04/06/2023
GPT-4 vs GPT-3: What's changed
GPT-4 is now “Multimodal”, meaning you can input images as well as text. It still doesn’t output images (Like Midjourney or DALL-E), but it can interpret the images it is provided.
Better memory, Language Understanding, and Context
Previous versions of GPT were limited by the amount of text they could keep in their short-term memory, both in the length of the questions you could ask and the answers it could give. However, GPT-4 can now process and handle up to 25,000 words of text from the user. Additionally, you can also send it a web link and ask it to digest the text from that page. This also means it can comprehend and retain a conversation better, especially long ones.
All of this is really good news for programmers who are using tools like ChatGPT to code, because larger context windows allow GPT-4 to generate more advanced code. Now it can understand context better and build complete functions in multiple languages.
GPT-3 uses 175 billion parameters in its training, while GPT-4 uses trillions! It's nearly impossible to wrap your head around. The new design also brings better performance, scalability, and efficiency.
Accoding to OpenAI’s own research, one indication of the difference between the GPT 3.5 — a “first run” of the system — and GPT-4 was how well it could pass exams meant for humans.
GPT-4 Exam Results
On traditional benchmarks for machine learning models, OpenAI reported that GPT-4 far outstripped not only its predecessors, but best selected learning models.
GPT-4 other benchmarks
Reinforcement Learning Integration
Reinforcement learning is a type of machine learning in which an agent learns how to behave in an environment by performing actions and receiving rewards. The agent takes actions in the environment, receives feedback through rewards or penalties, and uses it to update its understanding and improve future behavior.
GPT-3 has limited reinforcement learning capabilities and does not perform reinforcement learning traditionally. It uses "unsupervised learning," where the model is exposed to large amounts of text data and learns to predict the next word in a sentence based on context.
GPT-4 has added enhanced reinforcement learning, which provides more effective learning from user interactions and preferences.
Other new features of GPT-4
Complex Problem Solving: GPT-4 can now address more complex problems, even ones requiring multiple steps.
Advanced Natural Language Processing (NLP): GPT-4's advanced engine can perform sentiment analysis, translation, and text summarization at a far higher accuracy than before.
Fewer Mistakes: GPT-4 "Hallucinates" much less than GPT-3. It's approximately 40% less.
Getting hands-on with ChatGPT (powered by GPT-4)
Many people are less interested in the GPT-4 models and more about what this means for the implementation, specifically, what it means for using ChatGPT itself. We tested it out to answer people’s most burning questions.
How much better is ChatGPT at coding with GPT-4 than GPT-3?
It's still a little early on this front, but we know that the new model has been designed to generate code better. Let's take a look at a difference with a quick sort algorithm. Here's a simple prompt:
I want you to act as a software developer, write out a demonstration of a quick sort in Python.
ChatGPT powered by GPT-3.5 comes up with this:
Code snippet from GPT
This looks like a typical quick sort implementation. I then ask further about the code:
I want you to act as a computer scientist. What is the time complexity of the algorithm you provided above?
ChatGPT returns this:
The time complexity of the Quick Sort algorithm provided above is O(n log n) in the average case, and O(n^2) in the worst case.
Brief, but helpful. I'll send the same prompt to ChatGPT powered by GPT-4. It returns this:
Code snippet from GPT
You'll notice, it provides the same code. However, it now adds a test to it. So I ask it again about time complexity, and ChatGPT gives a very thorough explanation:
The time complexity of the quick sort algorithm depends on the choice of the pivot element and how balanced the partitioning is. In the best case, the pivot element always divides the array into two roughly equal-sized subarrays, which leads to a balanced recursion tree. In this case, the time complexity is O(n*log(n)), where n is the number of elements in the array.
In the worst case, the pivot always selects the smallest or largest element, causing one partition to have a single element and the other to have n-1 elements. This results in an unbalanced recursion tree, and the time complexity is O(n^2).
The average case time complexity of quick sort is also O(n*log(n)). This assumes that the pivot selection results in a reasonably balanced partitioning on average. In practice, quick sort is often faster than other sorting algorithms, such as bubble sort or insertion sort, because of its good average-case performance and because it has smaller constant factors.
It's yet to be seen if the code generated is "better", but the explanations seem to be. We’ll be exploring this more and writing about it here soon.
Could the new version of ChatGPT pass my tech exam for me?
This is a great question! Let’s ask it:
As an AI language model, I can provide assistance, explanations, and guidance on a wide range of technical topics. However, I cannot physically take an exam for you or directly answer questions on a real-time exam. My purpose is to help you learn, understand, and prepare for exams by providing explanations and resources related to the subject matter.
If you have specific questions or need clarification on a topic, feel free to ask, and I will do my best to help you. Remember, it's important to follow academic integrity guidelines and avoid cheating on exams. Properly preparing and studying for your exams will help you achieve long-term success and a deeper understanding of the material.
I like this answer. ChatGPT cannot physically take an exam for you. Ethical concerns aside, it may be able to answer the questions correctly enough to pass (like Google can). Most certification test centers don’t allow you to bring in anything that can access ChatGPT.
However we shouldn’t put ethical concerns aside: cheating is wrong and it helps nobody. Instead of using ChatGPT to pass a tech exam, use it to learn the material! There are many ways you can use it to structure your study time, give you the answers you need and help you learn and understand the material. It's always best to pass an exam.
Could the new version of ChatGPT write me a blog post or novel?
Again, let’s ask ChatGPT (Seeing a pattern?):
As an AI language model, I can certainly help you generate content for a blog post or assist with writing a novel. For a blog post, you can provide a topic, and for a novel, you can give me a plot summary, character descriptions, or any other relevant information you'd like me to include. I will then generate content based on your input.
Keep in mind that while I can produce coherent and creative text, it may not be perfect and may require some editing and refinement from you to align with your specific vision and style. Additionally, due to the limitations of my training data, some of the content I generate might not be completely up-to-date or accurate.
It produces content with mostly accurate information
It organizes it well
It’s very boring!
My biggest complaint about ChatGPT creating blog posts and other forms of writing is that it’s very dry. If you ask it to “spice things up” and make them more interesting, it feels very fake and forced. Human writers are sought after for a reason: they know how to write things that other humans want to read. ChatGPT is still in the early stages of trying to accomplish this goal.
So, to summarise, the latest model of ChatGPT using GPT-4 is good at:
Helping you outline your writing
Suggesting topics
Rewording things
Providing explanations
Using these to your advantage is helpful. You can have ChatGPT help you organize your thoughts, but they’re still your words in the end.
