Walid Mehanna is Chief Data & AI Officer at the Merck Group in Germany, where he also chairs the company’s Digital Ethics advisory panel. At Merck, he leads and helps deliver data & AI strategy, architecture, governance and engineering across the organisation. Born in Egypt and raised in three different states in Germany, Walid has harnessed his multicultural background to inform his commitment to diversity, equity, and inclusion in the workplace.
In today’s fast-evolving digital landscape, companies are seeking new ways to leverage AI tech that can streamline their operations. German-based organisation Merck KGaA leads the charge with the introduction of myGPT@Merck, a generative AI tool proving to be a game-changer for its employees.
In this post, the Walid talks to us about the development of this innovative new tool, and discuss other tech projects in the pipeline that are set to revolutionise the company:
Leveraging the immense potential of Data and AI is integral to both our present and future business success. To this end, we have crafted a comprehensive, organisation-wide data strategy and an integrated Data and Analytics Ecosystem, at the heart of which are Palantir Foundry and Amazon Web Services. We defined common ways of working for everything related to Data and Analytics and ignited a lot of excitement for the topic through various data culture activities. Our ambition is to pioneer digital ethics, and in pursuit of this, we have implemented a Code of Digital Ethics and assembled a Digital Ethics Advisory Panel.
As we continually refine our data strategy, we are now expanding it to encompass the AI domain. We also took our organisational setup to the next level and recently announced the formation of the Merck Data & AI Organisation which I have the pleasure to lead. We also have an AI Research team in our Science & Technology Office that is looking into new trends and emerging technologies.
We organise ourselves in a federated operating model that we call hub-hub-spoke. Everything that makes sense to only do once on the enterprise-level is done in the corporate center I head up. Then we also have hubs in the different business sectors and one for the group functions and, of course, many people that operate directly in the businesses and functions. We call them “spokes.” And we have a regular decision committee that we call the Merck Data Council. In this committee my peers in the sectors as well as representatives from IT, Security, Data Privacy, and I jointly shape and execute our data strategy.
myGPT @ Merck is an artificial intelligence-based digital assistant designed to support Merck employees in their daily work. It is powered by the GPT (Generative Pretrained Transformer) language models developed by OpenAI, and it can understand natural language queries and provide information and helpful responses. The current version of myGPT @ Merck uses the model gpt3.5-turbo.
We are a vibrant science & technology company with many curious minds. When ChatGPT came up, of course, those curious colleagues started to play around with it. We wanted to encourage them to embrace the new possibilities, but at the same time, we were concerned about our data security and the potential leakage of internal or even confidential information through a tool like this. This is why we decided to make this technology available in a Merck-compliant way and created our own GPT chatbot in collaboration with Microsoft.
The domain of Data & AI is full of exhilarating developments, making it difficult to pick just one. However, as someone deeply immersed in data, I find the growing recognition of the importance of high-quality data particularly gratifying. The increased interest in AI has made it clear that effective data management and governance are critical. After all, the quality of our insights directly mirrors the quality of our data – if we input garbage, we’ll output garbage. One promising aspect of AI is its potential to streamline and enhance our data supply through automation.
Equally exciting is the increasing integration of Large Language Models (LLMs) into various services via APIs. This promises significant advancements, although it requires rigorous testing and governance to ensure seamless and reliable operation. The potential benefits to streamline and automate operations, however, are tremendous.
Looking further into the future, the digital representation and simulation of drugs, materials, and, eventually, the human body, maybe represents the final frontier for us. This possibility, which may seem like science fiction today, could become a reality with the continued development of massive compute, data and AI. To me, it represents the ultimate challenge and aspiration for us as Data & AI professionals in science and technology.
Mathias Winkel leads the AI & Quantum Lab at Group Digital Innovation within the Merck Science and Technology organisation. With his team of experts in many scientific disciplines, he is continuously scouting for new and exciting technologies in AI and novel methods of computing. Matching them to real business problems, the team is propelling innovation powered by data and digital for the company, its customers, and patients.
Merck has more than 350 years of pharmaceutical and chemical tradition and corresponding expertise. The company’s ongoing success as a leading science and technology company has only been possible through permanent innovation. Today, this innovation imperatively needs to be powered by data and digital.
The AI & Quantum Lab within Group Digital Innovation at Merck is one of the essential building blocks to bring this aspiration to life. It is a team of highly specialised experts on the permanent lookout for new and exciting digital technologies that are documented in scientific publications, are developed by startups, or are industrialised by established companies. We assess methods in fields ranging from artificial intelligence and machine learning to quantum computing, fit them to existing business problems at Merck and transfer them into the company.
When the global hype about ChatGPT started, the technology behind it was not new: the potential of transformer architectures was already demonstrated by Google researchers at NeurIPS 2017 (Conference on Neural Information Processing Systems) and ChatGPT was preceded by several influential LLMs (large language models) developed by OpenAI as well as other wellkown companies. However, the specific way of training the model through RLHF (reinforcement learning with human feedback) with the specific goal to create a conversational AI that answers user prompts in a way, that cannot be distinguished from a human, has been a unique idea and finally lead to ChatGPT being the most influential development AI in 2022.
This tremendous impact resulted from three aspects. First, the training method and underlying data as well as the extreme size of the model with its 175B parameters made it indeed extremely powerful. For a long time, competing developments were unable to win against GPT-3.5, the underlying model of ChatGPT, in many relevant benchmarks.
Second, the very simple user interface made ChatGPT extremely approachable for everybody. Suddenly, artificial intelligence neither was some specialised black-box algorithm shipped as a hidden part of a huge software suite, nor some magic program that required special knowledge and a level of black art for getting it to do something useful. Instead, people who barely used a computer before, were able to interact with powerful artificial intelligence. Finally, these two properties, paired with initially very restricted access, lead to self-propelling excitement all over the world and an exponential growth in the user base.
“ChatGPT discussions will help us as humans to better understand ourselves and maybe even spark insights into what human intelligence is and how the brain is working.”
As mentioned before, Merck is positioned globally as a leading science and technology company. Equipping our employees with the most powerful and most efficient tools available for serving our customers and patients is a crucial cornerstone to fulfill this ambition. To also protect our customers and patients and their data as well as our company internals, it is however critical to ensure that usage of digital tools at the workplace happens in full compliance with our company’s regulations and within the applicable legal limits. Considering and balancing both aspects lead to the decision to set up myGPT, our own Merck-specific version of ChatGPT.
At the AI & Quantum Lab, we were already actively following the technology and analysed its capabilities and limits at least half a year before ChatGPT was announced. Already since then we were in very good contact with the developers and could build up internal expertise on large language models. This allowed us to understand their potential and limits early on and embark swiftly when we saw a perfect fit of the technology to our business needs. Today, we are actively developing internal applications using LLMs that go far beyond the well-known chat bot capabilities. The global excitement around ChatGPT has been a surprise, but due to our technological foresight we were well prepared.
In today’s rapidly changing world, this foresight is imperative not only to make best use of emerging technologies. It is also necessary to ensure business continuity. Just like digital cameras eradicated the market for classical photography only to be soon replaced by smart phones, every business model is under permanent risk of severe disruption due to technological advancements. This development has become even more serious during the last few years as traditional business limits are diluting more and more: developers of consumer devices look into autonomous driving, developers of electric cars build rockets, social media companies consider developing drugs, etc.
While innovation and technology scouting has to happen permanently at all levels and in all sectors at a company like Merck, it is the task of the enterprise level functions to take the perspective of the company as a whole to ask: which technologies are most promising for all sectors and which ones would be most disruptive for the company in general.
Answering this question is tough, because we are living in extremely exciting times where development cycles have become much quicker than before and nothing seems older than yesterday’s news about novel technologies. When concentrating on LLMs, I see three primary trends that we should expect to have some impact very soon. First, I am sure that these models will dramatically change the way we interact. We will be able to describe the desired output in natural language, and AI systems will spare us from most of the mechanic mouse and keyboard interaction to produce our digital creation.
Second, multi modality of models will again open a multitude of new application areas beyond natural language. Think about explanation and generation of images, audio, videos. Imagine the possibilities if these models enter the domain of science, start generating new materials or explaining fundamental laws in a way a human might never have thought of.
Finally, the human-like performance of these models on many tasks and their surprising capabilities as autonomous agents or when it comes to logical reasoning already lead to passionate discussions about General Artificial Intelligence, i.e. AI that can autonomously plan and solve problems it has not seen before.
Independent of their outcome and potential further developments in this direction, these discussions will help us as humans to better understand ourselves and maybe even spark insights into what human intelligence is and how the brain is working. It will also raise new ethical questions about the border between human and machine, about creativity that we consider a unique feature of intelligence, and about how we as humans interact with each other through these technologies. Thus, ChatGPT and its descendants and siblings will challenge humans on multiple levels.
Harsha Gurulingappa operates as Head of Text Analytics at Merck Data and AI Organisation and located at Merck IT Center, and was instrumental in establishing NLP as one of the core capabilities and service within Merck. In his current role, he is responsible for technological enhancements and adoption of NLP across business sectors within our organisation.
The ability to find precise information, facts or figures in time is one of the fundamental requirement for any business or individual. At Merck, we are an organisation with diversified businesses, functions and processes. Having capabilities to process complex data which is multi modal, multilingual and significantly unstructured, and transform them into actionable information and insights is a necessity.
The NLP team holds expertise in development and industralisation of data products and solutions leveraging practices of AI/ML/NLP works towards establishing NLP technical capabilities as part of Merck’s central Data and Analytics ecosystem. By consulting, partnering as well as upskilling citizen data scientists and data practitioners operating within or in proximity to business functions, we partner in rolling out sustainable NLP solutions which eventually gets integrated into business processes for generating distinguished value.
The success of ChatGPT and its strong adoption within the market raised an appetite within Merck to have comparable technology which is safe, secure, compliant and easily accessible within Merck.
A team of experts from different functions under the leadership of Chief Data & AI Officer, proactively assembled to drive this journey within a short timespan of less than two months. The team was a composition of colleagues from diverse expertise such as (but not limited to) cloud architecture, NLP engineering, AI/ML scientist, cybersecurity, data privacy, legal, communications as well as experts from Microsoft OpenAI service. myGPT @ Merck was designed, developed, tested, industrialised and released to a cohort group of over ten thousand users. It leverages secure Microsoft OpenAI service in the background and offers a custom designed front-end for the end users with features and functionalities similar to OpenAI’s ChatGPT.
On the other hand, partners from legal, data privacy and communications ensured development of internal policies and procedures, as well as training programs to ensure myGPT @ Merck is used in a compliant way. Altogether, myGPT @ Merck has been in production over a month and there is a surging adoption by the community of users for various purposes to boost efficiency and productivity in their daily business.
myGPT @ Merck is an iconic solution and symbolisation of Consumable AI delivered to be available for every Merck employee. However, as part of the NLP environment within central Data and Analytics ecosystem, there is availability of various generative models offered by industry leading technology providers such as HuggingFace and John Snow Labs.
These models can be finetuned on specific data and tasks. Our NLP environment allows deploying language models as APIs which can be consumed through data pipelines or custom applications requiring real-time response for end-users.
Large Language Models (LLMs) constitute a family of models which can cover different data modalities, domain specificity (e.g, general purpose or domain specific) and task specificity (e.g., Multi-tasking, text generation only, and more). They can range from having developed with few million parameters (e.g. BERT) to many billion parameters (e.g. GPT-3).
The NLP environment within our Data & Analytics ecosystem allows training, fine-tuning, and transfer learning language models. Such training or fine-tuning tasks have successfully been executed and deployed for various business cases. Most of the applications of LLMs fall in the category of information extraction from semior unstructured documents and retrieval augmented generation (RAG, aka. Retrieval augmented questionanswering).
From the LLM perspective, we are investigating leading proprietary API providers which can offer multilingual and multi-modal operations with the ability to fine-tune and deploy business specific data at viable costs. We are curiously observing the market for applications of GenAI and LLMs across various stages of pharmaceutical product development such as drug discovery, drug development, patient recruitment, regulatory process automation as well as post-market surveillance.
One other exciting development we are monitoring is integration of GenAI and LLMs into workplace technologies (e.g. email, Webex, Office tools, and more).
As the Global Head of Data Culture, Stefanie Babka is responsible for communication, change management and upskilling programs on Data & AI including the Merck Data & Digital Academy and the enterprise-wide Data & Digital community. The data culture activities are aiming to bring together artificial intelligence and human intelligence to drive meaningful insights and data-driven decision-making.
Data Culture is a skillset and mindset. You might have heard that saying that “culture eats strategy for breakfast”. According to studies culture is the biggest impediment for organisations to become data-driven. The digital transformation is not only about technology and processes. Indeed it is mostly about the people.
You need to take them with you otherwise the best technology will not bring any value. The beauty of myGPT @ Merck is that the entry barrier to the technology is very low which is why it is easy for people to embrace the tool.
I have to say it is a game changer. All of our activities in the data culture team aim to make people aware of the possible with Data & Analytics and AI. With the rise of generative AI tools like ChatGPT, all of a sudden people are able to understand the potential and the value that this can bring to them personally but also for our business. With my GPT @ Merck employees have a tangible application at hand where they can use conversational AI in a safe environment. They can experience it themselves. That is worth more than a hundred hours of theoretical training. And of course they get curious and want to learn more and hopefully they will develop the data creativity to embark on their own Data & Digital exploration journey looking into new ways of working with AI and creating new business models or products.
The most important rule is that the human person is always going to be in the driver’s seat and myGPT @ Merck is only an assistant helping to do the work in a more efficient way. It is required to validate the outcome and in some use cases also necessary to check back with our legal teams. The tool is available for confidential and internal data with the exception of personal information.
We currently focus on teaching our guardrails for the tool as well as making people understand how it works. This means we do trainings on demystification of AI that explain how machine learning, deep learning and large language models work. We also offer hands on training for prompt engineering and getting the best value out of this tool.
I think it is very interesting that this topic has now reached a broader audience and it has been discussed also in popular media for example and all of a sudden even my family comes and asks me about my opinion on, for example, ethical questions with regards to AI. The AI revolution will increase the need for general data literacy.
