Unlocking the Potential of Creative Machines

Generative AI is revolutionizing industries and unlocking unprecedented possibilities. In this article, written with input from two sessions by AI experts, Raza Habib and Moritz Schröder**, you can delve into the world of generative AI, explore its significance, potential applications, and the considerations startups should keep in mind as they apply this transformative technology.

How startups that leverage generative AI methodologies fuel innovative growth

Generative AI has garnered significant attention recently, fueled by its immense potential and the hype surrounding artificial intelligence and machine learning. However, there is a significant gap between people’s perceptions of AI and its true capabilities.

Many companies leverage AI without fully understanding how to harness its capabilities effectively. According to a survey among 2,500 companies conducted by one of our speakers, only 34% of companies were able to generate economic value from implementing AI. This is in stark contrast to 92% of companies in that same survey that consider AI to be an important business opportunity. Hence it is critical to measure the impact of AI projects and implementations, i.e. the Return of AI (RoAI). With that, the focus shifts from using AI for the sake of itself to utilizing generative AI methodologies to improve results and products.

To do this, a methodology for identifying which AI projects actually drive value needs to be created. This was accomplished with the creation of AI Canvas by Moritz’s team at Merantix together with the Universität St. Gallen (see below).

As part of a Whitepaper for critical and holistic evaluation of AI ideas, it brings all important aspects of using Generative AI together and visualizes the need for different units in companies to align here. It combines the most important questions you should ask yourself before implementing GenAI:

1. What value will the AI generate?
2. How will we measure the value?
3. How will it integrate into the company?
4. Which data will we use?
5. Is there a legal or ethical risk?
6. How will the solution run long-term?

One noteworthy example of the transformative impact of generative AI is the beta version of Firefly, an add-on for Adobe’s renowned Photoshop software. Firefly has undergone a remarkable transformation through the incorporation of generative AI technology, enhancing various aspects such as generated content and language interpretation. This integration enables written commands to be translated into images, ushering in a new level of creative possibilities. Firefly is just one illustration of how generative AI is reshaping familiar tools and pushing the boundaries of what is possible.

Additionally, generative AI has made its mark in widely used applications like the background filter on Zoom and Amazon’s recommendation engine. While seemingly simple, these AI-powered features have a substantial impact. The recommendation engine, in particular, has become a highly influential generative AI system, generating significant cash flow for Amazon.

🚧 Unveiling the Structure: A Deep Dive into an Impressive Generative AI Application

Most generative AI applications share a common structure consisting of three key components.

• The first component is the base AI model, which provides the raw intelligence and can be a large pre-trained model or a custom model trained on specific data.
• The second component is the prompt which serves as the instruction to the model for generating output.
• The third and final component is the data selection strategy which determines what data is provided to the model for additional context.

🔍 Components of Success: Lessons from a State-of-the-Art Generative AI Model

Creating a state-of-the-art generative AI application requires careful consideration of key components. Let’s explore the work of accomplished researchers who made significant strides in this field. These researchers developed an advanced generative AI model for image synthesis.

• They utilized the StyleGAN2 architecture, known for high-quality image generation. This model captured fine details and patterns, producing visually striking and realistic images.
• To empower users, the researchers designed a user-friendly interface allowing input customization, such as object categories, color palettes, and artistic styles. This input component offered control while embracing creativity.
• For accurate replication of desired visuals, the researchers curated a diverse dataset and employed advanced conditioning techniques like latent space interpolation and style mixing. This dataset served as a foundation for the model to learn and generate outputs that faithfully represented desired aesthetics.

The success of this generative AI model highlights the significance of component selection. The researchers’ expertise, coupled with the base AI model, user-centric input design, and curated dataset, resulted in visually stunning and realistic image generation.

Examining this work provides insights into the components behind successful generative AI models. It reinforces the potential of generative AI to revolutionize industries and unlock creative possibilities. This demonstrates the power of technical expertise, thoughtful component selection, and a deep understanding of generative AI principles.

🧠 Pitfalls and Best Practices

Based on the experience of working closely with tech companies building generative AI applications, several best practices have been developed for building reliable and differentiated AI applications. Let’s explore:

1. Objective Evaluation: Systematic evaluation is crucial for developing reliable generative AI applications and continuously improving them. Evaluating generative AI is more difficult than evaluating traditional ML applications as the output of generative AI models is often subjective in nature (e.g. generating a sales email). Deciding what feedback you capture to evaluate your AI application is an essential step. Also, it is imperative to incorporate evaluation at every stage of the development process from prototyping to production. Without an appropriate evaluation framework in place, it’s difficult to understand what is working and identify opportunities for product improvements.
2. Effective prompt engineering and management: The importance of prompt engineering, i.e. creating instructions for a generative AI model, is often underestimated. Continuously improving prompts can significantly improve model performance. Subject matter experts play a crucial role in that process, even if they are not technical. Additionally, by providing relevant contextual data to a generative AI model (e.g. through retrieval augmentation), the accuracy of your generative AI application can be improved further.
3. Avoiding Premature Optimization: Optimization techniques such as fine-tuning are often employed prematurely in the development process. However, premature optimization can make it more difficult to evaluate and adapt generative AI applications and can increase their maintenance burden before the application has even been validated. Therefore it is recommended to start with a powerful base model such as GPT-4 and to push the limits of prompt engineering (whilst continuously evaluating the generative AI application) before even considering more advanced optimization techniques. Only then should additional optimization techniques such as fine-tuning be considered. Complex technical setups such as LLM agents and LLM chains should be used only if they are essential for your use case.
4. Leveraging Fine-Tuning: Fine-tuning a generative AI model can yield significant performance improvements. Many developers shy away from investing effort in fine-tuning due to its perceived complexity or lack of data. However, as few as 100 data points are required to be able to fine-tune a model. By collecting and curating a dataset and fine-tuning the model on it, the model can benefit from incorporating proprietary data and be customized to a specific task whilst improving model accuracy, cost, and latency.

Conclusion

Generative AI, with its potential to unleash the creative power of machines, will reshape various industries. By understanding the significance of generative AI and how to best apply it, startups can leverage this transformative technology to drive growth and innovation.

While the hype around AI may have created misconceptions, generative AI has already demonstrated its enormous potential. The example of Firefly and the beta version of Adobe’s Photoshop add-on illustrate how AI can elevate familiar tools to new heights. Additionally, widely used applications like Zoom’s background filter and Amazon’s recommendation engine exemplify the impact generative AI can have on user experiences and business outcomes.

By following best practices such as systematic evaluation of AI applications and effective prompt engineering, startups can navigate the complexities of building generative AI products. Understanding — and therefore demystifying — generative AI and its uses will empower businesses to unlock the true potential of creative machines and achieve remarkable outcomes.

This article was co-written and edited by Stephanie Stein and Ferdinand Dansard with input from Moritz Schröder, ML Strategy Lead at Merantix Momentum, and Raza Habib, Ph.D., Co-Founder & CEO of Humanloop.