritizing cleaner air, universal accessibility, reduced transportation costs, quieter streets, and a significant decrease in greenhouse gas emissions.

Furthermore, enhancing walkability and bikeability can combat the growing epidemic of loneliness in Canada. Creating more walkable, bikeable, and transit-friendly cities fosters chance encounters, community interactions, and a greater sense of belonging, thus addressing both physical and social well-being.

We’ve already established a substantial carbon footprint with our daily technological usage, and with the escalating growth of AI, the environmental and human impacts of this technology deserve critical attention along with praise for its benefits.

AI’s Energy Hunger

As artificial intelligence technology gets smarter, it uses exponentially more energy. Data centres are the backbone of AI, powering everything from simple chats to climate forecasting models - in 2023, they produced about 2.5% to 3.7% of global carbon emissions and could use as much energy as countries like the Netherlands by 2027.

While initially built for large-scale networking, storage, and cloud computing, data centres are undergoing a major transformation to meet the escalating energy demands of AI. This shift towards more power-hungry applications has prompted the development of green data centres, which prioritize energy efficiency and renewable energy sources to mitigate their environmental impact.

Power Grids

Beyond data centres, our power grids are under immense pressure - the growing adoption of electric vehicles, the rise of cryptocurrency mining, and increased demand for air conditioning due to climate change are just some of the many factors contributing to additional strain on existing infrastructure, which is unsustainable.

Upgrading our power grids to accommodate an electric future and divert from fossil fuels is essential. This will require billions of dollars in new investments (1, 2) to develop and implement regionally tailored solutions, including new engineering designs for machinery and building materials, as well as advanced analytical models for optimizing electricity distribution.

Training AI

Training AI models is an energy-intensive process, requiring massive amounts of data to be fed into the models over several months to 'teach' them complex tasks such as language understanding, image and speech recognition, and decision-making. This demands immense computational power, resulting in significant carbon emissions that can vary greatly depending on the location of data centres and the type of energy used for training.

“Training GPT-3 once on a clean grid is equal to driving 1 car for one year, and training on a dirty grid (one reliant on fossil fuels) is equal to driving 40 cars for one year.”

The energy consumption of AI training is largely determined by the complexity of models, often defined by billions of ‘parameters’ – numerical values that fine-tune the AI's ability to learn and perform tasks. Many top performing models, developed by major tech companies, keep their exact number of parameters confidential for competitive advantage, while open-source models are transparent about their parameter count, enabling easier comparison and analysis.

Physically Resource Intensive

Data centres are not just energy-intensive, but also incredibly resource-hungry, relying on a wealth of materials for their construction and operation, including precious metals like gold, silver, and copper - some of which are classified as critical raw materials (CRMs).

Moreover, continuous upgrading and decommissioning of hardware like servers and cooling systems generates massive amounts of electronic waste, which can leach harmful substances like lead and mercury into soil and water if not handled correctly.

The Hidden Water Footprint of AI

The intensified computing power of AI-powered data centres generate significant heat, requiring vast amounts of freshwater for cooling to prevent infrastructure damage. For instance, training one large language model like Chat-GPT3 can use about 700,000 litres of freshwater - that’s enough to fill over 280 backyard swimming pools.

With a global freshwater crisis, it’s important to be mindful of how and where we use this scarce resource. AI’s growing water footprint deserves more attention, particularly in arid or drought-prone regions where data centres strain local water supplies, competing with residents and agriculture. Increased transparency is essential for understanding the full extent of water usage and taking steps to mitigate its impact.

AI as a Sustainability Catalyst

Electricity and heat production are the largest contributors to global greenhouse gas emissions, and AI has the potential to redefine how we generate and consume energy. Through innovation and increased adoption of regenerative practices, we can minimize resource depletion and identify efficiencies in virtually every domain — developing effective solutions that not only mitigate climate change but also deliver positive financial outcomes.

Innovating for Energy Efficiency

With growing awareness, the tech industry is working on ways to make AI more energy-efficient, like compressing models and optimizing how the next word is predicted (‘inference’).

As the market matures, consumers will have greater access to smaller, open-source models that can be run locally on personal computers; this allows customization tailored to user needs, privacy, security, and cost savings by reducing the need for expensive cloud-based processing.

Accelerated computing is another key innovation, utilizing specialized hardware like Graphics Processing Units (GPUs) and Google’s Tensor Processing Unit (TPUs), designed for machine learning, to offload work from the central processing unit (CPU) and improve overall energy efficiency.

Infrastructure and Manufacturing

In addition to optimizing power grids, AI is playing a crucial role in transforming the way we design and operate buildings. For example, AI-generated concrete formulas are a promising innovation that can lower emissions in the construction industry, which is particularly important for rapidly growing populations.

AI-driven smart buildings can optimize resource use by adjusting lighting, heating, and cooling based on occupancy and weather patterns, leading to reduced carbon emissions. In addition, intelligent IoT solutions are being deployed in commercial and industrial buildings to collect and analyze data on energy usage, occupancy patterns, and equipment performance, leading to increased cross-sector collaboration and more efficient energy management.

Global Logistics and Supply Chains

With the interconnected nature of our world, AI has the potential to revolutionize the way we move goods and people to make transportation more efficient and sustainable. By leveraging data analytics and predictive modeling, supply chains can optimize routes, schedules, and resource allocation, leading to reduced fuel consumption and emissions. However, the complexity of global supply chains requires thoughtfully designed solutions to avoid exacerbating existing inequalities and biases.

We’re already optimizing shipping routes for cargo ships to minimize their fuel consumption, which is vital for protecting our oceans. In aviation, airlines are optimizing flight paths and reducing delays, resulting in less fuel burn and unnecessary emissions — and on the roads, applications like Google Maps utilize AI to provide drivers with more fuel-efficient routes, effectively removing an estimated 250,000 fuel-based cars off the road for a year (which might finally be self-driving in a few decades).

Agriculture

With food production accounting for 40% of the world’s greenhouse gas emissions and a global population projected to increase by two billion by 2050, the need for sustainable agriculture is increasingly urgent.

Precision farming utilizes AI-powered drones and sensors to customize the management of individual crops and land areas, boosting overall yields and optimizing resources to meet the demands of a growing and hotter world. AI-driven research is also contributing to the development of new climate-resilient crop varieties, while tools for enhanced pest and disease detection, soil analysis, and livestock monitoring are further improving agricultural sustainability and productivity.

The agricultural sector's heavy reliance on water, which accounts for 70% of global freshwater withdrawals, makes efficient water use absolutely necessary, given the increasing demands of data centres and the ongoing global freshwater crisis.

Flood and Disaster Management

Warmer temperatures induced by climate change disrupt the natural water cycle, leading to an increase in extreme weather events. Among these, floods are the most common natural disaster, accounting for almost 10 million people displaced in 2023 alone.

AI is proving invaluable in predicting and responding to these devastating events, as well as other natural disasters such as earthquakes, hurricanes, and wildfires. By analyzing vast amounts of data from weather patterns, satellite imagery, and social media, early warnings can be identified that allow communities to prepare and evacuate in advance, potentially saving countless lives and reducing economic losses. In the immediate aftermath, AI-assisted satellite imaging and remote sensors can quickly assess damage and streamline the often challenging recovery process.

Achieving Equilibrium

At the recent COP28 climate change conference event, global leaders made a landmark declaration: the beginning of the end of fossil fuels. A historic agreement was reached among all nations to achieve net-zero global emissions by 2050, prioritizing a rapid transition to renewable energy and significant emission reductions. Notably, there was also recognition of AI’s essential role in tackling climate change, while ensuring it doesn’t widen the digital divide caused by unequal access to technology.

While international agreements set ambitious goals, the responsibility for achieving sustainability ultimately falls on national and local governments, as well as individual companies and consumers. Governments at all levels need to enact stronger regulations and incentives for renewable energy adoption, while big tech companies must prioritize transparency and invest in sustainable AI research and development, while prioritizing decarbonization efforts across all scopes of emissions.

The Path Forward

Collaboration and accountability across all stakeholders are necessary to ensure the responsible implementation of artificial intelligence; these technologies must be accessible, reliable, and beneficial for everyone while minimizing harm to both our communities and the planet.

In the short term, AI’s rapid growth presents major challenges in balancing competitive technological advancement with sustainability goals, despite many companies having made climate pledges in recent years. For example, Microsoft's greenhouse gas emissions were actually around 30 percent higher in 2023. Consumers can play a crucial role by demanding transparency from companies and prioritizing sustainability when choosing AI products and services, using accountability mechanisms like model cards, which function like nutrition labels to provide transparency into large language and machine learning models. However, most model cards currently lack carbon emission reporting.

The convergence of technological advancements in areas like IoT, materials science, renewable energy, and smart grid infrastructure, alongside increased awareness and policy action, offers cautious optimism for a sustainable AI future.

In this article, we’ll delve into the fundamentals of LLMs, explore the basics of prompt engineering, and preview the future of AI assistants.

How LLMs Work

Since the 1950s, researchers and scientists have been working towards enabling computers to understand, interpret, and generate human language, laying the foundation for the fascinating domain of Natural Language Processing (NLP).

Various advancements in computing over the last few decades — including machine learning, neural networks, cloud computing, and vast datasets — have significantly improved their language understanding, paving the way for powerful language models.

A pivotal moment occurred in 2017 with the introduction of transformer architectures and self-attention mechanisms. This breakthrough redefined the capabilities of language models, enabling them to understand the context and relevance between words in unprecedented ways.

Fundamentally, LLMs are highly sophisticated word predictors. They excel at predicting the next word in a sequence based on the vast amount of text they've been trained on, and their inherent probabilistic nature means that even with the same prompt, the output will vary slightly each time.

Crafting Effective Prompts

While today's LLMs can generate fluent, human-like text with ease, their brilliance has a blind spot: they rely on making assumptions based on patterns in their training data to generate responses. This can lead to potetial misinterpretations, incomplete results, and even fabricated information (known as “hallucinations”).

Prompt engineering offers a set of valuable techniques designed to maximize the effectiveness of LLMs and mitigate these risks. It’s the science of constructing the initial prompt to enhance the model’s understanding and guide it towards generating the desired output, essentially reducing the need for assumptions.

The field of prompt engineering is extensive and rapidly evolving, with new techniques and approaches emerging constantly. It's definitely worthwhile to explore and experiment to discover what works best for you. Here, we’ll focus on some basic but powerful methods to get you started:

Mastering the Art of Instruction

Clear and detailed instructions are essential for guiding the LLM to produce a desired output. Here’s a breakdown of key elements to include in your prompts:

• Define the Task (“What do I want?”)

  • State the task using actionable verbs and descriptive adjectives in a conversational tone.

  • Be clear and precise to minimize ambiguity.

  "Write an inspiring and informative article about the positive impact of community gardens on urban sustainability."

• Identify the Audience (“Who am I?”)
  Specify the intended audience to tailor the language and tone.

  “Explain climate change to a 5 year old.” vs. “Explain climate change to a scientist.”

• Adopt a Persona (“Who are you?”)
  Instruct the LLM to adopt a specific perspective or voice.

  "Write a poem from the perspective of a tree." or "Create a dialogue between two historical figures."

• Provide Context (“What should be considered?”)
  Offer relevant background information or guidelines.

  "Write a news article about a recent development in urban sustainability, focusing on its impact on affordable housing."

• Set the Format (“How do I want it?”)
  Define the desired structure or style of the output.

  "Summarize the key points in bullet points" or "Write a persuasive essay in five paragraphs."

• Establish Constraints (“What do I NOT want?”)
  Explicitly state any restrictions, exclusions, or limitations in a separate line within your prompt.

  Task: "Write an engaging blog post about the benefits of cycling for urban commuters that doesn’t mention any environmental benefits."
  Constraint: "Do not mention any environmental benefits."

Advanced Formatting Techniques

Strategic and consistent formatting help the model parse your instructions accurately; it can significantly enhance the quality and relevance of outputs, especially for more complicated tasks. Consider these techniques:

• Capitalization: Use ALL CAPS to emphasize keywords or phrases.

• Line Breaks (\n): Separate distinct parts/ideas with new lines.

• Bullet Points: List multiple items or steps in an organized manner.

• Numbers/Letters: Enumerate steps or options for structure.

• Quotes (" "): Denote direct speech, citations, or specific terms.

• Parentheses (()): Provide additional information or clarification.

• Square Brackets ([]): Create placeholders for information the LLM should fill in.

• Angle Brackets (<>): Indicate variables or parameters.

• Curly Brackets ({}): Denote specific data types or formats.

Zero-Shot and Few-Shot Prompting: The Power of Examples

One of the most powerful ways to improve LLM output is to provide examples, demonstrating the desired format, style, or content you’re looking for.

• Zero-Shot Prompting: This approach, where you provide a single instruction without examples, is often sufficient for straightforward tasks where the desired output is clear and doesn’t require a specific structure or style.

• Few-Shot Prompting: This technique involves providing a few examples of the desired output along with your instruction. It’s particularly useful when you need to:

  • Establish a specific format, style, or tone.

  • Guide the LLM’s creativity within certain boundaries.

  • Ensure consistency across responses in the conversation.

  • Handle nuanced or domain-specific tasks.
    

  By providing a few well-crafted examples, you essentially give the LLM a blueprint for the desired output, enabling it to better align with your expectations and produce more tailored and effective results.
  
  Here’s a practical example for a structured brainstorming session:

Task: I need help brainstorming approaches to improve public transportation and reduce car usage in our city. Please use the following format for each response:

Observation: [Your observation about a current challenge related to public transportation or car usage]
Proposed Solution: [A specific, actionable solution to address the observation]
Key Benefit: [The primary advantage of implementing this solution]
Potential Concerns: [Acknowledge any negative effects or consequences]

Prompting the Future with AI Assistants

A well-crafted prompt can often provide what you need, especially when you have a clear understanding of the task and relevant details. However, for open-ended or complex problems, starting with a shorter and focused prompt is a powerful strategy. Engaging in a conversation allows you to explore ideas, clarify ambiguities, and refine your approach. This intuitive, iterative process enhances our ability to utilize vast amounts of information.

Evolution of the LLM Agent

1. Planning and Reasoning: Advanced prompting techniques like Chain of Thought prompting break down complex tasks into smaller, manageable steps, mirroring human problem-solving. This enables them to tackle intricate challenges that a single prompt could not address effectively. Additionally, techniques like Reinforcement Learning from Human Feedback (RLHF) leverage human input to refine the model’s reasoning ability.

2. Memory: LLMs utilize memory to process information and make decisions. The LLM’s context window, its short-term memory, determines how much of the ongoing conversation it can “remember”. Long-term memory techniques like embeddings (vector representations of words and phrases) and vector storage allow LLM agents to remember information beyond the immediate context and produce coherent responses over extended conversations.

3. Tool Usage: Building upon their reasoning and memory capabilities, LLMs can leverage a wide range of tools like search engines, specialized databases, and even code interpreters. This expanded toolkit enables them to perform tasks like fact-checking, complex data analysis, up-to-date information retrieval, and even generating visual representations of data, making them even more versatile and valuable in diverse applications.

LLM agents are rapidly transforming into capable collaborators, progressively learning to understand our intentions, reason through complex tasks, and take appropriate actions independently. Along with advancements in image, video, and audio generation, we’re witnessing the rise of AI assistants seamlessly integrate into our work and daily lives, opening up a world of possibilities for increased productivity and creativity.

By better understanding the fundamentals of these techniques and the evolving capabilities of LLMs, we can continuously adapt to minimize the gap between technology and user needs, empowering us to realize the full potential of this transformative technology.

Growing up in the 90s, I was captivated by technological innovation. At the dawn of the Internet, we witnessed the birth of email, websites, and online chatrooms that connected us like never before. Colourful Apple Mac computers teased a vibrant, tech-enabled future, a spirit of progress embodied in shows like The Jetsons and Star Trek that showcased technological marvels.

Rapid urbanization has reshaped societies across the globe in recent decades, with larger populations migrating from rural to urban areas in search of greater economic opportunities and a better quality of life. This shift demands more attention to how we design, develop, and manage.

“One thing The Jetsons couldn’t have foreseen is how large cities would grow or how important sustainability would become to their prosperity. From just over 3 billion people in 1962, the world’s population has grown to more than 7 billion today and is on a trajectory to surpass 9 billion by 2050. Cities, in particular, are experiencing rapid change. Today, more than half of the world’s population lives in an urban area; by 2050, that fraction will grow to two-thirds.”

Fueled by our current economic systems, this growth has serious consequences. Overreliance on polluting industries, resource inefficiencies, and car-centric infrastructure are just a few ways we produce the excessive greenhouse gas (GHG) emissions that drive climate change. The record-breaking temperatures Canada experienced in 2023 make the consequences of our actions undeniable.

Yet, a shift towards sustainability is underway, and cities around the world are recognizing the urgent need to adapt.

Thinking like an urbanist

Urbanism examines ‘how inhabitants of urban areas, such as towns and cities, interact with the built environment’; it’s a key framework for city planners, urban developers, architects, politicians, etc.

By extension, citizens who enjoy city life and are inspired to respond to climate change are increasingly embracing an ‘urbanist’ mindset. They appreciate the built environment around them and its potential to enable sustainable and advanced lifestyles with less consumption, sprawl and pollution.

Here are some of the main functions of our communities that we can make more resilient:

• Transportation: access to low to no carbon options like walking, biking and taking public transit to move around and commute are essential. Heavy reliance on private vehicles enabled by car-centric infrastructure leads to traffic congestion and carbon emissions - for example, it accounts for over 40% of Toronto’s overall GHG emissions.

• Resource Efficiency: dense urban areas demand smart management of energy, water, and waste systems. They must adapt to increasing weather extremes like floods, rising sea levels, and hurricanes. We can all play a small part in the circular economy to better use resources and minimize waste.

• Food Systems: by adjusting our diets, promoting local food sources, advocating for sustainable production, and reducing food waste, we can collectively address 25% of global GHG emissions.

• Buildings and Infrastructure: creating energy-efficient and resilient buildings is essential for sustainable cities. Infrastructure development should prioritize energy efficiency, resilience to weather extremes, and human-centric design to ensure livability for a denser population.

• Social Cohesion and Culture: lack of awareness about sustainability issues stems from multiple aspects of our socio-economic reality, including rising costs of living and the spread of misinformation. Development decisions need to be equitable as much as they are effective and innovative.

Building livable cities

Our built environments are facing pressures from a complex world. Climate change disrupts our way of life with frequent extreme weather events, rising sea levels, and threats to core infrastructure. Population migrations strain cities, and rising costs of living threaten our health and well-being, diluting the economic diversity that give our cities their unique character.

Lower-income communities are disproportionately affected - and while international leaders acknowledge this, pledged funding to address climate damages remains elusive. Furthermore, our world becomes increasingly interconnected and travel becomes accessible to a wider population, with projections that a quarter of all carbon emissions could be from flying by 2050.

Though there’s a lot we can be optimistic about. Emerging technologies, evolving urban design and a growing emphasis on citizen engagement are driving positive change. The potential of AI technology, alongside other data-driven tools, holds real promise to improve urban planning and policymaking to drive better results for the public.

With a flexible, compassionate, and multidimensional approach, our communities can evolve to support both larger populations and enhanced living conditions. Trade-offs between growth and sustainability were once seen as inevitable and today, we’re finding innovative ways to overcome these perceived limits with a renewing relationship with nature.

Our challenges are monumental, but they should get easier with increased awareness and collective effort.

To stay updated, I recently attended the Product Operations Festival hosted by the Product Led Alliance, an online product management community. They hosted helpful speaker sessions with leaders sharing expertise from their experiences working at prominent tech companies.

Here are some common themes I observed, which resonate with experiences from my own work:

• Protect your time: often as a team of one, firmly prioritize your tasks and know when to say no. Delegate responsibilities where possible, and establish a clear system for tracking requests and progress that you can share with others.

• Team Cohesion: a shared understanding of product and company goals is necessary for effective alignment across the organization. Build trust through proactive communication and establish feedback channels that ensure inclusivity.

• Do more with less: with limited resources, finding and implementing efficiencies for multiple fast-moving teams is essential for role success.

Unlocking Efficiency with LLMs

Product Operations professionals analyze large amounts of varied data, including transaction data, customer feedback, product requirement documents, standard operating procedures, and more.

Large Language Models can process this complex information at scale and offload many time-consuming tasks; this empowers product teams to work on more strategic tasks and get the maximum value from their data, a key competitive advantage for companies now and in the future.

Hence, it looks like we can do more with less.

Let’s take a look at standard LLM use cases outlined on Nvidia’s website:

1. Generation (e.g., story writing, marketing content creation)

2. Summarization (e.g., legal paraphrasing, meeting notes summarization)

3. Translation (e.g., between languages, text-to-code)

4. Classification (e.g., toxicity classification, sentiment analysis)

5. Chatbot (e.g., open-domain Q+A, virtual assistants)

For each category, there are plenty of opportunities to further product operations outcomes:

Generation

• Draft communications announcements, requirements documents, training materials, etc.

• Brainstorm ideas for strategy, create frameworks for problem solving.

• Write a Google Apps, Python script to automate simple tasks.

Summarization

• Condense user feedback, meeting notes, and research reports into key insights.

• Explain complex technical specifications and problems to non-technical stakeholders.

Translation

• Translate documentation across languages for global products.

• Better research markets around the globe with diverse languages.

Classification

• Analyze customer support tickets and triage them more effectively by type, sentiment, etc.

• Better flag spam and fraudulent activity in daily operations.

Chatbot

• Internal knowledge management - context-specific chatbots can provide employees access to company data quicker.

  • ‘Agentic workflows’ are an exciting new development that allow LLMs to not only generate, but do as well. Starting with automating simple tasks, it has potential for much more.

LLMs and AI-powered tools offer innovative ways to optimize all stages of the product development lifecycle. The wide proliferation of “.ai” domain websites reflects our growing interest in AI innovation, including additional use cases for generative AI like image and audio tools.

As the technology matures, companies will need to adapt and product teams must carefully weigh factors such as data privacy, ethics, potential for bias, and maintenance costs when choosing between pre-built and custom LLM solutions to ensure responsible implementation.

Increased accessibility suggests that building custom tools will become a cost-effective and practical strategy.