A First Look at GPT-4

This model is “something else” 🤯

Heiko Hotz
3 min readMar 14, 2023
Image by author

What is this about?

This is blog post is about the first few tests I ran with the new GPT-4 model. Spoiler alert: It’s impressive!

Chain of thought reasoning

Let’s start with the title image. In this case we ask GPT-4 a very convoluted question that requires multiple chains of reasoning. Not only does it get the answer right, it also provides its reasoning step by step!

Logical fallacy

In the next one I try to trick GPT-4, and I already thought “Hah, gotcha!” when it non-chalantly added a sneaky footnote into the response 😂

Image by author

Logic / math — part 1

This type of question ChatGPT got wrong many times, but GPT-4 is very unimpressed by my attempt to trick it:

Image by author

Logic / math — part 2

Image by author

Logic / math — part 3 (with bonus points for the movie question)

This one blew my mind!

Image by author

AWS exam question — part 1

That’s correct!

Image by author

AWS exam question — part 2

Also … correct … Wait, am I out of a job now???

Image by author

Creating a sample architecture for a photo-sharing website

Yup, I’m definitely out of a job …

Image by author

Coding

I wonder if I can just let GPT-4 do my job and not tell anyone … 🤪 Can anyone confirm if this is correct, by the way?

Image by author

NLP

Of course it gets that one right 🙃

Image by author
Image by author

Current events

It still has the knowledge cutoff from September 2021, tho. But the geography part is correct, I believe (geo was never my strong suit …)?

Image by author

Conclusion

This model is something else, I believe. And we haven’t even explored the multimodal aspects of it. I will obviously stay up all night to experiment more and might update this blog post from time to time 😊

BECOME a WRITER at MLearning.ai

--

--

Heiko Hotz

Generative AI Blackbelt @ Google — All opinions are my own