And the democratization of AI implementation
When asked ‘How do you think AI will change us [humans]?’ Jeff Bezos prefaced his answer with a thought-provoking comment.
“Large language models in their current form are not inventions. They are discoveries.”
And to illustrate it with his examples, Galileo invented the telescope to later discover the moons of Jupyter. He anchors LLMs to the far end of the invention-discovery spectrum by arguing that “[as opposed to] a 787 [which] is an engineered object, LLMs are much more like discoveries. We are constantly being surprised by their capabilities”.
Large language models — and you can extend it to generative AI as a whole — are already driving excitingly innovative use cases everywhere. They are being wrapped in every shape and form and shipped to users as personalized customer service chatbots, writing assistants, and even embedded in new features at longstanding and widely used tools such as Adobe Photoshop. So it’s real technology. They are engineered tools — inventions already — right?
Well, sometimes the wrapper is just too thin. When you stop to reflect on Bezos’ comment¹ and start to look at each release of a new GPT by OpenAI or a new state-of-the-art open-source model by Meta solely as the outcome of research, you will notice what might be actually happening — we are getting highly proficient at bridging the gap from discoveries to products.
In this article, I reflect on how a new industry is democratizing generative AI implementation and how byproducts of research and broadly available tools are now closer than ever.
Generative Artificial Intelligence has never been the sole driver of disruptive implementations of AI, nor has it been a ubiquitous field in modern machine learning research. Quite the opposite, until 2020, although deep learning papers dominated Google Scholar’s list of most influential publications, computer vision research centred on object detection and image recognition attracted the highest number of citations, and “Attention is all you need”, the work that introduced the Transformers architecture, was only the 4th most cited in that year, even though Google was already…
