How 2022 became the year of generative AI
There has been a lot of enthusiasm and hype about generative AI (artificial intelligence) in 2022.
Updated
On this page
Updated November 2026. This was first published on 2 December 2022. The idea is the same. Any number below is what the original piece said then. It is not a new count.
Key takeaways
- There has been a lot of enthusiasm and hype about generative AI (artificial intelligence) in 2022.
- Social media sites such as Twitter and Reddit are packed with visuals made using generative machine learning models such as DALL-E and Stable Diffusion.
- Startups designing products on top of generative models are gaining investment despite the market slump.
There has been a lot of enthusiasm and hype about generative AI (artificial intelligence) in 2022. Social media sites such as Twitter and Reddit are packed with visuals made using generative machine learning models such as DALL-E and Stable Diffusion. Startups designing products on top of generative models are gaining investment despite the market slump. And Big Tech firms are incorporating generative models into their mainstream offerings.
With a few significant exceptions, most of the technologies we’re witnessing now have existed for several years. But, the convergence of multiple themes has made it possible to productize generative models and bring them to daily applications. The field still has many hurdles to solve. there is no question that the market for generative AI is likely to rise in 2023.
Generative AI became popular in 2014 with the emergence of generative adversarial networks (GANs), a sort of deep learning architecture that could build realistic pictures — such as faces — from noise maps. Scientists then constructed various kinds of GANs to do different tasks such as copying the style of one image to another. GANs and the variational autoencoders (VAE), another deep learning architecture, eventually ushered in the era of deepfakes, an AI approach that changes photos and videos to swap one person’s face for another.
2017 witnessed the birth of the transformer, a deep learning architecture supporting large language models (LLMs) such as GPT-3, LaMDA and Gopher. The transformer is used to create text, software code and even protein structures. A version of the transformer, the “vision transformer,” is also employed for visual tasks such as picture categorization. An older version of OpenAI’s DALL-E employed the transformer to produce graphics from text.
Transformers are scalable, which means their performance and accuracy improve as they are grown larger and fed more data. But more crucially, transformer models may be trained using unsupervised or self-supervised learning, meaning they require no or very little human-annotated data. has been one of the key bottlenecks of deep learning.
Contrastive Language-Image Pre-training (CLIP), a technology released by OpenAI in 2021, became crucial in text-to-image generators. CLIP is particularly good in learning shared embeddings between pictures and text by learning from image-caption pairs acquired from the internet. CLIP and diffusion (another deep learning approach for producing pictures from noise) were utilised in OpenAI’s DALLE-2 to produce high-resolution photos with outstanding detail and quality.
As we progressed near 2022, better algorithms, larger models and richer datasets helped enhance the output of generative models, making better graphics, writing high-quality software code and generating lengthy stretches of (mostly) comprehensible prose.
Generative models were initially offered as systems that could take up enormous portions of creative activity. GANs became famous for producing entire pictures with little input. LLMs like GPT-3 earned the news for composing complete essays.
But as the discipline has grown, it has become obvious that generative models are unreliable when left on their own. Many experts believe that existing deep learning models — no matter how huge they are — lack some of the essential components of intelligence, which renders them prone to committing unforeseen blunders.
Product teams are realising that generative models work best when they are applied in ways that allow greater power to consumers.
The last year has seen numerous products that use generative models in clever, human-centric ways. For example, Copy AI, a programme that employs GPT-3 to produce blog articles, features an interactive interface in which the writer and the LLM construct the outline of the article and flesh it out together.
Applications designed with DALL-E 2 and Stable Diffusion Also highlight user control with what it can do that allow for changing, renewing or customising the output of the generative model.
As Douglas Eck, chief scientist at Google Research, noted at a recent AI conference, “It’s no longer about a generative model that delivers a realistic picture. It’s about producing something that you produced yourself. Technology should fulfil our need to have agency and creative control over what we do.”.
In parallel with the algorithms and applications, the computational infrastructure and platforms for generative models have grown. This has allowed many organisations integrate generative AI into their applications without the requirement for the specialist expertise necessary to build up and run generative models.
Product teams with seasoned machine learning engineers can employ open-source generative models such as BLOOM and Stable Diffusion. Meanwhile, teams who don’t have in-house machine learning ability can select from a broad choice of solutions such as OpenAI API, Microsoft Azure, and HuggingFace Inference Endpoints. These platforms abstract away the complexity of building up the models and executing them at scale.
Also of interest is the emergence of MLops platforms. are making it easy to set up whole pipelines for obtaining feedback data, versioning datasets and models, and fine-tuning models for specific applications.
The generative AI sector still has obstacles to solve, including ethical and copyright problems.
But it is exciting to see the generative AI space grow. For the present, the biggest beneficiaries are Big Tech businesses with data, compute capacity and an established market and products to give the additional value of generative models. For example, Microsoft is taking use of its cloud infrastructure, its unique access to OpenAI’s technology and the vast market for its office and creativity products to bring the potential of generative models to its customers.
Adobe is also working to include generative AI in its video and graphic design tools. And Google also has numerous generative AI products in the works.
Down the future, though, the actual power of generative AI may express itself in new markets. Who knows, maybe generative AI may bring in a new age of applications that we had never conceived of before.
Questions and answers #
What is the main point? #
There has been a lot of enthusiasm and hype about generative AI (artificial intelligence) in 2022.
What else did the original piece say? #
Social media sites such as Twitter and Reddit are packed with visuals made using generative machine learning models such as DALL-E and Stable Diffusion.