ChatGPT-4 Explained

OpenAI, the company that take over the world by storm six months ago with the launch of ChatGPT-3 has announced that they have launched the updated version of their chatbot creatively named ChatGPT-4. Here are the things that you should know about the updated ChatGPT-4 and how it could help you (or replace your job).

What is ChatGPT-4

• It’s an updated chatbot: As the name suggests, GPT-4 is the fourth version of the automated chatbot developed by OpenAI, which has attracted investment from celebrities in the industry and most recently, $1 billion from Microsoft.
• It can accept images: The chatbot now can accept images as input. There’s a reddit going around that this person draws an outline of a jokes website and ChatGPT just spits out the code
• It can accept a lot of words as input: You can now put 25,000 words in the input prompt and ChatGPT-4 will analyze and do its magic. In theory, you can put an entire report and ask the bot to analyze it. No guarantees of accuracy though.
• It’s more creative: The researchers put some more creative spin on the model. In their example, the chatbot can explain the plot summary of Cinderella using every letter of the alphabet, in order, without repeating. Quite an impressive feat even for a human.
• Better reasoning: According to the researchers, the chatbot has less chance to fall for logical gotchas, which is one of the highlighted weaknesses of the chatbot. They also included some examples in their System card report.
• Better guardrails: When launching Bing Chat, which Microsoft uses ChatGPT-3, there were reports that reporters asked questions about relationship issues to which ChatGPT suggested the reporter leave his wife. Also stories about how people have misused ChatGPT to increase human biases. OpenAI claims that better guardrails are in place to prevent such situations.
• Better informed: ChatGPT-4 has 175 billion parameters compared to 25 million in ChatGPT-3 which OpenAI claims that the chatbot is now more informed. OpenAI demonstrated that ChatGPT has passed the Bar Exam at the 90th percentile, Medical Board Exam, and also university entrance exams.
• Plug-ins: On March 23, OpenAI announced availablity of plug-ins for ChatGPT-4. Some of the plug-ins that available are OpenTable, Wolfram and Expedia. In their demo, you can find restaurants and book a table using OpenTable plug-ins and ask the Wolfram plug-in to calculate the calaroies of the food that you are going to eat.

How to access

There are several ways to access such power:-

• ChatGPT Plus: A subscription model which starts at $20/month allows you access to ChatGPT-4. This is the easiest and fastest way to access it.
• API: Currently there’s a waitlist to access ChatGPT-4 API libraries. OpenAI stated pricing starts around $0.03/1k token (words) for the prompt (asking) and $0.06/1k token completion (answer).
• BingChat: BingChat, Microsoft’s search engine in chat form actually has been using ChatGPT-4 for the past few months as a hidden testbed. However, there’s a waiting list to actually use the feature.
• GitHub Co-Pilot X: Github is a popular software repository website owned by Microsoft. They have a AI-based help called GitHub Copilot which can help a developer to write out some code. Github is planning to launch Copilot X which takes it to the next level. With a plug-in in VS Code (Microsoft owned IDE), you can put in some prompt of a programming you planning to do, then boom!, code is generated. As of March 2023, it is available as a technical preview but you need to have a GitHub Copilot subscription ($10/month or $100/year).

What people say

Here are what people are saying about ChatGPT-4

• New York Times - Exciting, but scary.
• Wired - Smarter but still have the same flaws
• Arstechnica -

“If AI does turn out to be an existential risk, who will keep humanity safe? With US AI regulations currently just a suggestion (rather than a law) and AI safety research within companies merely voluntary, the answer to that question remains completely open.”

• New York Times - 10 ways it’s impressive but still flawed.

Pitfalls

However, nothing is ever perfect. Especially when you are trained on imperfect data, the model will always end up with biases and inaccuracies from the data.

• Math is better but still bad: In their own test, ChatGPT-4 scored 40-percentile on AP Calculus. Way better than the 0 percentile on ChatGPT-3.5, but don’t count on things.
• Still no novel way of thinking: Leetcode-hard ChatGPT-4 manage to score only 3 out of 40. OpenAI claims contamination, but as one would point out, it is easier to (re)create solutions if you have the world’s data on your hands.

Conclusion

The AI race is heating up over the last few years and the introduction of these very good chatbots is going to heat things up even more. OpenAI is in the limelight right now because of its aggressive release cycle, but that doesn’t mean that other companies like Google are not doing anything about it.

Putting a knowledgeable chatbot on your search engine is a novel way to search for things on the internet, but doing so, in the long run, might backfire. Right now, the internet is supported by ads, content, and ease of shopping. Having a super chatbot is going to change the first two things that you think about the internet.

Still, as the CEO of Microsoft would put it, GPT-4 will be remembered for making Google dance.

Resources

• Whitepaper - Technical Report by OpenAI.
• System Card - “Report Card” by OpenAI which demonstrates the guardrails that OpenAi claimed has put on the text generator.

Plug

Support this free website by visiting my Amazon affiliate links. Any purchase you make will give me a cut without any extra cost to you

Other Ecosystem Items

• Apple TV (Amazon) / Apple TV 4K (Amazon)
• AirTags (Amazon)
• HomePod Mini (Amazon)