The mics were live and tape was rolling in the studio where the Miles Davis Quintet was recording dozens of tunes in 1956 for Prestige Records.When an engineer asked for the next song's title, Davis shot back, I'll play it, and tell you what it is later.
Like the prolific jazz trumpeter and composer, researchers have been generating AI models at a feverish pace, exploring new architectures and use cases. Focused on plowing new ground, they sometimes leave to others the job of categorizing their work.
A team of more than a hundred Stanford researchers collaborated to do just that in a 214-page paper released in the summer of 2021.
In a 2021 paper, researchers reported that foundation models are finding a wide array of uses. They said transformer models, large language models (LLMs) and other neural networks still being built are part of an important new category they dubbed foundation models.
Foundation Models Defined A foundation model is an AI neural network - trained on mountains of raw data, generally with unsupervised learning - that can be adapted to accomplish a broad range of tasks, the paper said.
The sheer scale and scope of foundation models from the last few years have stretched our imagination of what's possible, they wrote.
Two important concepts help define this umbrella category: Data gathering is easier, and opportunities are as wide as the horizon.
No Labels, Lots of Opportunity Foundation models generally learn from unlabeled datasets, saving the time and expense of manually describing each item in massive collections.
Earlier neural networks were narrowly tuned for specific tasks. With a little fine-tuning, foundation models can handle jobs from translating text to analyzing medical images.
Foundation models are demonstrating impressive behavior, and they're being deployed at scale, the group said on the website of its research center formed to study them. So far, they've posted more than 50 papers on foundation models from in-house researchers alone.
I think we've uncovered a very small fraction of the capabilities of existing foundation models, let alone future ones, said Percy Liang, the center's director, in the opening talk of the first workshop on foundation models.
AI's Emergence and Homogenization In that talk, Liang coined two terms to describe foundation models:
Emergence refers to AI features still being discovered, such as the many nascent skills in foundation models. He calls the blending of AI algorithms and model architectures homogenization, a trend that helped form foundation models. (See chart below.)
The field continues to move fast.
A year after the group defined foundation models, other tech watchers coined a related term - generative AI. It's an umbrella term for transformers, large language models, diffusion models and other neural networks capturing people's imaginations because they can create text, images, music, software and more.
Generative AI has the potential to yield trillions of dollars of economic value, said executives from the venture firm Sequoia Capital who shared their views in a recent AI Podcast.
A Brief History of Foundation Models We are in a time where simple methods like neural networks are giving us an explosion of new capabilities, said Ashish Vaswani, an entrepreneur and former senior staff research scientist at Google Brain who led work on the seminal 2017 paper on transformers.
That work inspired researchers who created BERT and other large language models, making 2018 a watershed moment for natural language processing, a report on AI said at the end of that year.
Google released BERT as open-source software, spawning a family of follow-ons and setting off a race to build ever larger, more powerful LLMs. Then it applied the technology to its search engine so users could ask questions in simple sentences.
In 2020, researchers at OpenAI announced another landmark transformer, GPT-3. Within weeks, people were using it to create poems, programs, songs, websites and more.
Language models have a wide range of beneficial applications for society, the researchers wrote.
Their work also showed how large and compute-intensive these models can be. GPT-3 was trained on a dataset with nearly a trillion words, and it sports a whopping 175 billion parameters, a key measure of the power and complexity of neural networks.
The growth in compute demands for foundation models. (Source: GPT-3 paper) I just remember being kind of blown away by the things that it could do, said Liang, speaking of GPT-3 in a podcast.
The latest iteration, ChatGPT - trained on 10,000 NVIDIA GPUs - is even more engaging, attracting over 100 million users in just two months. Its release has been called the iPhone moment for AI because it helped so many people see how they could use the technology.
One timeline describes the path from early AI research to ChatGPT. (Source: blog.bytebytego.com) From Text to Images About the same time ChatGPT debuted, another class of neural networks, called diffusion models, made a splash. Their ability to turn text descriptions into artistic images attracted casual users to create amazing images that went viral on social media.
The first paper to describe a diffusion model arrived with little fanfare in 2015. But like transformers, the new technique soon caught fire.
Researchers posted more than 200 papers on diffusion models last year, according to a list maintained by James Thornton, an AI researcher at the University of Oxford.
In a tweet, Midjourney CEO David Holz revealed that his diffusion-based, text-to-image service has more than 4.4 million users. Serving them requires more than 10,000 NVIDIA GPUs mainly for AI inference, he said in an interview (subscription required).
Dozens of Models in Use Hundreds of foundation models are now available
