Leonidas Drakopoulos: As GenAI continues to scale and evolve, the industry must responsibly develop it

As with all novel technologies, Generative AI's potential for positive disruption is equalled by the possible abuses that it can be used for. GOLD magazine spoke with Leonidas Drakopoulos, Amazon Web Services (AWS) Tech Lead for Greece, Cyprus and Malta, to find out how the cloud services giant is leading the Generative AI charge while developing the technology responsibly.

To catch the Generative AI (GenAI) wave, Amazon has splashed out US$4 billion on Anthropic – its largest venture investment to date – while AWS has committed US$100 million for a centre to help companies use the technology. How does your strategic approach and focus differ from that of your competitors?

Indeed, the AWS Generative AI Innovation Centre connects companies with our Machine Learning (ML) and AI experts to help envision, design and launch new GenAI products and services. Amazon and AWS have been innovating in AI and ML for over 20 years, including the recommendation engine on Amazon.com, optimising robotic collection routes in Fulfilment Centres, large-scale product demand forecasting, Prime Air drone deliveries and the deep learning technology behind Amazon Go, where consumers can pick items off the shelf and simply leave the store.

Our main goal is to democratise AI, ML and GenAI technologies. We have more than 100,000 customers of all sizes and industries who trust and use our AI and ML technologies – including Intuit, AT&T, Thomson Reuters, AstraZeneca, Ferrari, Bundesliga, 3M and BMW – as well as thousands of startups and government agencies around the world.

By bringing Anthropic into the fold, the startup now primarily uses the AWS cloud, including Amazon Bedrock, unveiled last year. Can you briefly explain what Amazon Bedrock does and how your customers approach GenAI with the solution?

Amazon Bedrock is a fully managed service offering high-performing foundation models (FMs) from leading AI companies. Customers use it to create GenAI applications and solutions, with tens of thousands choosing it for their GenAI strategy. It provides access to models from AI21 Labs, Anthropic, Cohere, Meta, Mistral AI, Stability AI and Amazon, along with the necessary capabilities and enterprise security. It ensures seamless deployment, scalability and continuous optimisation, freeing customers from concerns about the underlying infrastructure. We focus on understanding the challenges and needs of our customers and then working to identify integrated solutions.

So, we always try to create services based on these needs and anticipate them to innovate before our customers even ask us to. For example, organisations of all sizes and different industries want to start by creating artificial intelligence to transform their operations, to rethink how they solve difficult problems and create new user experiences. Many businesses are eager to create solutions but are concerned about security and privacy.

They also want a variety of FMs to test and meet their needs while customers seek to maximise their data and fine-tune models for unique user experiences. Amazon Bedrock addresses these needs by offering robust security and a wide selection of models and capabilities. It helps organisations leverage their data strategy to fully utilize GenAI technologies.

Let's take a step back. A recent paper by Anthropic, "Sleeper Agents," revealed that LLMs can maintain deceptive behaviours despite extensive safety training. Can you provide insights into current efforts to address and mitigate this critical issue?

As GenAI continues to scale and evolve, the industry must responsibly develop it. The models fuelling GenAI are vast, open-ended and fluid – just one model can encompass billions of parameters – and there is active science underway to address some of the questions around accuracy, privacy, bias and more. We are committed to developing technology responsibly with a team of dedicated responsible AI experts, complemented by AWS engineering and development teams that continually test and audit our products for fairness and accuracy.

We are working alongside others like the OECD AI working groups, the Partnership on AI and the Responsible AI Institute to develop new approaches and solutions to identify and mitigate bias, address privacy questions and provide the right safeguards. Our AI and ML services inherit AWS’ enterprise-grade security and privacy best practices. Within Amazon Bedrock we have implemented a functionality called ‘Guardrails’ that helps organisations implement safeguards customised to specific use cases and responsible AI policies, enhancing the safety and privacy of user interactions by blocking harmful content.

We will not use proprietary data for fine-tuning and customisation to train base models or share that data with model providers – we always protect our customers’ valuable intellectual property. And within Bedrock, we offer a model that helps organisations evaluate, compare and select the best models for their specific use case, using either automatic or human evaluations.

Another particularly concerning application of GenAI is the creation of deepfakes, which can even disrupt democratic processes like elections. What is being done to prevent such misuse?

Again, we are committed to developing technology responsibly. In addition to the above, we engage with organisations like the National AI Advisory Committee and collaborate with policymakers such as the US AI Safety Institute. Also, our US$20 million investment in the National Science Foundation’s Fairness in AI grants programme, alongside partnerships with academia, aims to advance responsible AI science. And we innovate with our customers, staying at the forefront of trends and research, providing grants via Amazon Research Awards and scientific publications with Amazon Science.

While developing FMs, we adhere to responsible AI trends at each stage of the comprehensive development process. We consider several factors such as accuracy, ensuring that summaries closely match the underlying documents and that biographies are factually correct. Fairness is also a priority, ensuring that outputs treat demographic groups similarly and address intellectual property and copyright considerations. Additionally, we ensure appropriate usage by filtering out user requests for legal advice, medical diagnoses or illegal activities.

We also work to mitigate toxicity by removing hate speech, profanity and insults while prioritising privacy by protecting personal information and customer prompts. We build solutions to address these issues in our processes for acquiring training data into the FMs themselves and into the technology we use to pre-process user prompts and post-process outputs. For all our FMs, we actively invest in ways to improve their features and learn from customers as they experiment with new use cases. Together with academia, industry and government partners, we are committed to the continued development of GenAI in a responsible way.

On a personal note, what was the first time you were impressed – or even shocked – by the powerful capabilities of Generative AI? What did you ask it to do, and why did it evoke such a reaction?

My wife and I asked a GenAI application for details about her pregnancy and we were quite impressed with the answers delivered. We still double-checked with our doctor, though! On a work level, I was in awe when I first started using Amazon CodeWhisperer – now called Amazon Q Developer – that is a GenAI-powered assistant for software development. Hands down, I was able to develop code orders of magnitude faster!

On a different issue, despite recent reforms, the digitisation of Cyprus’ public sector can be best described as a work in progress – it lags behind many EU members in e-government services and the use of digital tools by public authorities. What practical steps should be taken before it can leverage GenAI?

Cyprus has made significant progress in its digital transformation. Ministries have proactively developed measures to boost economic competitiveness for future generations. The Deputy Ministry of Research, Innovation and Digital Policy established the National Digital Strategy, positioning Cyprus as a hub for sustainable business and international trade. We have signed a Memorandum of Understanding with the Government to support the modernisation of infrastructure and accelerate cloud services adoption in the private and public sectors.

The national digital agenda sets out a bold vision of a digital society and we are honoured to help make that vision a reality. We have seen many nations experiencing a digital skills shortage. In fact, the World Economic Forum predicts that, by 2025, over half of the workforce will require significant reskilling or upskilling to do their jobs.

So, a practical step to leverage GenAI is to encourage an environment of lifelong learning, which will be key to ensuring that innovation does not outpace the workforce. This includes upskilling students and reskilling current workers.

Staying with Cyprus, the island has emerged as a European hub for trading platforms and payment institutions. What actionable steps should companies operating in these segments take to start their Generative AI journey? And what are the top use cases at the moment?

Cyprus is indeed a hub for such institutions and we collaborate with several local trading and payments organisations. I would recommend that they take three actionable steps to start their journey. First, identify the right GenAI use cases that will deliver business value. Second, prove that value by delivering Proof of Concepts and designing your GenAI operating model and third, use the right infrastructure to go to production. In terms of use cases, we see everything across the board. In the FSI space, the NatWest Group uses Amazon Bedrock to help it detect if customers are being scammed.

By analysing customers’ behaviour with AI, the bank can spot unusual payment patterns earlier and intervene faster, reducing financial loss. Another example is Bridgewater Associates, which uses Amazon Bedrock to power its Investment Analyst Assistant to generate elaborate charts, compute financial indicators and create summaries of the results to accelerate the tedious parts of the research process. This enables analysts to spend more time on the difficult and differentiated aspects of understanding markets and economies.

Finally, perceptions surrounding the technology have moved from "it's rather implausible" to "we are getting there faster than we think". It begs the question: what's next for the technology and how soon will it be here?

We see emerging trends in technology, probably coming very shortly. One of them is Major Linguistic Models, which will be trained using cultural data resulting in a more nuanced understanding of the human experience and complex social challenges. In the coming years, culture will play a crucial role in how technologies are designed, developed and consumed, and the results will be significantly more evident in GenAI technologies.

We are also seeing the rise of AI Assistants, which will not only write and understand code but also be tireless assistants and teachers, personalised to the individual, the team or the entire company. Finally, we expect that higher education will evolve to keep pace with the speed of technology. Many new programmes will be developed that are more tailored to the needs of the industry and will resemble the experience one gains after many years of friction in the labour market.

(This interview first appeared in the February edition of GOLD magazine. Click here to view it.)