OpenAI released GPT-4 on March 14, 2023, and it’s already turning heads. Not only can it chat, hear, and speak, but it’s also showing the first hints of AGI and nuanced emotions. If you’re in a high-skilled job, you might want to pay attention.
This article has the most up-to-date GPT-4 statistics, facts, and everything you need to know about Open AI’s latest and greatest large language model. Let’s dive in.
I have also made a series of short videos discussing Microsoft paper on GPT-4 showing sparks of AGI and Open AI study on GPT capabilities and jobs most at risk.
What is GPT-4?
GPT-4 is the next version after GPT 3.5, which powers the free version of ChatGPT. Unlike its predecessor that only understands and responds in text, GPT-4 brings new capabilities like multimodality, video processing, and quicker responses.
Multimodality means the model can understand multiple types of input, such as text, images, and audio. It’s like being multilingual, but for different forms of data.
In simpler terms, GPT-4 can interpret images, text, and even audio, thanks to a recent update that added voice control in its mobile app. Additionally, it can respond using both text and images when paired with plugins.
How was GPT-4 trained?
GPT-4 was trained using web corpus data, which is a fancy way of saying it learned from a vast collection of text available on the internet. To fine-tune its abilities, it underwent a process known as Reinforcement Learning with Human Feedback, or RLHF. This means the GPT-4 model was improved by learning from both automated data and real human interactions.
Essentially, RLHF is like a two-step quality check: first, the machine learns on its own, and then humans step in to guide it towards better answers.
There were whispers that GPT-4 boasted an enormous 100 trillion parameters, dwarfing GPT-3’s 175 billion. However, OpenAI CEO Sam Altman debunked this myth, clarifying that the actual size is still under wraps. There are also new speculations about GPT-4 having around 220 billion parameters (compared to Google Palm 2’s 540 Billion parameter size!).
How to access GPT-4? (Free and paid methods)
- GPT-4 API can be accessed through an OpenAI account. It comes in several variants like GPT-4 Turbo with 128k token context, gpt-4-32k, and gpt-4-vision.
- GPT-4 is also available in ChatGPT plus (paid version of ChatGPT).
- Bing chat provides a free version of GPT-4 on its mobile app, which can be toggled on and off.
GPT-4 Models Available via API
| Model | Description | Context Window | Training Data |
|---|---|---|---|
| gpt-4-1106-preview (new) | GPT-4 Turbo. The latest model with improved instruction following, JSON mode, reproducible outputs, parallel function calling, and more. Not yet suited for production. | 128,000 tokens | Up to Apr 2023 |
| gpt-4-vision-preview (new) | GPT-4 Turbo with vision. Understands images, plus all other GPT-4 Turbo capabilities. Preview model not yet suited for production. | 128,000 tokens | Up to Apr 2023 |
| gpt-4 | Classic GPT-4. | 8,192 tokens | Up to Sep 2021 |
| gpt-4-32k | Classic GPT-4 with larger context. Snapshot of gpt-4-32k from June 13th 2023 with improved function calling support. | 32,768 tokens | Up to Sep 2021 |
ChatGPT Plus Subscribers (GPT-4 User Numbers)
- ChatGPT has 180.5 million active users, but only an estimated 1% subscribe to ChatGPT Plus plan which puts the number of ChatGPT Plus users at around 180,000-200,000. All ChatGPT plus users get access to GPT-4 at $20/month.
- Others can sign up for the GPT-4 waitlist to get early access to its API.
- Bing chat also offers GPT4 on mobile app.
What are GPT-4’s Capabilities?
GPT-4 is the most advanced AI model released by OpenAI till date. You can read the full technical report on GPT-4 here. Let’s dive into what sets GPT-4 apart from its predecessors, like GPT-3.5.
GPT-4 vs GPT-3: Enhanced Precision and Accuracy
- Precision and Relevance: GPT-4 provides more precise and relevant answers. A study found that GPT-4 has an accuracy of 85.5% in a 3-shot massive multitask language understanding (MMLU) test, compared to GPT-3.5’s 70.1% (source)
- Image Recognition: GPT-4 can see and describe images in great detail, a feature not present in GPT-3. For example, show it a photo of your fridge’s contents, and it can suggest recipes you can make.
- Vision and Voice: ChatGPT Plus allows users to talk to GPT-4 via voice control on its mobile app and to explain things to GPT by drawing.
- Model-Size & Parameters: Contrary to rumors, GPT-4 doesn’t have 1 trillion parameters as debunked by Sam Altman, Open AI CEO. The actual number could be closer to 220 billion.
- Mathematical Skills: GPT-4 is better at solving mathematical problems.
- Language Skills: GPT-4 outperforms GPT-3.5 in 24 out of 26 languages tested, including low-resource languages like Latvian, Welsh, and Swahili. It also excels in text editing, paraphrasing, and fixing grammatical errors.
- Humor: While GPT-4 is better at generating jokes, it still doesn’t fully understand humor.
- Better Safety Measures: GPT-4 is much better than GPT-3.5 at avoiding inappropriate, disallowed, and sensitive requests. OpenAI has decreased the model’s tendency to respond to disallowed content by 82% and improved its response to sensitive requests by 29% more often.
- Professional Tests: GPT-4 performs better in professional tests and exams than GPT 3.5.
GPT-4 Test and Exam Performance
GPT-4 has shown to excel in professional tests and exams. For instance, it scored in the 90th percentile in the Uniform Bar Exam and 99th percentile in the GRE Verbal section. Here’s a quick comparison table of performance of various GPT models at giving exams.
| Simulated exams | GPT-4 estimated percentile | GPT-4 (no vision) estimated percentile | GPT-3.5 estimated percentile |
|---|---|---|---|
| Uniform Bar Exam (MBE+MEE+MPT) | 298 / 400~90th | 298 / 400~90th | 213 / 400~10th |
| LSAT | 163~88th | 161~83rd | 149~40th |
| SAT Evidence-Based Reading & Writing | 710 / 800~93rd | 710 / 800~93rd | 670 / 800~87th |
| SAT Math | 700 / 800~89th | 690 / 800~89th | 590 / 800~70th |
| Graduate Record Examination (GRE) Quantitative | 163 / 170~80th | 157 / 170~62nd | 147 / 170~25th |
| Graduate Record Examination (GRE) Verbal | 169 / 170~99th | 165 / 170~96th | 154 / 170~63rd |
| Graduate Record Examination (GRE) Writing | 4 / 6~54th | 4 / 6~54th | 4 / 6~54th |
How is ChatGPT Plus (GPT4) different from ChatGPT Free?
“The difference (in performance) comes out when the complexity of the task reaches a sufficient threshold — GPT-4 is more reliable, creative and able to handle much more nuanced instructions than GPT-3.5″, says OpenAI.
| ChatGPT Free (GPT 3.5) | GPT-4 | |
|---|---|---|
| Input | Only text | Multimodal, accepting both text and images (requires mobile app) |
| Output | Only text | Only text |
| Model Size | 175 billion parameters, GPT 3.5 | Unknown, likely around 220 billion (Rumored to have 100 trillion parameters but debunked) |
| Availability | Free preview, Paid API | ChatGPT Plus, API waitlist, Bing chat |
| Applications | Language tasks like text summarization, text generation, writing and debugging code, etc. | Language tasks like text summarization, text generation, writing and debugging code, etc; Analyzing images and drawing conclusions from visual data. |
| Superiority | Inferior to GPT-4/ChatGPT plus | Better at avoiding inappropriate requests (safer), less likely to hallucinate, superior at exams |
I took GPT-4 around for a spin in my comprehensive ChatGPT plus review . In a nutshell, GPT-4 is leaps ahead of GPT 3.5 in almost every task requiring multi-step thinking.
Gpt-4 API models vs GPT 3.5 API
| Aspect | Best GPT-4 Models | Best GPT-3.5 Model |
|---|---|---|
| Model Name | gpt-4-1106-preview & gpt-4-vision-preview | gpt-3.5-turbo-1106 |
| Description | gpt-4-1106-preview: Latest GPT-4 with advanced features like 128k context, data analysis, etc. gpt-4-vision-preview: GPT-4 Turbo with image understanding. | Updated GPT-3.5 Turbo model with improved capabilities. |
| Key Features | – Improved instruction following – JSON mode – Reproducible outputs – Parallel function calling | – Cheaper and faster. |
OpenAI Dev Day for GPT-4
OpenAI announced several features related to GPT-4 on OpenAI dev day:
- The Assistants API allows developers to create agent-like experiences in their applications, using generative AI for various tasks including coding and data analysis. These are called GPTs.
- GPT-4 Turbo features an expanded128k token context window and has capabilities for image input and human-quality speech output, with different versions for text-only and text-plus-image understanding.
- Custom GPTs can be created through conversational programming, making it accessible for individuals without coding experience.
- The GPT Store is a forthcoming platform where creators can list their GPT applications, with a revenue-sharing model.
- Copyright Shield aims to protect enterprise and API customers of ChatGPT from copyright lawsuits.
- The new interface for ChatGPT offers a simpler design with a dark theme and functionality to switch between ChatGPT and DALL-E 3.
- ChatGPT’s updated information now includes data up to April 2023 and the ability to search through PDFs and other documents.
- Startup Mentor is a demonstration program that can create a GPT to advise startup founders, showcasing the ease of creating specialized GPTs.
What are limits of GPT-4?
- Context limits: With API access, you get to choose between two models – GPT-4 8k token model and GPT-4 32k token model. This means GPT-4 can process up to roughly 25,000 words of text.
- Message limits: GPT-4 messages are limited to 50 messages every 3 hours on ChatGPT Plus.
- Knowledge limits: GPT-4 has access to all the latest internet data because of Bing integration and plugins. However, it can still hallucinate.
Apps Using GPT-4
- Sudowrite is the best AI story generator on the market, which uses fine-tuned GPT-4 model to brainstorm characters, expand plots, give AI feedback to writers, etc.
- ChatGPT Plus uses GPT-4 to answer user queries.
- Duolingo: GPT-4 helps create real-life conversation scenarios for language learners, making practice more engaging and relevant.
- Be My Eyes: GPT-4 describes objects and reads text aloud to assist visually impaired users in their daily tasks.
- Stripe: GPT-4 handles customer service queries and detects fraudulent activities by analyzing user behavior.
- Morgan Stanley: In wealth management, GPT-4 analyzes financial data to provide insights like market trends and investment strategies.
- Khan Academy: GPT-4 generates educational content and interactive learning modules, including quiz questions and topic explanations.
- Government of Iceland: GPT-4 translates documents into Icelandic and generates content to help preserve the language.
People’s Reaction to GPT-4
- Despite the rumors about 1 trillion parameters of ChatGPT-4, OpenAI has chosen not to reveal the actual size of GPT-4.
- Experts are disappointed by the closed nature of OpenAI’s GPT-4, saying it is slowing down research.
- Read our posts on impact of GPT-4 in education and its impact on business.
- Google has launched Gemini line of AI models as an answer to GPT-4.
FAQs
Does GPT-4 Hallucinate and Invent Facts?
Yes, GPT-4 does hallucinate and invent facts. Despite scoring 40% higher than GPT-3.5 on internal adversarial factuality evaluations, OpenAI has explicitly warned users that while GPT-4 has made improvements over its predecessor GPT-3.5, it can still “hallucinate” facts and make reasoning errors. For example, once GPT-4 invented references, DOI links, and/or facts without any identifiable link to a real article.
Is GPT-4 AGI?
No, GPT-4 is not fully an Artificial General Intelligence (AGI), but it’s getting closer. According to a paper by Microsoft Research titled “Sparks of Artificial General Intelligence: Early experiments with GPT-4”, GPT-4 exhibits more general intelligence than previous AI models. And it can solve a wide range of tasks that span multiple domains such as mathematics, coding, vision, medicine, law, and psychology without needing any special prompting. Its performance is often close to human-level and vastly surpasses prior models like ChatGPT 3.5.
Which Jobs Could be Replaced by GPT?
According to a study by OpenAI, jobs that require higher education credentials like engineers, writers, and accountants are more exposed to GPT while physical jobs like carpentry and plumbing are less affected. The study reveals that about 80% of the U.S. workforce could have at least 10% of their work tasks affected by GPTs. Moreover, around 19% of workers may see at least 50% of their tasks affected.
How Much Data Did GPT-4 Train On?
The exact amount of data GPT-4 was trained on is not publicly disclosed. However, it’s speculated to be significantly larger than GPT-3, which was trained on 570GB of text. GPT-4 is rumored to have 1.76 trillion parameters, indicating a massive dataset. Source: Wikipedia
How Good is GPT-4 at Math?
GPT-4’s mathematical capabilities are enhanced through plugins like Wolfram Alpha. This allows it to solve complex equations and perform advanced calculations, making it more reliable for mathematical tasks than its predecessors.
GPT-4 Parameters – Is it 100 Trillion?
Contrary to some rumors, GPT-4 does not have 100 trillion parameters. Sam Altman has debunked this, stating that the true size is a secret. However, some sources suggest it might be closer to 220 billion parameters.
Does GPT-4 Have 2023 Data?
While GPT-4’s training data only goes up until January 2022, its plugins can search the internet, providing it with access to the latest information, including data from 2023.
What Is OpenAI Evals?
OpenAI Evals is a framework designed to evaluate the safety and performance of language models like GPT-4. It ensures that the model is both effective and safe for various applications.
Final Thoughts
GPT-4 is a huge leap forward for natural language understanding and generation, as well as for content creation and consumption. GPT-4 could revolutionize your search experience by allowing you to see videos based on your queries, instead of just text or images.
Given the right plugins, it could also enable you to create your own videos from text descriptions, such as scripts, summaries, or captions. Imagine being able to watch a movie based on your favorite book, or a video tutorial based on your instructions.
GPT-4 is not just a tool for entertainment, but also for education, research, and business. It could help you learn new skills, discover new information, and communicate better with others. It could also help you generate high-quality content for your blog, website, or social media.
We hope you enjoyed this blog post about GPT-4 and its amazing capabilities. If you want to learn more about GPT-4 and how to use it, check out my how-to article on using GPT for validating business ideas. Thank you for reading and have a wonderful day! 🙂
- 5 Inspiring No-Code Startup Examples w/ Tech Stack - January 26, 2024
- AI Sleeper Agents: Latest Danger to AI Safety (Anthropic Research) - January 20, 2024
- LOVO AI Review: Complete Guide & Alternatives [2024] - December 11, 2023