In my role as field CTO, I sometimes speak with analysts and the media. Conversations often focus around a predetermined area of interest of set of questions. The follow text is based on my preparation for an interview with Techgoondu, Sinagpore.
Q1. What’s the Asia-Pacific AI landscape like currently? What AI technologies are enterprises adopting and where do they stand in relation to other regions?
A1. In APAC, almost 90% of organisations are using or plan to use AI applications during the next 12 months. During this period, AI spending will grow by almost 25% CAGR and is expected to reach US$49.2 billion by 2026.
Within the field of AI, we are seeing particularly strong growth and interest from customers around Generative AI during the last six months. Organisations are looking for ways to safely and economically get value from Large Language Model (LLMs) across multiple lines of business.
Recent polls indicate that close to two-thirds of organisations in the region have either invested in or plan to invest in Generative AI by the end of 2023.
Generative AI and LLMs are creating so much interest due to the almost unlimited number of potential applications across every function within an organisation. They are being used for everything from improving developer efficiency, to providing analysts with summaries of complex dense reports and improving the efficiency and effectiveness of customer call centres.
While the potential for use cases for LLMs is almost limitless, organisations are proceeding cautiously to protect sensitive information. Usage policies are being carefully developed and self-hosted LLMs are increasingly being deployed to complement the consumption of SaaS based LLMs.
In parallel, organisations continue to focus on ethical and responsible AI with governments and regulatory bodies playing an increasingly important role.
Q2. What are some of the key risks and limitations of AI today, in relation to enterprise use of this technology? Can you explain the impact of these risks for an organisation?
A2. There are several risks or concerns associated with the adoption of AI. They include ethical issues with bias and discrimination being of particular concern. They also include data privacy, data security, transparency and explainability, societal concerns around job placement and the accuracy and relevance of answers or guidance from AI systems.
AI systems certainly do have limitations, but with rapid advances within Generative AI, limitations are constantly being overcome. A few years ago it was inconceivable that we’d be able to engage in a natural language conversation to perform complex tasks or create valuable original content.
Current limitations for LLMs include:
- Limited performance and usefulness when asked to generate content that’s unrelated to the data in its training set.
- Models may generate content that’s repetitive
- Without suitable context, results may be less relevant
- Model behaviour may lack explainability and transparency due to complexity of the underlying structure
- Models may demonstrate bias
The impact of these risks are real and need to be managed. If an AI model provides the wrong information to a customer, it can impact an organisation’s brand and service reputation.
Bias can lead to several bad outcomes. It may lead to lost service or product revenue and in the worst case have legal consequences. There have been cases of bias influencing credit limits and insurance policies to name but two examples.
Some processes required explainability and transparency. If decisions are made within an opaque black box, it may introduce risk that can violate industry guidelines and data protection regulations. There was a case in the Netherlands of workers that were dismissed without either suitable human intervention or transparency and explainability in AI supported processes. The employer was found to be in violation of article 22 of GDPR.
Q3. What can organizations do to mitigate such risks?
A3. Let’s focus on the risks of data privacy, contextual related performance and ethical or responsible AI.
To mitigate data privacy risks, organisations need to classify data and provide clear guidelines on usage and use technology to help enforce it. For example, while it may be acceptable to use a SaaS-based LLMs to analyse and summarise a report that’s already in the public domain, the policy will almost certainly prohibit the use of a public SaaS service to analyse a sensitive internal document. Similarly, LLMs that help developers document or explain blocks or code may be acceptable, but if an engineer is working on technology that’s highly sensitive and protected, pasting a chunk of code into an external SaaS service will almost certainly violate organisational data management policies.
But policies, guidelines and technology to control data privacy is only part of the solution. Organisations need to augment SaaS based solutions with their own privately hosted solutions that provide comparable performance.
In terms of contextual relevance and performance, organisations need to ensure that suitable context is injected into the conversation (prompt) with LLMs they use. A reliable way to do this is to carefully control access to the prompt and programmatically inject relevant context. Context is typically provided from a suitable knowledge base and through a process called embedding. This contextual augmentation of knowledge is often referred to as Retrial Augmented Generation (RAG).
One of the key foundations to trusted, contextually relevant data and AI is strong data management controls across the entire data lifecycle. This is a core focus of the Cloudera Data Platform (CDP). When combined with Cloudera Machine Learning (CML), organisations can deliver trusted AI on premises or in the cloud while still maintaining full control over their data assets.
Responsible or ethical AI is multifaceted, with bias being a significant element. In order to address bias, we need to first understand the bias in the training data. This can be done by analysing the data for patterns of unfairness or discrimination, but of course this is only possible if you have access to the training data set. When using a pre-trained model, organisations need to understand in-built biases in the model. This is difficult to do and may require referring to analysis on how the models behave.
Responsible or ethical AI is a large and complex topic, so I’d recommend that organisations connect with the governing body within their industry. For example, financial institutions within Singapore should look to MAS for guidance.
Q4. I understand that trust is critical for AI, why is that so, and how can organizations build trustable foundational data and models?
A4. Trust is a foundational element of any service whether it be customer facing or internal. Trust is at the foundation of being a data driven organisation. If you do not trust your data, how can you confidently make decisions or provide services based on it?
Trusted AI is the result of having strong controls over your data across its lifetime, from the edge to AI. We need to carefully control who or what has access to the data and especially how the data is transformed or changed over time.
Data Lineage and provenance is required at each stage of the model training and augmentation process. When consuming closed-source AI models, especially LLMs, organisations will have limited access to the specific details of how models are trained and what data was used. This introduces risk and unpredictability that needs to be mitigated through careful testing, monitoring of performance and suitable context.
Q5. What are some best practices for organizations to take note of when adopting enterprise AI?
A5. I would start by having a clear usage policy, strong data management controls and scalable, reliable approach to MLOps:
The foundation of any analytical use case, be it traditional data warehousing or predictive analytics as provided by AI is strong data management. AI models, in particular ML and Deep Learning (DL) models generally perform better with more good quality data. Even the most sophisticated AI algorithms cannot compensate for poor training data. Even if we start from a training model, it often requires contextually relevant data to be useful for a specific purpose. A strong data management capability is therefore crucial.
Next, it’s important for organisations to have clear data usage policies. This requires data, algorithms, models and services to be classified and approved for certain use cases. Personnel need to be educated and updated regularly as policies are revised to accommodate new capabilities. For example, SaaS based developer productivity services may be restricted to only a subset of non-sensitive development projects. However, with the introduction of a privately hosted LLM based on the Open Source StarCoder LLM, the policy is extended to include this new capability for sensitive development projects.
When it comes to AI models, it’s important to understand the licensing and fair usage policies. The specifics should be discussed with suitable legal counsel.
As previously mentioned, data ethics and responsible AI should be influenced by the relevant industry body.
Lastly, most AI models struggle to get out of the lab and into production efficiently and at scale. This is the focus of MLOps. It’s a multifaceted domain but one that’s at the foundations of Cloudera Machine Learning (CML). MLOps covers everything for data exploration, data engineering, model training, model tuning and subsequently making those models available for consumption. It also includes the process of monitoring model performance and retraining of models when appropriate with suitable human oversight.
Q6. What are enterprises in Asia-Pacific doing to prepare their business to be “AI-ready”?
A6. From a data perspective, organisations are focussed on the three areas I just mentioned. When I think about organisations that are leading in this space or were able to rapidly take advantage of recent developments in LLMs, they all have strong, mature data management and data platform capabilities.
For example, OCBC bank, the winner of this year’s Cloudera “Best Transformative AI” award, has talked about the reason they were able to quickly incorporate LLMs into their on-premises business because they carefully laid the data platform foundations several years ago. Within only a few days they were able to replace existing developer code assist tools with privately hosted services based on the open source StarCoder LLM. In doing so they reduced operating costs by 80% and made the service more contextually specific with their own coding standards. They were also able to efficiently roll out new AI augmented services to help customer support staff via speech-to-text and sentiment analysis and document summarisation that reduced the effort to analyse documents by 95%.
Developing a strong technology capability is a necessary, but in complete step in preparing to be AI-ready. Building strong Data Science and Data Engineering skills is necessary to take full advantage of the algorithms and models available today. OCBC started their journey four years ago with a four-person team and have grown quickly and organically since then. They are a good example, that great results can be delivered with a modest initial investment and a focus on delivering business value as a gating function to future growth and investment.
Due to the computational resources required to both train models and perform inference, organisations are also looking to the cloud for flexible payment options for variable, unpredictable and bursty workloads. Easy access to specialised, expensive hardware as required by AI continues to push the hybrid cloud agenda.