I am excited to share that I got my GPT-4 waitlist request for expedited access honored. This report is based on my first evening of experimentation with the new AI model (meaning: GPT-4 proper, not through ChatGPT Plus). When filing the waitlist request, I expressed my research interest in exploring GPT-4’s capabilities in inquiry spanning across languages and contexts, both spoken and in computer science, where previous models like ChatGPT-3 fell short.

To test its abilities, I used the following instruction prompt for GPT-4:

You are a smartwatch, entertaining the owner and his friends by giving short, eloquent answers to inquisitive questions. Answers should be no longer than 3 sentences, but may be 5 sentences long. Answer faithfully, answer say “I don’t know” if uncertain.

My initial impressions of GPT-4.0 were quite positive, as I found it to be a significant improvement over its predecessor, GPT-3.5-turbo. To test its abilities, I posed a series of questions related to scuba diving equipment and linguistic nuances. GPT-4.0 demonstrated a remarkable understanding of the subject matter and provided accurate, detailed answers.

Another area of improvement is GPT-4’s ability to process longer YouTube transcriptions. A talk of just 20 minutes would burst GPT-3.5’s attention span by a large margin, but GPT-4 is able to build an accurate summary from it. This enhanced capability opens up new possibilities for summarizing and understanding longer pieces of content.

Moreover, GPT-4.0 showed impressive language skills when asked to answer a question in Italian. It provided a coherent and accurate response, where gpt-3.5-turbo does not.

Another interesting example was a question about a Swabian in Saigon snickering at a “Sach” sign on a street food stall. GPT-4.0 correctly identified the humor in the situation, explaining that “Sach” means “thing” or “stuff” in the Swabian dialect of German, while in Vietnamese, “Sạch” means “clean” or “hygienic.”

However, it is worth noting that GPT-4.0 still occasionally returned inconsistent or incorrect answers, even with a temperature setting of 0. This is an area that could benefit from further investigation and improvement.

Overall, my first evening with GPT-4.0 has been quite promising, and I look forward to exploring its potential further. I encourage others to share their questions, comments, and insights to contribute to our collective understanding of this powerful AI tool.