How AI writers work

In recent years, the field of artificial intelligence has experienced a monumental leap forward in the quest to create intelligent machines. This breakthrough can be attributed to the emergence of Large Language Models (LLMs), which are based on research aimed at replicating the intricacies of the human brain. LLMs have ushered in a new era known as generative AI, where software exhibits the remarkable ability to craft text, images, and computer code with a level of sophistication that closely mimics human aptitude. Join us on a journey as we explore the fascinating world of LLMs, their transformative potential, and the challenges they pose.

The Birth of LLMs

LLMs, underpinned by the transformative transformer model, have sparked a revolution in the AI landscape. This technology, introduced by Google researchers in 2017, serves as the foundation for LLMs and is instrumental in their exceptional capabilities. Slav Petrov, a senior researcher at Google, acknowledges the profound impact of transformer technology across various domains, from healthcare and robotics to enhance human creativity.

One of the most touted benefits of LLMs is their capacity to enhance productivity through text generation and analysis. However, this very capability poses a dual threat to human employment, with Goldman Sachs estimating that it could automate the equivalent of 300 million full-time jobs across major economies, potentially leading to widespread unemployment.

The Intricate Process of Text Generation

To understand how LLMs generate text, we must first delve into their intricate inner workings. The journey begins with words, which are broken down into tokens, fundamental units that can be encoded. These tokens are derived from vast training datasets, consisting of billions of words collected from the internet. The model then processes this data, creating word embeddings—lists of values that quantify various aspects of a word's meaning.

These word embeddings, often composed of hundreds of values, allow LLMs to grasp the nuances of word meanings. For example, words like "mountain" and "hill" may have different contexts, but their embeddings reveal their close linguistic relationship. By reducing these values to just two dimensions, the model can visualize the proximity between words more clearly.

The Power of Transformers

However, what truly sets LLMs apart is the transformative power of transformers. These models process entire sequences, whether sentences, paragraphs, or entire articles, in a holistic manner. This ability to capture context and patterns simultaneously significantly enhances their text generation accuracy and efficiency. The transformer model, introduced in a groundbreaking research paper by a group of Google AI researchers in 2017, marked the dawn of the generative AI era.

A central concept of transformers is self-attention, which enables LLMs to understand the relationships between words. Unlike previous AI translation methods that processed words sequentially, self-attention allows the transformer to compute all words in a sentence simultaneously. This contextual understanding empowers LLMs with advanced language comprehension capabilities.

Examples of Self-Attention in Action

Consider the sentence: "She plays a sweet melody on the piano." With self-attention, the model recognizes that "melody" is related to "piano." If we alter the sentence to "He plays a sweet melody on the guitar," the model adapts its understanding accordingly. Even when combining sentences, the model maintains its contextual awareness, discerning the correct meaning of each word.

This self-attention functionality goes beyond words with multiple meanings. For instance, in the sentence "He likes to run in the park," self-attention correctly identifies "run" as an action in a recreational context. Changing "park" to "gym" results in the model associating "run" with exercise.

The Arrival of GPT-4 and Its Impact

One of the most advanced LLMs to date is GPT-4, created by OpenAI. This model exhibits "human-level performance" on numerous academic and professional benchmarks, such as the US bar exam and SAT school exams. With the capacity to generate and analyze vast volumes of text, GPT-4 has reshaped the tech industry, prompting major players like Google, Meta, and Microsoft to enter the race alongside smaller startups.

These LLMs, including Google's PaLM, Anthropic's Claude, Meta's LLaMA, and Cohere's Command, are already being adopted across various industries. However, they face legal challenges related to their use of copyrighted text, images, and audio scraped from the web.

The Challenge of Hallucination

Despite their incredible capabilities, LLMs are not infallible. Their predictive nature can lead to "hallucination," where they generate fabricated information, including numbers, names, dates, and even entire articles. This phenomenon poses challenges, as users have reported instances of misleading or false information generated by LLMs.

To address this issue, researchers are working on "grounding," a process that cross-checks LLM outputs against web search results and provides citations for verification. Additionally, human feedback, known as reinforcement learning by human feedback (RLHF), helps improve output quality.

Beyond Language: The Transformer's Real Power

Transformers have unlocked a host of cutting-edge AI applications that extend beyond language. These models can recognize and predict patterns in various domains, from analyzing medical images to predicting stock market trends and assisting in drug discovery.

In the words of Aidan Gomez, CEO of AI startup Cohere and co-author of the transformer paper, "Take this simple model that predicts the next word, and it...can do anything." These models, trained on vast datasets, have become the driving force behind AI's next frontier, surpassing anything that came before.

As we embrace the era of LLMs and generative AI, the possibilities are boundless, and the challenges are formidable. These intelligent machines are reshaping industries and pushing the boundaries of what AI can achieve. While they may have their limitations, the transformative potential of LLMs is a testament to the relentless pursuit of AI excellence.