Unique Features in GPT-4 in a Nutshell
GPT-4 is now released, and its new features put the GPT-3 or 3.5 to shame even those they were extremely advanced already. OpenAI hosted a developer demo on YouTube on March 14th, showing new features and various tasks that GPT-4 can do better than the previous versions. GPT-4 API has not been released yet, but you can join the waitlist.
Let's dive in to a short list of what GPT-4 can do or do better.
OpenAI describes GPT-4 as being capable at "human-level performance" on various tasks, and they are not lying. Compared to GPT-3, GPT-4 scores far better in standardized tests in various subject matters.
Using the newly released Chat Playground, OpenAI demonstrated the difference in verbal and logical reasoning capabilities of GPT-3 and GPT-4.
When a long article was provided and GPT was asked to summarize the article using only the words starting with G, GPT-3.5 Turbo (the model that is powering ChatGPT) simply summarized the article without understanding the constraint, while GPT-4 was able to summarize it using only words starting with G. The word "AI" doesn't start with "G" but when the demonstrator told GPT-4 of the mistake, it did give a correct answer.
This is perhaps the biggest improvement over GPT-3.5 - that GPT-4 is a multimodal model, not just a language model. GPT-4 can take in prompts composed of both text and images, and this opens a new door to how humans and AI can interact. For example, if you don't know why an image is funny, you can ask GPT-4 its opinion.
Some practical tasks that can be accomplished include
In contrast to ChatGPT which can take in 3,000 words as a context, GPT-4 can take in up to 25,000 words. What this allows is that GPT-4 can analyze a long body of text and do problem-solving on it. For example, the tax code is notorious in its length and complexity. No one wants to spend hours looking at the tax code, trying to figure out what's going on. When you feed the tax code to the GPT-4 and ask a question, it can answer your question to a reasonable accuracy because of the ability to take in extended body of text.
One of the problem with GPT-3 was that from time to time, it gave insensitive or offensive responses. This presented a potential liability problem for businesses that wanted to use GPT-generated answers without human review. To mitigate the risk, businesses had to implement custom filters to filter out offensive content, put a human reviewer to review the output before showing to end-users, or put a disclaimer on their website or app.
According to OpenAI, GPT-4 became safer and less risky for businesses to use as it has less likelihood of producing inappropriate responses. Compared to GPT-3.5, GPT-4 is 82% less likely to respond to inappropriate request from end-users and 40% more likely to produce factual answers. While no AI system is perfect, this improved safety standard should be a boon for businesses utilizing GPT-4 to communicate with their customers.
GPT-4 API is not available to developers, yet. That means you cannot create applications using GPT-4 as of today. But, knowing how fast OpenAI is rolling out new features, it shouldn't be too long before it is made available. Momentech has a team of capable developers experienced in GPT technology and other AI solutions. Let us know your idea and we'll see how to make it happen.
Let's dive in to a short list of what GPT-4 can do or do better.
GPT-4 has a far better reasoning capability than GPT-3.5
OpenAI describes GPT-4 as being capable at "human-level performance" on various tasks, and they are not lying. Compared to GPT-3, GPT-4 scores far better in standardized tests in various subject matters.
- Uniform Bar Exam - GPT-3: 10th percentile, GPT-4: 90th percentile
- LSAT - GPT-3: 40th percentile, GPT-4: 88th percentile
- SAT Math: GPT-3: 70th percentile, GPT-4: 89th percentile
- SAT Evidence-Based Reading & Writing: GPT-3: 87th percentile, GPT-4: 93rd percentile
Using the newly released Chat Playground, OpenAI demonstrated the difference in verbal and logical reasoning capabilities of GPT-3 and GPT-4.
(Image credit: OpenAI)
When a long article was provided and GPT was asked to summarize the article using only the words starting with G, GPT-3.5 Turbo (the model that is powering ChatGPT) simply summarized the article without understanding the constraint, while GPT-4 was able to summarize it using only words starting with G. The word "AI" doesn't start with "G" but when the demonstrator told GPT-4 of the mistake, it did give a correct answer.
GPT-4 can take in both text and image
This is perhaps the biggest improvement over GPT-3.5 - that GPT-4 is a multimodal model, not just a language model. GPT-4 can take in prompts composed of both text and images, and this opens a new door to how humans and AI can interact. For example, if you don't know why an image is funny, you can ask GPT-4 its opinion.
(Image credit: OpenAI)
Some practical tasks that can be accomplished include
- Extracting information from a picture and processing the information subsequently
- Describing an image or look for specific data for visually impaired people
- Learning assistant for young children who have not attained mastery of reading, yet
- Analyzing information contained in an image and reporting the outcome
- Problem-solving tasks involving both images and text such as providing possible diagnosis by looking at a patient's CT scan and medical history
GPT-4 can take in up to 25,000 words
In contrast to ChatGPT which can take in 3,000 words as a context, GPT-4 can take in up to 25,000 words. What this allows is that GPT-4 can analyze a long body of text and do problem-solving on it. For example, the tax code is notorious in its length and complexity. No one wants to spend hours looking at the tax code, trying to figure out what's going on. When you feed the tax code to the GPT-4 and ask a question, it can answer your question to a reasonable accuracy because of the ability to take in extended body of text.
(Image credit: OpenAI)
Less inappropriate responses by GPT-4
One of the problem with GPT-3 was that from time to time, it gave insensitive or offensive responses. This presented a potential liability problem for businesses that wanted to use GPT-generated answers without human review. To mitigate the risk, businesses had to implement custom filters to filter out offensive content, put a human reviewer to review the output before showing to end-users, or put a disclaimer on their website or app.
According to OpenAI, GPT-4 became safer and less risky for businesses to use as it has less likelihood of producing inappropriate responses. Compared to GPT-3.5, GPT-4 is 82% less likely to respond to inappropriate request from end-users and 40% more likely to produce factual answers. While no AI system is perfect, this improved safety standard should be a boon for businesses utilizing GPT-4 to communicate with their customers.
Want to implement GPT-4 to your business?
GPT-4 API is not available to developers, yet. That means you cannot create applications using GPT-4 as of today. But, knowing how fast OpenAI is rolling out new features, it shouldn't be too long before it is made available. Momentech has a team of capable developers experienced in GPT technology and other AI solutions. Let us know your idea and we'll see how to make it happen.
