Just as crafting wine is a complex process involving purposeful steps, successfully developing and implementing artificial intelligence demands proficiency at different levels. This article outlines a framework for understanding the key phases involved - from foundational hardware and data collec-tion through packaging trained models and optimizing their real-world use.

This article is written by Abdel el Amrani ([email protected]) and Chung Do ([email protected]). Abdel and Chung are both part of RSM Netherlands Business Consulting Services with a specific focus on Tech-nology Consulting. 

The Winemaking Process of AI Development

Creating an effective AI model requires meticulous steps like producing high-quality wine. First and foremost is sourcing: for wine, this means selecting the optimal grapes based on factors like weather, soil and intended style. Analogously, AI depends on vast amounts of real-world data tailored to the desired application. Whether text, images or other media, acquiring a comprehensive, representative dataset is the foundation of any model. Once raw materials are obtained, preprocessing begins. For winemaking, this entails crushing grapes and extracting juice in preparation for fermentation. AI paral-lels involve cleaning and structuring information - tasks like tokenization, normalization and augmenta-tion strengthen what the models learn from. Both endeavors then reach an important design phase: winemakers pick the ideal yeast while AI researchers select model architectures matched to the data and goals.

Long-term development comes next, changing the character of initial inputs. Wine undergoes fermen-tation, transforming sugar into alcohol content. AI models require extensive training, refining parame-ters against examples to learn patterns and generate targeted outputs. Regular evaluation and testing guide improvements, paralleling how winemakers check variables and make amendments.

Finally, the results require packaging. Finished wine gets safely stored in bottles. Well-tuned AI needs responsible deployment through applications and APIs to prove its utility. Quality assurance remains critical too - clean facilities prevent wine spoilage, while data validation ensures model robustness. With equivalent diligence in each linked process, both pursuits can yield rewards greater than their sum.

Towards AI Excellence: Understanding the Four Levels of Mastery

There are four levels of mastery required to successfully produce and utilize AI models. The four main stages of AI development can be described through an analogy to winemaking - from the equipment suppliers needed to "ferment" models, to the expert "winemakers" conducting advanced training, all the way to productization and real-world application. Each level builds upon the previous and pro-gressing through all four stages is necessary to achieve excellence with AI.

Level 1 – Equipment sellers

Just as winemaking requires equipment to handle the fermentation and aging processes, developing sophisticated AI also relies on powerful processing capabilities to handle the immense volumes of data involved. Companies like NVIDIA have invested heavily in developing specialized GPUs, or Graphics Processing Units, and chips optimized for neural network training and inference. Their goal is to empower more organizations and individuals to embark on their own AI efforts by selling the hardware required to get started. Other firms contribute by developing distributed training frameworks and cloud services that allow machine learning workloads to seamlessly scale across massive com-puter clusters. Without the right tools and technology, the ability to train advanced AI models would simply not be possible given current computational constraints.

Level 2 – Wine makers

As with growing grapes and blending wine, developing high-quality, useful AI models takes significant time, resources and expertise to iteratively refine. While startups and smaller enterprises have demon-strated the ability to create basic models targeted at narrow problems, the largest tech players like Google, Amazon, Microsoft, and OpenAI currently have unmatched access to both the vast amounts of data and compute infrastructure required to train state-of-the-art massive language models involv-ing billions of parameters. Their teams of highly skilled researchers and engineers spend months or years iteratively improving models through successively larger pre-training runs to achieve superhu-man-level performance across a wide range of capabilities like conversation, translation, summariza-tion and more.

Level 3 – Packaging and promotion

The next level focuses on how to package trained models and promoting their usefulness. However sophisticated the underlying technology, AI systems offer little real-world value if they are not produc-tized and distributed in ways that are easily applicable. Leading companies excel at operationalizing their advanced techniques by building polished applications, tools and programming interfaces cen-tered around core trained models. For example, OpenAI makes their powerful language models ac-cessible through products like ChatGPT while also providing tooling to integrate capabilities like ChatGPT API and GPT Builder into other services and software. They also publish documentation to teach developers, researchers and vendors how to build upon, extend and harness AI components developed in-house. This serves to both expand adoption of their technologies while enabling a thriv-ing ecosystem of third-party AI applications and services.

Level 4 – Wine servers

The final level sees mastery in integrating AI techniques into real operational workflows to drive tangi-ble benefits. While the first three levels focused on developing foundational technologies, training capabilities, and distributing mature products - the true purpose of AI is how it can be holistically adopted by businesses and other organizations. Leaders in this final level apply AI as a means of streamlining processes, elevating customer and employee experiences, and gaining powerful insights from data. For instance, leveraging generative models for automating routine customer service inquir-ies or applying computer vision to improve factory workflows. The goal is to derive real, measurable value from AI investments across entire enterprises.

Which Level of Mastery Should Company Focus On?

Just as hobbyist winemakers can also create drinkable wine, companies can experiment with basic AI models for internal use. However, to produce quality products at scale requires extensive control over the full development cycle. Different businesses have varying goals, so their optimal AI strategies should also differ. Leaders need not aim for mastery across all levels - specializing in particular phas-es allows maximizing strengths.

Level 1 involves developing critical hardware like computer chips and GPUs. This allows models to be trained efficiently using massive datasets. However, cutting-edge hardware research requires huge investments that many companies cannot afford. Focusing efforts elsewhere may be more realistic.

Level 2 mastery is training sophisticated internal AI models. As these models are continually improved through ongoing use, feedback, and adjustments, it creates a competitive edge over time. Model performance incrementally increases, capturing more value. Yet few have the vast resources needed to rival technology giants in large-scale training capabilities. Level 2 suits ambitious startups able to dedicate extensive engineering resources to refining powerful algorithms.

Level 3 packaging centered around taking trained models and integrating them into user-friendly ap-plications and APIs. This opens revenue streams from selling software access, while outsourcing much of the data preparation and training workload. Success means competing with many other ven-dors also leveraging external models. Well-established product companies tend to find this a practi-cal option that matches their core competencies.

Finally, Level 4 sees AI integrated directly into regular operations and workflows to streamline pro-cesses. While saving costs by avoiding proprietary model training, complete reliance on external technologies introduces risks. Data privacy and security becomes a bigger concern, as does overde-pendence on third parties for ongoing support and improvements. Most traditional businesses are better suited to this optimized application-focused approach.

In each case, companies must match their strengths and goals to the level demanding resources and expertise available internally. Focusing expertise generally leads to a competitive advantage over attempting to master every facet of AI. Business needs to assess their competencies and determine where focusing mastery - at model development, product creation, or applied usage – to yield maxi-mum benefits. Specializing skillfully at one level is frequently more viable than attempting breadth across all.

Forward thinking

Just as the wine industry offers countless varietals and brands, AI presents massive diversity in how it can be applied. With so many options at each step, properly selecting and using generative AI tech-nologies is complex. In the end, AI adoption comes down to crafting the right "cocktail" for your indi-vidual needs. As people have personalized preferences around flavors, visual appeal, and experienc-es, so too will companies differ in what they need from AI to address unique challenges and opportu-nities. It is crucial to carefully allocate resources, time, and focus on the level of AI mastery best suit-ed to long-term objectives and strengths. Clarity on priorities, from model development to workflow integration, ensures investments maximize benefits.

Most importantly, AI should not be seen as a threat replacing jobs or businesses. Rather, those who skillfully leverage AI - through the framework of the four mastery levels - will be best positioned to thrive innovation in their industries. Ongoing learning and adaptation maintain an advantage as new generations of generative technologies continually emerge. AI holds great potential to augment, not disrupt. By finding the right "fit" at each stage of development and use, stronger possibilities are unlocked. 

RSM is Thought Leader in the field of Technology Consulting. We offer frequent insights through training and sharing of thought leadership that is based on a detailed knowledge of regulatory obliga-tions and practical applications in working with our customers. If you want to know more, please reach out to one of our consultants.