Exploring Generative AI and Law: ChatGPT, Midjourney, and Other Innovations | Pre-Conference Primer

Introduction

In this section, Harry introduces himself and the agenda for the session. He also provides an overview of what will be covered in the session.

Agenda

• Harry introduces himself and his role at Silk and Flatirons.

• The agenda for the session is discussed, which includes an introduction to generative artificial intelligence, chat GPT, large language models, how GPT works, word vectors, neural networks, pre-training transfer learning Transformer and attention mechanism instruction fine-tuning.

What is Generative AI?

In this section, Harry explains what artificial intelligence is and defines generative AI.

Definition of Artificial Intelligence

• Artificial intelligence is defined as using computers to solve problems that require higher order cognitive activities such as driving or chess.

Definition of Generative AI

• Generative AI is a specific set of artificial intelligence that focuses on creating creative outputs such as art, music or text.

Generative Text

In this section, Harry discusses generative text and its importance in changing society.

Importance of Generative Text

• Generative text is the biggest breakthrough among generative art forms because it can generate text using models like chat GPT or GPT in large language models.

• Chat GPT is a chat-based interface to an underlying technology called GPT which generates text.

Understanding GPT: Generative Pre-trained Transformer

In this section, the speaker introduces GPT and explains what each letter in the acronym stands for.

What is GPT?

• GPT stands for Generative Pre-trained Transformer.

• It is considered generative because it generates text one word at a time based on all the other texts that have come before it.

• It is pre-trained using vast quantities of text from the internet and books to learn patterns about reasoning and how language works.

• It uses lots of language during training to learn different patterns of language, making it known as a large language model.

• The Transformer is an architecture or way of doing something called Deep learning that allows models like GPT to understand the context of words being asked around it.

History and Development of GPT

In this section, the speaker discusses the history and development of GPT.

History and Development

• OpenAI developed GPT by building upon Google's Transformer architecture.

• The original version of Chatbot came in November 2022, which was a huge breakthrough based on two improvements made by OpenAI called instruction fine-tuning and reinforcement learning from human feedback.

• The latest version, GPT4, was released last month. It is incredibly advanced but only available to those who pay for it or use Bing chat.

Advancements in AI Technology

In this section, the speaker highlights how much AI technology has improved over time.

Advancements in AI Technology

• The speaker emphasizes how important it is to understand how bad the technology was just last year compared to now.

• The speaker shares a funny poem about sea law generated using GPT.

• GPT can generate text one word at a time based on all the other texts that have come before it, making it considered generative.

• Pre-training is an expensive process of getting AI to learn patterns of human thought and language.

• The Transformer is an architecture or way of doing something called Deep learning that allows models like GPT to understand the context of words being asked around it.

Open AI Playground

The Open AI Playground allows users to look at old models and play with them. These models are frozen in time, allowing users to see what they were capable of in the past.

Common Sense Questions

• Common sense questions have been used to test machines' reasoning abilities.

• A question like "how many legs does an apple have?" would be difficult for a machine to answer correctly.

• In the past, machines would almost always fail these types of questions.

Advancements in AI

• In November 2022, GPT emerged with huge advancements in its ability to engage in reasoning and problem solving.

• This was unexpected among AI researchers who thought it would only be able to generate text.

• GPT can now solve hard problems that involve coins and puzzles.

• It can also generate legal documents such as patent applications and contracts.

Understanding Language

• GPT's ability to understand language is shocking because it can take any input and respond sensibly as if it understood what was going on.

• However, we should not anthropomorphize these systems since they are still just machines.

GPT's Ability to Generate Legal Documents

GPT has made significant advancements in generating legal documents such as patent applications and contracts.

Generating Legal Documents

• GPT can generate a first draft of a patent application that could pass muster at the patent office.

• It can also generate a first draft of just about any legal document, although it may not be suitable for submission directly to a judge.

• While it does make mistakes, it is still amazing that it can generate such high-quality legal documents in just one year.

Conclusion

• GPT's advancements in language understanding and document generation are impressive and unexpected.

• It will be interesting to see what other advancements will be made in the future.

Breakthrough in Artificial Intelligence

In this section, the speaker discusses a breakthrough in artificial intelligence that he believes is one of the biggest in the last 20 years.

Word Vectors

• Word vectors are also called word embeddings and were invented by Google and Stanford.

• The idea behind word vectors is that you can encode the meaning of words mathematically as a list of numbers.

• Each column represents a different aspect of the word, such as whether it's an animal or has fur.

• The model learns to associate words with similar contexts and starts to learn mathematical similarities between them.

State-of-the-Art AI Model

• The speaker describes how he tested an AI model's reasoning abilities by asking it about an apple and someone eating it in another room.

• He was impressed with the model's ability to understand intentions and emotions, calling it a "theory of mind."

• The speaker believes this breakthrough is one of the biggest in artificial intelligence in the last 20 years.

Conclusion

• The speaker concludes by saying that he will explain how this AI model works and encourages viewers to take a deep dive into understanding it.

Animals and Math

In this section, the speaker talks about how animals are grouped together mathematically and how pets are nudged closer to each other. The speaker also discusses how word meaning can be encoded using numbers.

How Animals Are Grouped Together

• The x-axis and y-axis are used to group animals together.

• Pets are nudged closer to each other in the grouping.

• Cows are not considered pets.

Encoding Word Meaning Using Numbers

• Word vectors or embeddings encode word meaning using numbers.

• This breakthrough has allowed for GPT (Generative Pre-trained Transformer) technology to exist.

Neural Networks and Deep Learning

In this section, the speaker talks about neural networks, deep learning, and how they work.

Neural Networks

• Neural networks were invented in the 1940s but weren't perfected until 2012.

• They learn patterns from data and encode them.

• They're loosely inspired by the human brain.

Deep Learning

• Deep learning is an area of machine learning that involves scaling up neural networks on lots of data.

• It's a super flexible technique for encoding patterns.

Committees and Neural Networks

In this section, the speaker explains how neural networks work by comparing them to committees.

Committees in Neural Networks

• Each connection in a neural network represents how important it is to the next connection.

• It's like a series of committees voting on what comes next in a sentence.

• Any pattern can be encoded in a neural network.

GPT Technology

In this section, the speaker talks about how GPT technology works.

How GPT Works

• GPT uses neural networks to learn patterns from data.

• Word vectors are used to represent words mathematically.

• GPT was trained on the whole internet and two million books.

• It predicts the next word in a sentence and learns from its mistakes.

Training and Mathematics of Language Models

In this section, the speaker explains how language models are trained using billions of parameters and the mathematics behind it. The speaker also introduces the Transformer, which is Google's invention that takes advantage of representing words as vectors.

Training Language Models

• The process of training a language model involves nudging billions of parameters up and down to teach it English language and reasoning.

• When a language model predicts an incorrect answer, it compares what it predicted with what it knows the text should be. It then demotes the wrong prediction and promotes the correct one.

• After adjusting its weights or parameters, the model is tested again to see if it produces better results.

Mathematics Behind Language Models

• Google's invention called Transformer takes advantage of representing words as vectors or numbers.

• Words with multiple meanings can be represented mathematically by nudging them towards one meaning or another based on context.

• The Transformer architecture has 96 layers that encode context mathematically along every word to understand what was asked and what was said so far.

Nudging Words Mathematically Using Transformers

In this section, the speaker explains how Transformers nudge words mathematically towards their intended meaning based on context.

Nudging Words Mathematically

• Transformers take into account context by encoding each word mathematically based on what was asked and what was said so far.

• The Transformer nudges words mathematically towards their intended meaning based on context.

Breakthrough in AI: Transfer Learning

In this section, the speaker discusses the breakthrough in AI called transfer learning and how it differs from other AI systems.

Transfer Learning

• OpenAI's breakthrough was the idea of transfer learning.

• Unlike other AI systems that were designed for narrow tasks, transfer learning involves training a general model on everything and then applying it to almost any task.

• The model is trained on the entire human textual output compressed into a neural network of 175 billion parameters.

• GPT-3 had capabilities that were lurking within it, which researchers at OpenAI believed they could pull out by clever engineering.

Instruction Fine Tuning and Reward Model

In this section, the speaker explains how instruction fine-tuning and reward models were used to improve GPT-3's performance.

Instruction Fine Tuning

• OpenAI used something called instruction fine-tuning to nudge GPT-3 into chat GPT.

• They gave GPT-3 thousands of examples of good questions and answers so that it could learn what a good output looks like.

Reward Model

• OpenAI came up with an even smarter idea where they got humans to rank five outputs produced by GPT-3 from best to worst.

• They then took those scores and stuck them in another AI system called a reward model, which taught the system what good output looks like.

• The reward model trained up another AI system, which produced chat GPT by going a billion times good output bad.

Chat GPT and Its Capabilities

In this section, the speaker discusses how chat GPT was created and its capabilities.

Chat GPT

• Chat GPT is the result of using instruction fine-tuning and reward models to improve GPT-3's performance.

• It is not perfect and makes mistakes called hallucinations, but it is amazing and the biggest breakthrough in AI that the speaker has seen in their 20 years of studying it.

Capabilities

• Transfer learning allows for applying chat GPT to almost any task, including poetry and coding.

• While disruptive for jobs, it will also accelerate research and create new opportunities.