Future of Generative AI: A Frontline Practitioner’s Take on Adoption Trends
Introduction
We’re living in a time where terms such as ‘Generative AI’, ‘large language models’, and names like ‘ChatGPT’, ‘Llama’, and ‘Bard’ are popping up all over our newsfeeds and sparking countless discussions. The innovative technology surrounding us is profound and the potential uses for these technologies are staggering. However, along with these opportunities come heated debates about the future implications of Generative AI, and the potential risks it could introduce into our lives.
In the past decades, my career intertwined with the evolution of data and analytics space, and I’ve been fortunate to have a front-row seat to these transformations. This unique position has allowed me to formulate my perspectives about the future landscape of data analytics, especially with the emergence of Generative AI.
In this post, I’d like to share my insights on the following topics:
- A brief history: The evolution of data analytics over the past decades.
- The road ahead: Projecting the future landscape of data analytics in the era of Generative AI.
- Building bridges: How can we establish trust and mitigate risks in Generative AI applications?
- Forward-thinking: As a Chief Data Officer or team lead in data analytics, how can you prepare for the upcoming transformative journey?
Evolution of data analytics
Data and analytics isn’t a fresh idea — it dates back to ancient times. We could spend all day discussing the basics, but for this post, let’s dive into its journey since the big data era.
Stage 0 — The Data Lake Era
Most businesses initiated their modern data analytics journey with the creation of a data lake, a vast reservoir storing tables, schemas, and databases that are crucial to their business operation. In its infancy, the industry-standard technology was the Relational Database Management System (RDBMS), which was restricted to handling structured data like numbers, attributes, and dates. A big downside was the cost of maintenance, which increased alongside the growing data volumes. Fortunately, the rise of modern cloud-native data warehouses helped address scalability and cost-effectiveness issues.
There’s a humorous term for when a data lake gets out of hand — ‘data swamp’. At this stage, data analytics teams often find themselves stuck managing data before they can even begin the real analytical work.
Stage 1 — Descriptive Analytics
With the data neatly stored and managed, it’s time to dig into the wealth of information it can reveal.
Business Intelligence (BI) tools are your best friends here — they can summarize and visualize data, helping us understand what happened in the past. Modern tools like Tableau, PowerBI, and Qlik offer excellent dashboard capabilities and user experiences. They enable business analysts to better comprehend their data and present compelling insights to decision-makers.
Sometimes, BI tools can’t directly consume data from a warehouse due to factors like formatting or access restrictions. That’s where Robotic Process Automation (RPA) steps in, acting as a preliminary data processor for BI tools.
RPA can also be very useful for tasks that are repetitive and manual, like reformatting Excel sheets as per specific guidelines, or sending identical emails to multiple recipients. It streamlines operations and helps save significant time and effort.
Stage 2 — Predictive Analytics
In the descriptive analytics stage, business insights are distilled and handed over to decision-makers. But what if we could use tools not only to generate insights but also to make decisions?
Welcome to the era of predictive analytics, where machine learning is the star of the show. Machine learning uncovers hidden factors within historical data that can predict future outcomes, acting as a powerful tool for businesses to stay ahead in their industry. It augments decision-making processes, enabling businesses to make better and quicker decisions, and swiftly adapt to market changes. And it doesn’t just provide outcomes — numerous model explanation techniques help demystify machine-driven decisions and offer actionable insights to human decision-makers.
Deep learning, a subset of machine learning, is a key driver in predictive analytics. It uses artificial neural networks with many layers (hence ‘deep’) to understand complex patterns in data. Capable of handling diverse data types like images and text sequences, many breakthroughs in predictive analytics are in the deep learning domain. In fact, deep learning forms the foundation of generative AI, which we’ll be exploring in our upcoming discussions.
It’s important to clarify that predictive analytics is not intended to replace humans in decision-making. Instead, it serves as a more objective and intelligent benchmark, enhancing our ability to make better decisions.
Future Outlook: The Augmented Data Analytics Landscape Shaped by Generative AI
The rise of generative AI has unveiled a new frontier for data analytics, offering the ability to learn from vast internet resources and generate content — from text and images to music and voices, although still a very new concept, with few large-scale corporate success stories. As an AI practitioner in the data analytics space, I’d like to share my perspective and outlook of how Generative AI will evolve. and I’m keen to hear your thoughts as well.
Augmented Data Layer with Generative AI
Generative AI increases the scope of data available for analytics, with all types of content — documents, files, images, voice, and videos — now potentially serving as input.
I believe the true enhancement of the data layer comes with foundation models. These large-scale models, trained on an extensive and diverse dataset, offer a more powerful means of data access. Examples include language models like GPT-3, Llama, and text-to-image models like DALL-E.
To understand the impact of foundation models on data layers, consider this analogy: Just as the modern data warehouse allows us to understand data intuitively and structurally instead of deciphering bits, foundation models represent another leap forward in how we access and interpret data.
Augmented Analytics Layer with GenerativeAI
In my perspective, we can still classify the types of analytics into descriptive and predictive, albeit with a fresh interpretation in the context of Generative AI
Note: Given that the essence of all GenAI-related analytics revolves around content creation, I won’t separate content creation as an individual category.
i. Generative AI for Descriptive Analytics
The first wave of Generative AI adoption will likely focus on descriptive analytics. A prominent application is the Knowledge Base, which uses foundation models to enhance data accessibility and interpretation. With this approach, you can collate information from various sources and receive tailored responses to specific queries.
Early adopters, like Morgan Stanley Wealth Management, have begun using GPT-4 to manage their extensive knowledge bases. (source)
Another promising field is Robotic Generative AI Automation. For instance, when creating this article, I used Midjourney to generate all the images. Crafting effective prompts for Midjourney is an art in itself, but I found that using GPT-4 to provide concise, comprehensive explanations greatly helped to visualize the images in my mind. Robotic Generative AI Automation could integrate multiple applications and orchestrate tasks which will exponentially boost your Generative AI productivity
ii. Generative AI for Predictive Analytics
As discussed previously, predictive analytics provide a more objective and intelligent benchmark for human decision-making. Generative AI elevates this capability to an entirely new level.
By harnessing Generative AI with industry-specific knowledge, organizations can develop virtual advisors offering tailored advice and recommendations to their customers. For example, Bloomberg recently launched BloombergGPT, a 50-billion parameter large language model designed specifically for finance NLP tasks(source). Creating such a model from scratch requires significant computational resources, time, and expertise. An alternative, more feasible, yet still costly approach involves ‘fine-tuning’ foundation models using industry-specific data.
Another area that I personally devoted lots of time on is to leverage Generative AI (especially ChatGPT) to make zero-shot predictions. Imagine asking ChatGPT to determine if a customer’s behavior is suspicious based on a prompt detailing their actual and expected transactional activities. In one of my blog posts, I conducted such an experiment, using DataRobot(Automated AI/ML Platform) to replicate the ChatGPT-driven decision and highlight the key factors that influenced it.
I understand that I’ve used quite a few Generative AI-specific terms in our discussion. If you’d like a more simplified, ELI20(‘Explain-Like-I’m-20’) style of these concepts, you can check out this blog post.
Risk, Governance, and Trust: Looking Forward with Generative AI
While this post mainly focuses on the exciting opportunities of Generative AI, I’ll briefly touch on my views regarding its inherent risks and governance. I’ll delve into this topic more deeply in a forthcoming post.
Governance is not a novel concept in the realm of data analytics. We already have regulations like the GDPR (General Data Protection Regulation) for data access and privacy, and frameworks like SR 11–7 from the Federal Reserve Bank for model risk management.
Even the most fundamental statistical analyses and visualizations are not risk-free — take Simpson’s Paradox for example — even the manner of slicing data could dramatically alter data interpretation.
That said, Generative AI introduces unique risks and challenges. These arise due to its cutting-edge nature and the adoption of new technologies.
Potential Risks and Mitigation Strategies
- Data Privacy and Security: Large language models can inadvertently access and generate sensitive data. To mitigate this, enterprises should integrate comprehensive data privacy and security measures when deploying such models. Also Enterprise grade LLM integration, such as Azure OpenAI could also help mitigate such risks.
- Bias: If a model’s learning data is biased, the generated output may also show bias. Employing bias detection and fairness assessment tools can help identify and correct such biases.
- Model Explainability: Understanding how a model makes its decisions is crucial. The development and application of model explanation techniques, including those for large language models, is vital. Some methods may leverage existing machine learning explanation techniques, while others may require novel approaches.
- Monopolies due to Access to Compute Resources: Only a few organizations with extensive resources can train large models, potentially leading to monopolies. To prevent this, governments and industry bodies must establish a regulatory framework that fosters fair competition.
Advice for Data and Analytics Executives
Having led teams of data scientists for many years and working closely with executives in data, analytics, and machine learning domain, I’ve gathered some insights that may help you prepare for your Generative AI journey:
- Empower Your Talent: Fear and hesitation often stem from not fully understanding new technologies. Encourage your analytics team to learn about Generative AI. This access to knowledge will break down barriers and cultivate growth in the long term.
- Start with Descriptive Generative AI: Beginning with knowledge bases and smart searches can offer cross-functional benefits with less operational and reputational risk. As we’re still in the early stages of this field, teaming up with an established company with applied AI expertise can give you an advantage.
- Keep an Eye on Regulations: Stay updated with ongoing discussions about regulations. Even better, join these conversations and contribute to the formation of industry standards on Generative AI with other thought leaders in the sector.
© 2023 Ray Mi. All rights reserved. Unauthorized use or reproduction of this blog post without express written permission from the author is strictly prohibited. Information is provided as-is, without warranties. Views are those of the author alone.