Get startedGet started for free

The evolution of generative AI

1. The evolution of generative AI

Let's explore the key factors and milestones that have brought generative AI into products that people around the world use every day.

2. Generative AI burst on the scene in 2023

Generative AI burst onto the global mainstream in 2023, with the launch of multiple generative AI consumer products. The chatbot ChatGPT made history when it reached 100 million monthly users in only two months, something that took viral social media platforms like TikTok and Instagram several times longer to achieve.

3. Key factors driving development

Leading up to that historical moment, the field of generative AI evolved over many years, enabled by increased computing power, massive datasets for training, market and geopolitical competition, and novel model designs and tools.

4. Computational power allowed large models

By 2023, models already required 100 million times more computing power than models from 10 years earlier. Innovations allowing multiple computations to happen simultaneously, known as parallelization, allowed the training of larger and more complex models. It was enabled by specialized processors, such as Graphics Processing Units, or GPUs, and Tensor Processing Units, or TPUs. Cloud computing has also driven generative AI forward, enabling researchers to access and scale computing resources as needed. Finally, improvements in software frameworks and libraries optimized the utilization of computing power.

5. Models improved with massive datasets

As data availability has exploded exponentially in recent years, data-hungry generative AI models have had more to train on. Additionally, breakthroughs in techniques for creating synthetic data that can augment and enhance real data further scale the availability of training data.

6. Competitive pressures encouraged faster development

Big Tech companies and governments also spurred generative AI development to gain commercial or political advantages against their peers. But the core of generative AI's evolution is innovation in models.

7. GANs unleashed high quality generation

We're already familiar with the first model innovation: generative adversarial networks, or GANs. Remember that these models consist of generator models and discriminator models that compete with one another. Introduced in 2014, GANs brought a massive leap in the quality of results that generative techniques could produce.

8. Transformers brought context and coherence

The second model innovation we'll discuss is the transformer. This powerful model type is designed to understand and process text by considering multiple words and their relationships at once, rather than focusing on individual words one at a time. Consider the sentence, "The animal didn't cross the street because it was too tired." Here, the model recognizes that "it" refers to the animal, emphasized by the blue coloring. Now, take the sentence, "The animal didn't cross the street because it was too wide." In this case, the model understands that "it" refers to the street.

9. Transformers brought context and coherence

Transformers excel at grasping the context of a given text, which allows them to generate more coherent responses. By analyzing the relationships between words and seeing the text as a whole, transformers can generate responses that feel natural and informative, like an educational chatbot that keeps a student on track to complete a lesson.

10. RLHF engaged user feedback

A third important innovation is Reinforcement Learning with Human Feedback, or RLHF. This technique improves models by applying feedback from users. Reinforcement learning teaches models through trial-and-error interactions. This allows them to learn how to achieve complex and specific goals. The human feedback part of RLHF comes from users scoring model responses. This feedback becomes part of a retraining process that helps the model outputs better match what users score highly.

11. RLHF engaged user feedback

Image generation AI company Midjourney allows users to rate images. Midjourney can then use this feedback to steer their model toward higher-rated responses and away from lower-rated ones. Public products can accumulate massive amounts of feedback from their users, creating a flywheel that speeds up development further.

12. Let's practice!

Time for some exercises!

Create Your Free Account

or

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.