Unpacking GPAI Summit 2023

For those who couldn’t attend the Global Partnership on Artificial Intelligence (GPAI) Summit 2023 or those who attended but became engrossed in exploring the expo, networking, and organizing (like myself), missed a few sessions, couldn’t make it to all the three days, etc. Long story short, if you’re an AI enthusiast who couldn’t catch all the action across the 3-day GPAI Summit, I’ve got you covered.

It’s going to be a bit of a lengthy read, so let’s make a deal: you give me 20 minutes, and I’ll give you all the learnings and insights from the summit that Prime Minister of India inaugurated & was attended by delegates from 29 member countries, representatives from international organizations like OECD, UNESCO, ISO, WEF, World Bank, UNDP, Commonwealth, etc., along with 150+ global AI experts, industry leaders, startup founders, veterans from academia, AI practitioners, researchers, students, and officials from Central and State Governments across India. Sounds good? Let’s dive in.

Background

The Global Partnership on AI (GPAI) is a multi-stakeholder initiative established to bridge the gap between AI theory and practice, supporting cutting-edge research and applied activities on AI-related priorities. Launched in June 2020 with 15 members, GPAI originated from an idea within the G7. As of now, it has 29 members, including Argentina, Australia, Belgium, Brazil, Canada, Czech Republic, Denmark, France, Germany, India, Ireland, Israel, Italy, Japan, Mexico, the Netherlands, New Zealand, Poland, the Republic of Korea, Senegal, Serbia, Singapore, Slovenia, Spain, Sweden, Turkey, the United Kingdom, the United States, and the European Union.

In November 2022, India was elected as the Incoming Council Chair of GPAI, receiving more than two-thirds of first preference votes. India assumed the Lead Chair position on December 12, 2023, for the term 2023-24 and hosted the annual GPAI Summit, bringing together all 29 GPAI members to discuss AI-related matters. Here, I’ll try to skip as many names and designations as possible and focus on the learnings and insights that I gained from the deliberations at the summit. 

The discussions primarily focused on healthcare, Large Language Models (LLMs), and agriculture. We will delve into these topics first and then cover the solutions, innovations, challenges, and developments shared across other subjects throughout the summit.

Healthcare

The Global AI Healthcare Market size was USD 11.2 billion in 2022 and is projected to reach USD 427.5 billion in 2032. The top 10 AI trends identified in healthcare include healthcare analytics, medical diagnostics, telehealth, medical robots, hospital management, clinical decision support, clinical trials, public health management, cybersecurity, and personalized healthcare. The ‘Global Initiative on AI for Health’ is currently driven by the International Telecommunication Union (ITU), the World Health Organization (WHO), and the World Intellectual Property Organization (WIPO).

Dr. Ricardo Baptista Leite, CEO of HealthAI, shared HealthAI’s new strategic direction, which I found compelling:

• Build national regulatory mechanisms to validate AI Tools:
  • Each country they partner with helps build the capacity within the regulatory agency of that country
  • Specialized teams in AI, capable of defining standards internationally, are set up.
  • Then, these teams are trained and certified to ensure compliance with international standards.
• Establish a global regulatory network:
  • Currently, if something goes wrong with AI, we lack a mechanism to know about it promptly.
  • Implementing an early warning system that can raise a red flag if something goes wrong is essential
  • Resolving the volume issue of validating and approving various AI tools.
  • Facilitating a cross-country validation approach.
• Create a global public repository of validated AI solutions for health:
  • Providing global exposure to local solutions.
  • If an Indian regulatory body approves a technology, it will be automatically showcased to the world.

India heavily invests in the health sector, with notable initiatives such as bringing all health data together by Ayushman Bharat, the Health Technology Assessment system in the Department of Health Research, and individual-level initiatives like https://www.niramai.com/. Implemented in 15 districts in Punjab, 22 cancer patients have already been treated, and a Cervical cancer detection tool is in pilot testing in Gujarat.

Given the significant diabetes problem in our country and its negative impact on the eyes, even potentially damaging vision, startups are working to create images of the back of the eye. This innovation aims to make it easy for people with zero or minimal training, given the scarcity of ophthalmologists in India, to produce images suitable for running AI tools. The application of AI tools on these images enables accurate predictions of conditions such as diabetic retinopathy, glaucoma, and ARMD.

It was established that validation poses a challenge. How do we understand validation? As a user, what and whom do I trust? The evolution of the validation environment for India is going to be crucial. 

Then, four areas of AI intervention in healthcare were identified:

1. Surveillance for diseases, illnesses, pathogens, and pandemic prediction (We need to build such tools under one health mission in India).
2. Ensuring seamless logistics in hospitals – ensuring drugs, devices, and apparatus.
3. Research center development – describing diseases, clinical research, and describing prognosis.
4. Enable Electronic Health Records: Rules of writing an EHR. An EHR is written by a doctor, and with 14 lakh doctors, standardization is needed.

Under healthcare, a special focus was placed on neurotechnology. AI and neurotechnology complement each other. AI enhances neurotech’s ability to interpret complex neural data, and neurotech provides insights into human cognition that can inform AI development. Government investment in neurotechnology has surpassed $6 billion since 2023, as estimated based on available information. Private annual investment has increased 22-fold from 2010 to 2020, reaching $7.3 billion and totaling $33.2 billion by 2020.

Seven countries have national projects running on neurotechnology, namely the Canadian Brain Research strategy, US Brain Initiative, EU Human Brain Project, Japan Brain/Mind Project, China Brain Project, Korean Brain Initiative, and Australian Brain Alliance. 

Neurotechnology applications could revolutionize the diagnosis and treatment of neurological disorders and mental health issues, leading to the development of advanced brain-computer interfaces for more intuitive interactions with technology in healthcare and other areas. The potential for enhancing cognitive abilities like learning, memory, and decision-making is significant. Other applications include multimodal neuromodulation, brain-computer interfaces and limb rehabilitation, seizure prediction, sleep optimization, brain performance assessment, and advanced brain imaging.

Large Language Models (LLMs)

Another domain of key deliberations was LLMs. For those new to the AI domain, Large Language Models (LLMs) are advanced computer programs trained on extensive datasets to understand and generate human-like language. The term “Large” indicates the significant amount of data they learn from to mimic human communication. The biggest challenge with LLMs regarding India is the absence of large datasets available for training on Indian languages. There are 60 lakh articles in the English language and only 1 lakh in Indian languages. The English language also has Wikipedia as a database (extensively used by GPT), but the knowledge repository of India does not really exist in digital form. This lack of data also increases the cost per token, as the tokenizer doesn’t have enough Indian language content, so if GPT-4 produces an output of 100 words in the Hindi language, it is equal to 1000 tokens for it (which ideally should be 100). The model treats every Indian character as a token, and it is a problem because we pay per token.

To tackle this problem, two approaches were pitched. The first one was to take a powerful LLM and teach it more about India, instead of building things from scratch, so that it becomes more aware of the Indian context. It was proposed to take the best open-source model, add Indian language data to it, and continually pre-train. We need to feed Indian data in Indian languages so that the native capacity of LLM in these languages becomes better. This is in contrast to taking the English route or treating them as languages that just have character-level tokens. For the same, Project VeLLM (UniVersal Empowerment with Large Language Models, a multistakeholder project by Microsoft, for culturally complex countries like India is in action. 

Another solution pitched was the development of IndicNLP. Dr. Partha Talukdar, Senior Staff Research Scientist at Google Research, India, highlighted the vibrant landscape of NLP research in the country, where various institutions are actively engaged in work and categorized the efforts into four buckets.

1. Sanskrit NLP: The Computational Paninian Grammar Framework, detailed in a research paper authored by Akshar Bharati and Rajeev Sangal, explores supertagging, presenting a unique method of combining linguistic and statistical information.
2. Multi-lingual & Indic NLP: In the multilingual and Indic NLP domain, there are 22 scheduled languages, 1300+ languages, and 60+ languages with 1M+ speakers, reflecting the complexity of a multilingual society with code-mixing challenges. Open-source multilingual language models have been developed to support multiple languages within a single model, finding application in various contexts. The IndicNLPSuite research paper delves into Monolingual Corpora, pre-trained multilingual language models tailored for Indian languages, and presents an evaluation benchmark.. 
3. Knowledge & Information NLP: It focuses on knowledge and information extraction through techniques such as assigning tokens using Markov Conditional Random Fields. Many organizations have contributed to this area. India has taken the lead in Open Information Extraction (OpenIE) with initiatives like Iterative Grid Labelling and Coordination Analysis for open Information Extraction receiving support from IIT Delhi and IIT Bombay. 
4. Low Resource Ecosystem NLP: The challenge lies in the scarcity of content in Indic languages on the internet. To address this issue, Project Vaani, a collaborative effort by IISc, Bangalore, and ARTPARK, funded by Google, aims to capture the true diversity of India’s spoken languages. The project involves capturing the speech landscape in India across 773 districts, 154K speeches, 15.4L transcribed, and approximately 1 million participants. This is being done in a region-anchored manner for effective anchoring.

To enable both the solutions mentioned above, the best repository or database that was pitched was All India Radio (AIR). AIR broadcasts three bulletins every day in 40+ languages. These are news bulletins where a person is talking, and there are transcripts of the same. We need to collect these transcripts for the last 20 years and feed them as Indian knowledge to LLMs. Another valuable source is Supreme Court judgments locked in Indian languages in PDFs.

Agriculture

While discussing agriculture-related challenges and solutions, numerous innovative ideas emerged, and individuals along with startups actively working on the ground with farmers shared their experiences. Solutions that caught my eye include

1. Surabhi ID: Out of the 380 million cattle in the country today, less than 4% are insured. The nose of a cow, called the muzzle, is as unique as a human fingerprint. Therefore, Dvara Surabhi is utilizing this distinctiveness to create a cattle health fitness certificate. They aim to replace invasive RF ID with muzzle scanning, contributing to the enhancement of cattle insurance.
2. Google Agriculture Initiative: The Google Agriculture initiative aims to organize agricultural information worldwide at an individual farm field level which will serve as a foundational layer for stakeholders in the agricultural ecosystem. More information will be available soon on www.agri.withgoogle.com 
3. Google & Protean Collaboration: The collaboration between Google and Protean has resulted in the development of the Pro Kisan App. Powered by Google Cloud generative AI, this app automatically generates catalogues for farmers’ produce. It provides advisory services throughout the crop lifecycle in regional languages with voice capabilities.
4. Cattle Health Monitoring Solution: India has a cattle population of 380 million (30.8 crores), consisting of about 53 breeds. Unfortunately, valuable knowledge about these breeds is diminishing with each passing generation. Areete offers a comprehensive solution for cattle care across breeds and provides all the necessary farm equipment through its portal.

Solutions built/being build across other domains

1. Language Translator in Japan: Mr. Hiroshi Yoshida, Vice-Minister for Policy Coordination at the Ministry of Internal Affairs and Communications, Japan, shared that a Language Translator is being developed for 17 languages in Japan to overcome language barriers. This translator, compatible with smartphones, will feature Automatic Translation and Simultaneous Interpretation. The work is ongoing, and a prototype demonstration will be made in 2025.
2. Continental Training Datasets: There are four places globally for continental training datasets, of which two are unlocked (US & China), and two remain to be unlocked (Europe & India). India’s approach to unlocking continental datasets for training models, allowing Indian companies to create world-class models, will be enabled by the DEPA Training cycle. DPI, called DEPA, has two cycles: the training cycle and the production cycle. The production cycle is rolling out soon, as shared by Mr. Sharad Sharma, Co-Founder, iSPIRT Foundation.
3. “AlphaFold” by Google: “AlphaFold,” developed by Google, has successfully constructed 3D structures of proteins through AI, providing 400 million years of progress in a matter of a few weeks.
4. Deepfake Precaution: To mitigate risks in driving AI innovation, Google has developed the “About this image” tool. This tool helps determine the authenticity of a specific image by uncovering relevant facts.
5. Flood Forecasting by Google: Google’s research team has created a technological flood forecasting model that predicts floods seven days in advance. The forecasting process utilizes machine learning and physics stimulation, incorporating two models: the Hydrologic Model (estimating water flow in a specific location) and the Inundation Model (predicting water levels in a place). They have now introduced a Global Hydrologic Model that learns hydrology patterns through machine learning. This allows predictions even in locations with limited historical data. The model is currently operational in flood hubs across over 80 countries, including 23 countries in Africa, where data is scarce.
6. Traffic Light Optimization – Project Green Lights by Google: Google’s Project Green Lights focuses on optimizing traffic light schedules at intersections. Given that carbon emissions at traffic intersections are 29 times larger than on open roads, data analysis without sensors and AI optimization can reduce stop-and-go events by 30%. This reduction translates into a 10% decrease in carbon emissions.

AI implementation

Once AI development is figured out, implementing it remains a challenge. In the realm of AI implementation, adopting a customer-centric approach is paramount. Think about one specific person and what they need. Craft your product with a specific individual in mind, delving into the context by posing relevant questions about the topic. The deeper the relevance, the higher the likelihood of user adoption. However, adoption is merely the initiation; it just signifies a willingness to engage, nothing more, nothing less. But once they do, pay close attention to what they say. Think about where & what can be improved, and keep working on making it better. Doing this over and over makes the AI better and builds trust with the people using it.

To facilitate AI implementation, tailor your messaging strategy to different stakeholders. Pitch the idea of inclusion to policymakers, advocate the democratization of technology to technologists, and promote market expansion to business professionals. By weaving these narratives around a common set of principles, it becomes possible to bring these diverse perspectives to a harmonious convergence.

Solutions Shared

1. Alignment on Standards: Work is currently underway in ISO and sub-committee 42 on AI, focusing on the development of a risk management framework.
2. Common Regulatory Approaches: Emphasis should be placed on adopting a risk-based approach rather than a one-size-fits-all strategy.
3. Cooperation Around Research: Introduction of micro-specialization in AI for students, integrating computational aspects into various subjects to enhance subject-specific knowledge.
4. Reliable Detection of AI-Generated Content: Challenges arise in differentiating between machine and human-created content, necessitating reliable detection tools. Proposed solution: Companies producing AI generators must demonstrate and supply a reliable detection tool as a condition for public release.
5. Data Transfer and Consent (Health Pass Book): A concept similar to a health passbook in Taiwan, enabling consented data transfer to third-party apps via a government-run health passbook.
6. Talent Development (Industries and Academic Institutions): Industries and academic institutions are pivotal in developing future-ready talent, with government support limited to shaping the talent pipeline.
7. Making AI Accessible: Efforts should be directed toward creating lightweight AI for future use on smartphones/laptops, particularly in medical applications, for energy efficiency.
8. Resilient Ecosystem for Data Collaboration: The focus should be on creating a resilient ecosystem for data collaboration rather than mere data sharing.
9. Energy Consumption and Sustainability: Minimizing energy consumption and waste from servers is vital for sustainability. Transitioning from air-cooled servers to liquid cooling and eventually to immersion cooling, using oil to efficiently cool servers without additional energy consumption is essential. 
10. AI Lifecycle Challenges (Model Monitoring): Building an AI algorithm is not the most challenging aspect of the AI lifecycle. Deploying and monitoring it is. Continuous monitoring is crucial: tracking variables, identifying inputs not seen in the training data. The area with significant scope for innovation is tools for model monitoring, an emerging field where most companies currently rely on in-house tools. If specific model monitoring tools become available, they can enhance efficiency across the ecosystem and facilitate easier compliance with new standards.

Hidden Challenge

The labor component of the AI supply chain remains a poorly understood and often overlooked aspect. While the tech sector invests billions of dollars in data collection, the true workforce—the data workers—often operate without the recognition they deserve. It’s essential to forge new pathways that establish a fair market for labor while leveraging the significant momentum to construct ethical AI practices that address labor issues. 

Prioritizing fair wages for workers is crucial, as higher wages result in superior data sets and substantial enhancements in AI models. Empowering workers with ownership of their contributions can yield phenomenal AI data sets. Additionally, addressing the challenge of significant salary disparities between the private and public sectors is essential.

Regulations Across The Globe

Countries exhibit fundamental differences in their approaches to AI regulation, broadly classified into three models:

• Horizontal Regulation (e.g., European Union, India):
  • Scope: Applicable to all AI applications across various domains and sectors.
  • Governance: Regulated by a central government body, functioning in addition to sector-specific regulations.
  • Enforcement: Stringently enforced by the designated authority.
• Vertical Regulation (e.g., US, UK):
  • Scope: Limited to specified industries or applications of AI.
  • Governance: Enforced by sector-specific authorities, operating at the federal, state, or private citizen level.
  • Enforcement: Implementation varies based on the specific area of law.
• Non-binding Guidelines (e.g., Japan, Singapore, Israel):
  • Nature: Non-binding standards with broad definitions.
  • Governance: Lack strict enforcement mechanisms; adoption is voluntary.

These varied regulatory models highlight the nuanced approaches taken by different nations to address the challenges and opportunities posed by AI technologies.

International Business Scenario

Amidst diverse regulatory approaches, the international business scenario grapples with distinct challenges, including varying AI definitions, diverse regulatory applicability across use cases, and disparate compliance requirements. For a country aspiring to market its products globally, navigating and aligning with regulations from different countries pose significant hurdles. The critical question emerges: How can a company ensure compliance with diverse regulatory landscapes worldwide?

To address this challenge, the GPAI working group is actively engaged in a comprehensive initiative. This includes mapping diverse regulations globally, conducting surveys involving affected stakeholders, pinpointing relevant fields of compliance, identifying support mechanisms worldwide, and creating an extensive library cataloging existing approaches, activities, and resources. The ultimate goal is to construct a globally approachable platform, facilitating seamless compliance for businesses operating on an international scale.

Key Developments across the Globe shared-

1. Council of Europe’s Binding Treaty on AI: The Council of Europe is taking a pioneering step by developing the first binding treaty on AI (The Council of Europe is not the European Union; the CoE has 46 member states, whereas the EU has 27 member states.)
2. UN’s Global Digital Compact: The United Nations is in the process of formulating the Global Digital Compact, a document at the leader’s level addressing various digital challenges and opportunities, including those posed by AI.
3. AI4SME Portal Initiatives: The AI4SME Portal, launched in Singapore and Poland, and set to be launched in France, Germany, and Serbia, is designed to enable Small and Medium Enterprises (SMEs) to enhance their businesses through AI. This platform involves three key stakeholders: SMEs, who are looking to improve their businesses through AI adoption; Solution Providers, who offer solutions and aim to expand their user base; and Portal Operators, responsible for managing the portal and onboarding solution providers.

Books recommended

In all the discussion and cross-pollination of ideas, three books were recommended by people whose names I forgot to note down because I got busy googling the authors of the books they recommended, Here are the books:

1. “Power & Progress: Our Thousand-Year Struggle Over Technology and Prosperity” by Daron Acemoglu and Simon Johnson.
2. “The Coming Wave: Technology, Power & 21st Century’s Greatest Dilemma” by Mustafa Suleyman and Michael Bhaskar.
3. “Tools & Weapons: The Promise & Peril of the Digital Age” by Brad Smith.

Conclusion

In summary, the summit underscored the imperative for collaboration in healthcare, innovation in agriculture, and the tackling of challenges posed by Large Language Models, especially in the Indian context. The stress on a customer-centric approach to AI implementation, along with key solutions such as alignment on standards, demonstrates a commitment to responsible AI practices. With a global focus, the summit positions AI as a transformative force with the potential for positive change. The overarching challenge for the international community is to bridge the gap between defining high-level principles and establishing accountability chains, fostering trust in AI technologies worldwide for their responsible and inclusive deployment.