The Latest Version of AI Falters at Vocabulary Guessing Games
The latest version of artificial intelligence, ChatGPT4, has failed to win a popular vocabulary guessing game called “Wordle.” Despite being trained on about 500 billion words from various sources such as Wikipedia, public domain books, scientific articles, and text from websites. Even feeding the entire Library of Congress into the AI would not improve its performance for now. The neural network uses a complicated mathematical function that maps inputs and outputs, which must be numbers for the program to function. ChatGPT4 cannot reason about letters or see words as text because the deep neural network does not have access to them. Words inputted into the AI must be translated into numbers using a tokenizer program that converts common sequences of characters found in text.
Ways to Improve ChatGPT’s Performance in Word Games
To improve ChatGPT’s performance in playing word games, one solution could be augmenting its training data to include mappings of every letter position within every word in its dictionary. Another possibility is programming future language models like ChatGPT to generate specific code for games like Wordle. However, despite these efforts, there are still limitations in the AI’s ability to understand language.
Prompt engineering is another way to improve the output of ChatGPT with interesting and useful results. By adding extra lines of instruction or words, specialists can fine-tune generated responses on various topics. Testing with GPT4 can also provide better results than older versions of ChatGPT. For instance, by asking for information or creative ideas like meal ideas and ingredients, ChatGPT can provide responses in tabular form or game ideas and equipment and even produce them in Excel format.
By carefully prompting ChatGPT with follow-up prompts and natural language, the AI can make changes to tables, output text in the style of different authors, and even produce responses in a certain number of words or paragraphs to spot mistakes from a humanized lens.
Open Source Controversy and ChatGPT
OpenAI’s GPT3 language model is exclusively licensed by Microsoft. The company initially released ChatGPT 3.5 as a research preview, but it became very popular after OpenAI made it easier to access. Microsoft invested $10 billion in OpenAI, which is no longer an opensource non-profit company, but an entity focused on making a profit. Elon Musk, one of OpenAI’s co-founders, intended for the company to be an open-source non-profit counterweight to Google.
While OpenAI is legally permitted to use open source licenses, ChatGPT itself is unlikely ever to become open source due to its commercial leaning. Despite these challenges, developers continue to explore ways to enhance artificial intelligence capabilities, and with each new development comes new possibilities.
Image Source: Wikimedia Commons
