Generative AI is a branch of artificial intelligence that creates new content, such as text, images, music, code or synthetic data, instead of only classifying or predicting from existing data. It works by training large neural networks on huge datasets until they learn the patterns well enough to produce new, realistic outputs on request. Its potential is broad: faster creative and engineering work, new products, and a large share of future demand for computing hardware. It also brings real risks around accuracy, ownership and misuse.
This post was first written in September 2023, less than a year after ChatGPT was released to the public in November 2022. The potential and the challenges below are as the author saw them then, with some background added on how the technology works and why it matters to the semiconductor industry.
Key takeaways
- Generative AI learns the statistical structure of its training data and samples new content from it.
- The main model families are generative adversarial networks (GANs), variational autoencoders, transformers (behind large language models) and diffusion models (behind most image generators).
- Applications span creative work, education, healthcare, business and entertainment, and increasingly engineering, including chip design.
- Challenges include ethics and ownership, accuracy and control, compute cost, and cultural impact.
- Training and running these models needs enormous computing power, which is why it drives demand for GPUs, accelerators and high-bandwidth memory.
What is generative AI?
Generative AI focuses on creating new content or data from scratch, such as images, text, music and code. It uses techniques such as deep learning, natural language processing, computer vision and generative adversarial networks (GANs) to learn from existing data and generate novel and realistic outputs.
How generative AI works
All generative models learn, in some form, the probability distribution of their training data, then produce new samples from it. The main families differ in how they do that:
| Model family | Basic idea | Known for |
|---|---|---|
| GANs (introduced by Ian Goodfellow and colleagues in 2014) | A generator makes fake samples and a discriminator tries to tell them from real ones; each improves against the other | Realistic faces, image style transfer |
| Variational autoencoders | Compress data into a smooth latent space, then decode new points from it | Image and molecule generation, data compression |
| Transformers (from the 2017 paper Attention Is All You Need, Vaswani et al.) | Predict the next token in a sequence using attention over all previous tokens | Large language models such as GPT, code assistants |
| Diffusion models | Learn to remove noise step by step, so they can turn pure noise into an image | Text-to-image tools such as Stable Diffusion and DALL-E |
A simple way to picture a language model: given the words so far, it assigns a probability to every possible next word piece, picks one, adds it to the text and repeats. Do this with a model of billions of parameters trained on a large share of the public web, and the output reads as fluent, often useful prose. It also explains a key weakness: the model produces what is likely, not what is verified, so it can state errors with confidence.
The potential of generative AI
Generative AI has a lot of potential applications in various domains, such as:
Creative industries
Generative AI can help artists, designers, writers, musicians and other creative professionals to produce original and diverse content, such as paintings, logos, poems, songs and stories. Generative AI can also augment human creativity by providing inspiration, feedback or collaboration.
Education and research
Generative AI can help students and researchers to learn new concepts, generate summaries, write essays and create presentations. Generative AI can also help teachers and educators to design personalised and engaging learning materials and assessments.
Healthcare and medicine
Generative AI can help doctors and patients to diagnose diseases, generate medical reports, design candidate drugs and create realistic simulations. Generative AI can also help researchers and scientists to discover new insights and innovations in biotechnology and genomics.
Business and marketing
Generative AI can help entrepreneurs and marketers to create new products, services, brands, logos and slogans. Generative AI can also help businesses and organisations to optimise their operations, strategies and decision making.
Entertainment and gaming
Generative AI can help gamers and developers to create immersive and interactive experiences, such as virtual worlds, characters and stories. Generative AI can also help entertainers and media producers to generate realistic and captivating content, such as movies, shows and podcasts.
Engineering and chip design
The same techniques are being applied to engineering work. In semiconductors, examples discussed by EDA vendors and chip companies include assistants that answer questions about design documents, draft RTL or testbench code, explain tool logs and help with verification planning. Machine learning was already used before generative AI for tasks such as placement optimisation. Any generated design content still needs the same review and verification as hand-written work.
How big is the opportunity?
McKinsey's June 2023 report, The economic potential of generative AI: The next productivity frontier, estimated that generative AI could add the equivalent of 2.6 to 4.4 trillion US dollars a year across the 63 use cases it analysed, with much of the value in customer operations, marketing and sales, software engineering and R&D. That is an estimate of potential, not a forecast of realised value, and it depends heavily on adoption.
These are just some of the examples of how generative AI can transform various fields and industries. However, generative AI also poses some challenges and risks that need to be addressed carefully.
Challenges and risks
Ethical and social issues
Generative AI can raise questions about the ownership, authorship, responsibility and accountability of the generated content or data. Generative AI can also affect the privacy, security, trustworthiness and fairness of the data sources and users. Convincing fake images, audio and video (deepfakes) are one visible example.
Technical and practical issues
Generative AI can require a lot of computational resources, data quality and domain expertise to produce high-quality and relevant outputs. Generative AI can also be difficult to evaluate, interpret, explain and control. Language models can produce plausible but wrong answers, often called hallucinations, so outputs in technical, legal or medical settings must be checked.
Human and cultural issues
Generative AI can impact human creativity, identity, expression and values. Generative AI can also influence the cultural diversity, norms and expectations of society.
Why it matters to the semiconductor industry
Training a large model involves trillions of arithmetic operations repeated over enormous datasets, and serving it to millions of users needs fast inference at scale. This work runs mainly on data centre GPUs and dedicated AI accelerators, each paired with high-bandwidth memory (HBM) and fast interconnect. In 2023 this made AI one of the strongest sources of demand for leading-edge logic, HBM and advanced packaging, and it pushed chip designers to optimise for low-precision arithmetic, memory bandwidth and energy per operation. Smaller models are also moving onto phones and PCs, which brings neural processing units into mainstream consumer chips.
FAQ
Is generative AI the same as ChatGPT?
No. ChatGPT is one product built on a large language model. Generative AI also covers image, audio, video and code generators.
Does generative AI understand what it writes?
It models patterns in data very well, but it does not check facts the way a person does. That is why its output needs review.
What hardware does generative AI need?
Training uses large clusters of GPUs or AI accelerators with high-bandwidth memory. Inference can run on similar hardware in the cloud, or, for smaller models, on NPUs in phones and laptops.
Summary
Generative AI is a powerful and promising technology that has a lot of potential benefits but also some potential drawbacks. It should be developed and used in a responsible and ethical way that respects human dignity, rights and interests.
Your comments will be moderated before it appears here.