Member-only story
OpenAI Quietly Released GPT-3.5: Here’s What You Can Do With It
Some ideas to make the most of this mind-blowing tech
OpenAI’s GPT-3, initially released two years ago, was the first to show that AI can write in a human-like manner, albeit with some flaws.
The successor to GPT-3, likely to be called GPT-4, is expected to be released soon, possibly as early as 2023.
In the meantime, OpenAI has quietly rolled out a series of AI models based on GPT-3.5, an improved version of GPT-3. The first of these models, ChatGPT, was unveiled at the end of November.
ChatGPT is a fine-tuned version of GPT-3.5 that can engage in conversations about a variety of topics, such as prose writing, programming, script and dialogues, and explaining scientific concepts to varying degrees of complexity.
GPT-3.5 was trained on a blend of text and code published before the end of 2021, so its training stopped at this point, meaning it’s not able to access or process more recent data, events, etc.
It learned the relationships between sentences, words, and parts of words by ingesting huge amounts of data from the web, including hundreds of thousands of Wikipedia entries, social media posts, and news articles.
Rather than releasing GPT-3.5 in its entirety though, OpenAI used it to create several systems that are fine-tuned to achieve specific tasks. One of these is text-davinci-003, which is better at both long-form and high-quality writing than models built on GPT-3. It also has fewer limitations and scores higher on human preference ratings than InstructGPT, a family of GPT-3-based models released by OpenAI earlier in 2022.
What can GPT-3.5 do?
GPT-3 is accessible via the OpenAI Playground, which provides a neat user interface anyone can use.
At its simplest level, it lets you type any request directly in this front-end. There are several enhanced parameters to the right-side of the screen, including a number of models, each with their own features. The latest, text-davinci-003, is a significant improvement and creates approximately 65% longer outputs compared to the previous version (text-davinci-002).
