Digital business excellence

Generative AI

Technology Summary

Generative Artificial Intelligence (AI) is a subset of machine learning that focuses on algorithms that generate new content based on existing content, such as text, images, or music . The resulting artifacts are authentic-looking, however original and newly produced.

Generative AI can be usefully structured across three dimensions:

  1. The output that is generated: Text, images, video, 3-D models, music, code, and ‘other’. ‘Other’ is a growing list of domain-specific results. Generative AI is valuable wherever input datasets of huge complexity and variety can be consumed as a basis for output generation. In biology, it is for example used to generate DNA and protein structures.
  2. The input that is consumed during model usage (inference). This is dominantly text, but also e.g. images.
  3. The specificity of the AI solution for a given domain or task. For example, ChatGPT can be considered a general-purpose text generation, while Donotpay is an AI lawyer generating legal advice. This dimension is not binary (generic or specific), but a continuum.

The table below is a deliberately simple summary of important Generative AI use cases. It includes a few non-exhaustive examples that are leaning (mostly) to the generic side .

Output Category Use Case Examples
Text Generate short texts, ‘conversational’: Generate short-text interactions like social media chats (Twitter, Instagram…) or individual chats (WhatsApp…).
Generate medium-sized texts, ‘answering’: Generate summary responses for any question, in a couple of short paragraphs. ChatGPT
Generate long texts, ‘text generation’: Generate essays and other more substantial pieces of text, often interactively with a human writer. Jenni AI
Jasper AI
Domain-specific texts: Similar to the above, but specialized, high-quality advice for a specific topic. DoNotPay
Summarizing: Generate input text summaries. This may be preceded by a speech-to-text converter accepting e.g. an audio recording of a meeting. Fireflies
Images Generate art and photorealistic images. Dall-E-2
Stable Diffusion
Video Generate videos, often interactive with a human video creator. This is in an early stage, but progressing fast. Gliacloud
Music Generate music. Ampermusic
3D models Generate 3D models. point-e
Code Generate code. This includes extremely advanced autocomplete functions supporting human coders, based on natural language. Tabnine
OpenAICodex CodeT5
Other Example: AI-based prediction of protein 3D structures. Alphafold

Generative AI produces genuinely new content. For example, Stable Diffusion generates a genuinely new picture from a random ‘white noise’ starting point. However, the term ‘Generative AI’ also encompasses data transformation across domains and styles. The result is also ‘generated’ but hardly as ‘original’ as an image generated from white noise. The most prominent examples are:

  • Speech (audio recording) to text (e.g., Nuance Dragon)
  • Text to speech (e.g., Murf), text to a speaking video avatar (e.g.,, Synthesia)
  • Language translation (e.g., DeepL)
  • Upscaling of images (Example: Gigapixel AI). This is important because image generators often generate pictures of a medium resolution, e.g. 512 x 512 pixels. These can then be upscaled with upscalers.

Image of a medusa iteratively generated from a white noise picture (Stable Diffusion)

Image of a medusa iteratively generated from a white noise picture (Stable Diffusion)

Technology Evaluation

What caused the current breakthrough progress of Generative AI?

Artificial Intelligence had a slow start. Fundamental ideas were developed in the 1960s, but many predictions on its maturing speed were not met. Instead, for about half a century there was an up and down of expectations followed by disillusionment. The ‘downs’ were also called ‘AI winters’.

However, since roughly the 2010s there is a period of unprecedented progress. A key contributor is certainly the scientific advances that were made. But arguably of at least the same importance was the actual implementation of very large-scale model training. First of all, this required financial investment. For example, OpenAI (creators of ChatGPT) received funding in excess of USD 1 billion. Second, it required large amounts of training data. Data collection organizations, often non-profit ones like CommonCrawl or Laion, crawl billions of web pages and collect petabytes of data to train the models with. Finally, massive advances in hardware (especially GPUs) provided the processing power required for all that data. To a large degree, it is the dizzying growth of model sizes which led to the breakthroughs we see today.

Market – Current Adoption

Impact on the IT Industry

Today, Generative AI quickly becomes mainstream. ChatGPT’s publishing in November 2022 was certainly a breakthrough in terms of public attention to Generative AI. In a nutshell, the 2010s were the decade of massive technological advances in AI, and in all likelihood, the 2020s will be the decade of far-reaching impact on society.

While general purpose models and solutions are quickly spreading, the market of domain-specific solutions that can be built on top of them is nearly untapped. Sequoia Capital calls this the upcoming ‘wave of killer apps’ and compares today with the early times of mobile phone app stores: Then as now, creative people can come up with great new ideas about how to use the new possibilities, filling niche after niche. So, on the professional side, we are on the brink of a new wave of software creativity.

Impact on society

Generative AI is civilization-altering. It is comparable to the invention of the wheel, book printing, or power-generating machines.

A good analogy can be made to power-generating machines. Since the beginning of mankind, humans had almost exclusively muscles (human or animal muscle) to perform physical labor. The invention of the steam engine that rotated a shaft changed that and started the industrial revolution. Different kinds of power-generating engines were invented later. By now, the bulk of physical labor is done by engines.

Today, a significant percentage of workers are information workers. It is not clear yet to which level this can be replaced by AI, for example, if AI can write scientific articles beyond what a human could, or write a movie script more entertaining than a human-written script. But even if these heights should not be reached any time soon: It is already clear that the bulk of human information work can drastically be reduced by Generative AI. Its results are good enough for the 90+% of tasks that are not at the tip of human creativity.

Risks of Generative AI

Short- to medium-term impact on the labor market

Similar to the industrial revolution, the social impact will be profound. In the transition phase, many people’s lives will be negatively impacted before the benefits reach them. Unemployment, social unrest, and worse could result.

Difficulty in assessing reality

‘Pics or it didn’t happen’ (#POIDH) is a well-known meme. We sometimes demand pictures as proof of reality, but we also accept them to be the same. This will change. To the degree that AI-generated pictures becomes undistinguishable from a photo, we can no longer believe our eyes and ears in the way we do today. On the other hand, algorithms for identifying generated content are being developed, a battle for reality with currently unclear outcome.

Inability to verify human intellectual achievements in writing

For schools and universities, it becomes impossible to verify the original author of a written result, like a homework essay, seminar paper, or even a master’s thesis . Already today, copy-and-pasting is common, but plagiarism checkers do a reasonably good job of discovering that. This may become nearly impossible in the future, as AI-generated text is essentially original.

Image generated with Stable Diffusion. The true revolution lies in the fact that this was generated by the author in five minutes with a few tries, without any design or drawing skills

Image generated with Stable Diffusion. The true revolution lies in the fact that this was generated by the author in five minutes with a few tries, without any design or drawing skills

Proliferation of weapons

In six hours, a simple AI model trained on publicly available data generated 40,000 potentially deadly molecules, based on the nerve agent VX. While having such a result is not the same as producing the actual physical end product, this serves as a reminder of the power that could fall into the hands of malicious actors.

Opportunities of Generative AI

Provision of a much-needed productivity boost

There is no risk that humanity will run out of work. On the contrary, the world is facing an aging population and thus major problems due to a dramatically shrinking workforce. To avoid prosperity shrinking with it, human work must become much more productive. Generative AI may be a key enabler for that.

Contribution to solving major problems

Generative AI can lead to breakthrough scientific results, e.g. on topics like cancer research and many others. It might even contribute to solving humanity’s most pressing issues, in particular, environmental changes and finding a clean and sustainable energy source.