What is a model?
A model is a type of prediction tool in the artificial intelligence field, typically designed for a specific kind of scenario. These models are used for anticipating a variety of outcomes - from weather patterns and stock market trends to sports results and image content identification. The common thread among these models is their ability to take a certain input (like current weather data) and output a prediction (like tomorrow's forecast), often accompanied by a confidence level.
The accuracy and reliability of these predictions can vary widely. Before the advent of GPT models, there wasn't an efficient model capable of accurately predicting the continuation of a given "text input", there wasn't an efficient model capable of accurately predicting the continuation of a given "text input".
However, OpenAI pioneered a new category of models known as generative pre-trained transformers (GPT), which significantly improved this. These GPT models can effectively 'extend' an input text in many circumstances, matching or sometimes surpassing the speed and proficiency of an average human.
Since "text input" is a broadly applicable concept, these GPT models can be employed for a range of tasks, from answering queries and adhering to editing/formatting guidelines to even writing code.
So, in essence, a GPT model is a forecasting tool for textual content.
History of the different models of OpenAI
Before understanding the differences between those different models, let’s take a trip down the (short) memory lane and understand the GPT models that openAI released over the last 7 years of their existence.
- GPT (Generative Pretrained Transformer): Released in 2018, the GPT model was a breakthrough in the domain of language understanding and generation. It was pretrained on a diverse range of internet text, but it was not without its limitations, including generating creative but sometimes nonsensical responses.
- GPT-2: Introduced in 2019, GPT-2 was a direct upgrade to its predecessor, offering more parameters and better performance. Initially, due to concerns about potential misuse, the full model wasn't released. However, it was fully open-sourced later that year.
- GPT-3: The third iteration, released in 2020, increased the model capacity significantly, boasting 175 billion machine learning parameters. This model was capable of generating impressively coherent and contextually relevant responses, even outperforming its predecessors in translation, question-answering, and other tasks.
- GPT-3 fine-tuned variants: OpenAI also released some GPT-3 models that have been fine-tuned for specific tasks. For example, the ChatGPT model (technically GPT-3.5), a version of GPT-3, was fine-tuned specifically for generating conversational responses.
- GPT-4, released in 2023, is the most advanced model so far with 1 Trillion parameters. In internal testing, the model managed to score within the 94th percentile on the SAT (end of high-school exam in the USA), the 88th percentile on the LSAT (Law School Admission Test), and the 90th percentile on the Uniform Bar Exam (Lawyer Exam). These results are remarkable when compared to the previous model, GPT-3.5, which achieved scores in the 82nd percentile on the SAT, the 40th percentile on the LSAT, and just the 10th percentile on the Uniform Bar Exam.
So what are the differences Between GPT-3, GPT-3.5, and GPT-4 Models?
Models from the GPT-3 range are designed to generate text in response to specific directives, but they aren't as effective for generating fluid, conversational responses. The peak performer in this class is text-davinci-003, though it does come with a higher price tag. At the very beginning of GPT Workspace, all of our services used text-davinci-003, which is also quite slow.
The GPT-3.5 series, also known as ChatGPT, which was launched on the 1st of March, 2023, is designed with a focus on conversation. These models can also handle instructional cues quite competently, similar to text-davinci-003. However, in certain situations, they may generate responses that are somewhat more verbose or imaginative than necessary. However those are not real drawbacks since this model performs extremely well on most tasks, is also extremely fast and cheap to run. Compared to the latest release of Bard AI, the competitor of the GPT models from Google, it tends to be more logical and “hallucinates” (invents imaginary answers) way less.
The GPT-4 models, is the last model of openAI, it is currently the cutting edge of the GPT models (available only for our Premium, Team members) and most likely the smartest model available on the market right now. Extremely good at reasoning it also excels in writing, inventing and deductive reasoning. OpenAI also announced multimodality (the ability to run queries via images and other types of media, not only text). It is also capable of doing mathematical calculations extremely well, one part where GPT-3.5 just was not capable. His only and biggest drawback is that it is… slow. And also expensive, so much so that OpenAI limits (at the time of writing) to even its paying customers at 25 messages every 3 hours.
So which model to choose?
For productivity tasks, the short answer is: GPT-3.5 Turbo. It is particularly good with everything related to Google Sheets to provide quick results and fill in cells, lists and tables quickly.
However for a professional utilization and for people that don’t mind the latency, using GPT-4 will outperform it mostly everywhere. Many people are waiting for GPT-4-Turbo, but it will probably be a while before it's released.
The older models like text-davinci-003 have been simply deprecated from GPT Workspace, as they were not providing the performance expected from our users.
In conclusion, OpenAI's series of GPT models have revolutionized the field of language understanding and generation. From the initial GPT model to the latest GPT-4, each iteration has brought significant improvements in terms of parameters, performance, and specific capabilities. While GPT-3 models excel in generating text in response to specific directives, the GPT-3.5 Turbo shines in tasks related to conversation and productivity. GPT-4, despite being slower and more costly, is unparalleled in its intelligence and reasoning capabilities. The choice between these models largely depends on specific use-cases and requirements. As we anticipate the release of GPT-4 Turbo, we can only imagine the advancements it will bring. Understanding the unique features and capabilities of each model can help users make informed decisions and fully leverage the power of these AI tools. We hope this article has provided you with a deeper understanding of OpenAI's GPT models and their evolution and will help you better choose a model when using GPT Workspace.
Add-on GPT for Sheets,
Docs, Slides and Drive ↑