Artificial Intelligence (AI) went from a niche tool to having permeated business, technology, art, and culture in a matter of months. Having worked on Natural Language Processing ourselves before the craze, we’ve seen the evolution of theory and application. The sudden explosion in AI adoption was the free, unlimited access to ChatGPT by OpenAI at the end of 2022. The base for the application, GPT 3, was already 2 years old by the time it was made available for public use.
The rapid, massive-scale, and generally positive feedback has encouraged many competitors to develop their own models and systems. The sudden spike in demand encouraged production of special-use hardware, bringing costs down.
What We'll Cover
What was AI
You’ve probably heard AI defined several times already, but it’s important to remember that we’ve had artificial intelligence for decades. It was originally common to ascribe “AI” to any program that does tasks humans would do. Unlike a repetitive task, coded with a macro, for example, pre-2022 AI was typically relegated to gaming and some algorithms that mimicked human behavior.
Even before 2010, some chat-bots entered the market that would frame answers with a degree of flexibility that approached human-like replies. These early chat-bots, however, often fell in an uncanny valley of awkward discussion. They would readily go off topic, provide incorrect or inappropriate information, or would communicate with bizarre grammar.
So what’s changed?
In the latter half of the 2010s, we observed increased attention to generative AI. The “trick” to a successful model was avoiding the uncanny valley of mostly-human responses, or noisy images. New techniques for machine learning improved training times and yielded increasingly useful results. We took on a project leveraging the keras python library to develop an LLM with church bulletins. The output was a bot that could provide talking points for certain times of the liturgical year.
Modern AI relies on a large pool of data with which to synthesize something unique when prompted by someone. Because the base data came from people’s interactions, the output is more familiar – as if a human provided the result. This is generative AI.
So, why aren’t we in the news? Well, at the end of the day, successful AI requires an immensely massive amount of data to train models from – more than we could collect and use on our own. Today’s major players in AI use their existing infrastructure to train their models. They’ve created a massive pool of data from which users can fine-tune for what they care about.
Key Terms
A number of new terms and abbreviations have entered the mainstream when talking about AI. Sometimes, you’ll find them strung together in a sentence, which can cause some confusion as to why one tool is different from another. In this primer, let’s go over some common terms:
Large Language Model (LLM)
LLMs are a specific type of generative AI designed to understand and generate human language. They are trained on vast amounts of text data. Fine tuning LLMs enable specialized outputs for specific tasks like translation, summarization, or chat-bots. They excel at tasks involving context, such as conversation and content generation. Frankly, we believe the key to a good AI tool is a good LLM. Put good in, get good out.
Vision Model
The equivalent of an LLM for image generation can be called a “vision model” or “image model”. Vision models generate images rather than text. It’s worth noting here that companies and enthusiasts are creating models for music, video, and more all the time. Many products are hitting the market every month.
Fine-Tuning
The process of adapting a pre-trained model to specific tasks. Fine-tuning is a crucial process in machine learning (ML) where a pre-trained model is adapted to perform a specific task or domain. This is typically done by people or companies who use ai products.
Machine Learning (ML)
Machine learning uses algorithms trained on data sets to create models. These models enable machines to perform tasks that would otherwise only be possible for humans. Machine learning focuses on the use of data and algorithms to imitate the way humans learn. It gradually improves accuracy by learning from examples. This process is done by the providers of AI products.
Hallucinating
In the context of AI, a hallucination is when a bot responds with incorrect information or veers off topic. This is a common concern among users and regulators who suspect AI could promulgate misinformation. For programmers, hallucinations can generate ineffective code and thus prove counterproductive. More mainstream AI tools, such as from Google and Microsoft, provide sources in their replies for fact checking and further reading.
Transformer
A neural network architecture that excels at handling sequential data by considering context. A neural network is the model trained on an LLM which provides the context.
Generative Pre-trained Transformer (GPT)
Stands for Generative Pre-trained Transformer. The “Generative” part refers to its ability to generate text or other content, the “Pre-trained” part indicates that the model is initially trained on large amounts of data before being fine-tuned for specific tasks, and “Transformer” refers to the specific architecture used in the model, which is particularly effective for natural language processing tasks.
Is AI The future?
Yes and no. We predict it will resemble the hype cycle of social media platforms. At first, they did a fantastic job at their intended purpose. Then they added various features and guardrails. Regulations then come into force. Then, international alternatives came online. Social media today is a diverse landscape where different platforms cater to different audiences.
Social media remains a useful tool for business, but today there are a lot of limitations, and tiers of functionality available for a fee. Accountability for platform owners lies with shareholders and regulations. People can see their accounts frozen or deleted for violating terms of use, which always change.
We see AI going this route – remaining useful but becoming more muted for general audiences. It has already changed dramatically with watermarking, source citations, and guardrails. Chat-bots provide clearer answers with fewer hallucinations. Popular AI tools are becoming available in free and paid versions. The same way businesses adopted websites in the 90s, went on social media in the 2010s, the 2020s may be the decade we see AI – in some form or other – present in every modern enterprise.
