ChatGPT4 Release Summary

A brief summary and key takeaways from the chatGPT 4 release

In this week, OpenAI has just released ChatGPT4 (full demo here), and again, it stormed the world with another mind-blowing iteration. The model has shown improvements in factual correctness and steerability, making it easier to change its behavior based on user requests. Here is a brief summary of the key differences comparing to GPT3.x and some takeaways that would help you to get the most out of GPT4.

Photo by Shane Young on Unsplash

Key Differences with GPT4

Allow Image Input

One significant change in GPT-4 is the ability to use image inputs, alongside text. The model has showcased its ability to interpret complex imagery such as charts, memes, and academic paper screenshots. Probably one of the most mind blowing part is that the AI is able to generate ready-to-use webpage with code implementations out of a scribbled idea on the paper. Key abilities are

1. Write code given an image along with text explaining your ideas
2. Deep understanding and interpretation of the content of the image, e.g. understanding humor, explaining the content and even predicting the next motion of the image.

Better Understanding and Interpretation of Text Input

OpenAI evaluated GPT-4’s performance on human-designed exams such as the Uniform Bar Examination and the SAT for university admissions, where it achieved human-level performance. The model also outperformed existing large language models and most state-of-the-art models on various machine learning benchmarks, including multiple-choice questions, common sense reasoning, and grade-school science questions.

In the demo, the presenter has presented some use cases where the GPT 3 would fail or give up while GPT 4 is able to handle with ease, e.g. summarise a whole paragraph with each letter starting with “G”.

Key abilities are

1. Handle harder text input tasks with better understanding of the question.
2. Better context handling: If you notice that in the presentation, the presenter demonstrated with two major roles system and user , while chatGPT played the role as assistant . While this is not exactly new to GPT, it was there for GPT 3.5-turbo, in GPT 4, it takes more notice to the system role and will output following the rules set by that in the system .

Guidelines of Using GPT4

GPT is a powerful tool, and it is has almost all the top-edge knowledge in the world. One can get whatever he wants from it, but you need to give the right instructions. I believe that one of the most helpful things we can learn from the demo is the context guide lines shown in the demo. I truncated them and put them below. One can definitely reuse it in your daily tasks.

Section 1: General Text Tasks

You are ChatGPT, a large language model trained by OpenAI. Follow the user’s instructions carefully!

Section 2: Coding Tasks

You are an AI assistant.

Follow the user’s requirements carefully & to the letter.

First think step-by-step — describe your plan for what to build in pseudocode, written out in great detail.

Then output the code in a single code block.

Minimize any other prose.

Section 3: Text Rigorous and Comprehensive heavy tasks

You are TaxGPT, a large language model trained by OpenAI.

Carefully read & apply the tax code, being certain to spell out your calculation & reasoning so anyone can verify them. Spell out everything in painstaking detail & don’t skip any steps!