The Big Thing in 2023: ChatGPT

Since years now, everyone is talking about and tagging their software with Artificial Intelligence labels – but most of those are not much more than speech or text recognition combined with key word assignments, decision trees and maybe some heuristics. Trying to have a context-aware interaction even with Alexa, Siri or Cortana clearly unveils those shortcomings. If it’s getting complicated, they will offer you just a web search based on your keywords –?not very intelligent and far from a “personal digital assistant”.

This is about to change drastically: since a month now, “ChatGPT” from OpenAI was released, which is far more a game-changer than any Web3 or Metaverse thing in the near future.

What is ChatGPT and OpenAI?

As an AI language model, ChatGPT is able to provide information and assistance on a wide range of topics. This includes answering questions, providing definitions and explanations, offering advice and suggestions, and helping to clarify information. ChatGPT can also engage in conversation and carry out dialog on various subjects.

ChatGPT is based on GPT-3 and therefor has one of the largest language models which enable it to perform many language tasks, such as translation, summarization, question answering, and text generation, often with a high level of accuracy.

GPT-3 has 175 billion parameters and was trained on 570 gigabytes of text. For comparison, its predecessor, GPT-2, was over 100 times smaller at 1.5 billion parameters.

This increase in scale drastically changes the behavior of the model – GPT-3 is able to perform tasks it was not explicitly trained on, like translating sentences from English to French, with few to no training examples. Additionally, human feedback is taken into account by the model to refine and improve its results.

OpenAI is a research organization that aims to promote and develop friendly artificial intelligence in a way that benefits humanity as a whole. It was founded in 2015 by a group of high-profile tech leaders, including Elon Musk and Sam Altman, with the goal of advancing AI research in a responsible and safe manner. OpenAI is now a profit-capped organization with also a large investment from Microsoft.?

ChatGPT is currently having a waiting list for new users due to the heavy demand and high resource requirements to cater to those demands, but it’s worth waiting and soon there will be coming many implementations of ChatGPT within other platforms and applications.

What can ChatGPT do?

One of the most remarkable abilities is to create human-like texts for all kind of purposes like writing an email, official letters, text or book summaries which are not just a random aggregation of web search results, but actually well-written text with a good structure, great language and also offering different viewpoints and additional aspects on related topics.

If you ask for example “Why can’t pigs fly in space?”, ChatGPT will provide you with an elaborated answer over three paragraphs discussing different aspects of physical inability, inherent difficulties for space flight itself and a summary result.

Even more interesting is, if you ask further: in this example continuing with “Would it be possible, if they had wings?”, ChatGPT will go deeper into the discussion of details (e. g. required time and resources to train pigs to adhere to flight control instructions). The most remarkable thing about this is the context-awareness: you can actually discuss topics and go deeper into details, which is a major step towards an actual helpful personal digital assistant.

Aside from effectively writing text for you, an even more efficient usage is to have it create programming code: instead of searching Google for code snippets, evaluating the result relevance and adapting some example to your needs, you can just ask ChatGPT to create the required code in any programming language – and again have it refine its own results based on your continued feedback. This is incredibly good and efficient to be integrated directly into any program, Excel macro or web page.

What’s the impact?

Upon wider adoption of ChatGPT and integration into popular platforms like smartphones, this has the potential to really become an actual “Digital Personal Assistant”: it can write official emails, send personalized birthday greetings, create complete presentation outlines or texts – all just based on a single request line (or “prompt”). All of this is already possible and implemented in different web tools.

While the OpenAI based image and video creation in “Dall-E 2” that was largely discussed in 2022 (and used for the illustration of this article) “only” mostly impacts efficiency and jobs of media creators, ChatGPT has a much wider base of potential users and impact –?basically everyone writing text or searching for specific answers or explanations to a topic.

What are the risks and limitations?

“Everyone writing text” also includes school and university students: getting a summary of “Romeo & Juliet” or explaining “Nuclear Fusion processes, their risks and advantages” gives you a 300+ words essay within a few seconds. Asking for more detail returns another 300+ words on a more detailed level without replicating any of the initial answer. This means it’s even easier for students to just “create & paste” with even less effort than searching the web or creating something themselves.

Similar to search engines to detect plagiarisms in texts, there are now also web services to detect AI-generated texts to enable detection of these. And again, there are already web services to re-write AI-generated texts to make them sound even more humanly created. The most popular coding Q&A site “Stack Overflow” recently banned ChatGPT answers from its site as it was flooded with those AI generated answers which often were not correct or appropriate answers.

While the resulting output quality is already very good, of course it’s not perfect and sometimes even wrong (e. g. “The mammal that lays the largest eggs is the elephant”). It also happens that results vary largely if the prompt is just rephrased slightly. The problem is that ChatGPT will always sound perfectly right and reassured about its answers and even give you logic and detailed argumentation to an incorrect answer. So, in any case, results should always just be a “starting point” to be reviewed and adapted to your needs.

Don’t believe me? This article also contains multiple paragraphs that were written by ChatGPT. Can you guess which ones?

There are also topics like politics or dangerous topics (“How to build a Molotov cocktail?”) that are not answered but produce an answer pointing to the model’s neutrality in opinions and related risks. However, sometimes those limitations can currently still be circumvented by asking more often or with prompt variations.

Another limitation is the training set cutoff: ChatGPT currently doesn’t know anything from after 2021, which will only change when the model is re-trained with additional data at a later point of time. This is a common AI pattern: it’s still “learning” from user feedback and getting better in its answers by that, but the underlying data in the language model is a fixed set at a specific point in time and is not continuously extended.

How can I use it?

As stated above, there’s currently a waiting list to join ChatGPT, but I would assume this will become wider available throughout 2023.

Additionally you can use platforms that are already available and leverage ChatGPT to create PowerPoint presentations (Tome: beta.tome.app), writing assistance (Notion AI: notion.so), complete websites (Durable: durable.co)??etc.

#AI #ChatGPT #OpenAI #2023