OpenAI comes back with a bang with the next iteration to GPT-3, which comes shortly after the introduction of ChatGPT earlier this year. OpenAI, ever since its first GPT model, has been setting the trend in terms of Large Language Models, which has had big shots like Alphabet and Meta on their toes, risking their reputation to push architectures to compete with OpenAI.
In November 2022, we talked about how close we are to the introduction of GPT-4 by the organization but shortly deviated with the introduction of ChatGPT and what followed it with Microsoft Investment and integration into their services. But now that it is here, how about visiting that article and seeing how correct our predictions were and what new the architecture brings to the table?
What is GPT-4?
GPT-4 is the next iteration of the well-known Generative Pre-trained Transformer by OpenAI. The model comes from a line of transformers that are known to be the best when it comes to language models. These models can generate text, make conversations, and, even in the case of DALL E 2, make sense of the context of human exchanges and sentences.
GPT-4 is creating such a huge buzz and is, in fact, such a big deal because of the series of generative models we have been seeing both in the sector of image and language generation, which started back last year with DALL E 2, then Stable Diffusion, GATO, ChatGPT, etc. The model itself catches the attention of all the “Tech Bros” in Silicon Valley because of the sheer money big corporations such as Microsoft and Alphabet are ready to invest in this sphere and use case. Back in January, Microsoft invested over $10 Billion in OpenAI to acquire this language model and its variants for their products.
The model has now set up a new precedent for Large Language Models, especially in the category of Autoregressive Transformer Language Models, which will both serve as a competition and headache for researchers worldwide trying to come up with the next revolution in language.
How does it compare to GPT-3.5?
With the great confidence, OpenAI has in its model, its documentation and the official announcement are not even comparing the model to other architectures in the market but its last iteration of the architecture, GPT-3.5. With OpenAI refusing to disclose how big of a parameter count we are looking at when running GPT-4, its difficult to breakdown and compare the models on a technical level, but behold, they have presented a comparison that should be relevant to any VP or Researcher when choosing their next language model.
Comparing the models purely based on exams we humans have to sit for. With the model making a difference in more real exams, we can genuinely see the generative ability of the model. The model can achieve 90 percentiles in the Uniform Bar Exam compared to the ten percentiles conducted by ChatGPT, a model which runs on GPT-3.5. However, we still see the model lacking in exams such as Calculus.
This kind of performance may put the estimated number of parameters of the model north of 500 Billion parameters. When put into wise, this is an astonishingly large yet conservative estimate.
The model comes with a very new and innovative feature that sets it quickly months ahead of the competition with the introduction of “Vision”. GPT is no longer just a language model, accepting over 25,000 words compared to the mere 3,000 ChatGPT can get. The model can even conceive images as its input to generate text.
GPT-4 API
As far as accessing the model goes for individuals like us, it is very readily available with ChatGPT Plus, which is a matter of question if that will be the case going forward. However, for the people who miss waiting for the API like in olden times, you can still sign up for the waitlist and get access in the traditional way.
The model, however, still has some biases in regards to social biases, hallucinations, and adversarial prompts, the API is, according to OpenAI, ready to integrate into users’ pipelines and generate unique and innovative projects.
What are the use cases of GPT-4?
GPT-4, even before its official release, has been given to some of the top clients of OpenAI and Microsoft to test/deploy the model in real-life situations. Some of such releases can be seen in the following:
1. Duolingo: The go-to if you are looking to learn a new language, Duolingo comes powered with GPT-4 now to increase the quality of conversations between the user and the application to make you learn a new language faster.
2. Be My Eyes: The coveted Visual Aid application enabling the challenged with state-of-the-art image-to-text tools comes powered with GPT-4 for better context understanding and text generation which is even out for everyone on IOS to try.
3. Stripe: Stripe is one of the leading payment streamline platforms for business’s around the world requires an astonishing number of conversations and documentation support which is now more efficient with the introduction of GPT-4 in the pipeline.
4. Morgan Stanley: Stanley’s large Wealth Management System has become instantly easily traversable with the GPT-4’s ability to act as an experienced financial advisor and deliver relevant information faster.
5. Khan Academy: The world-class free education provider integrates the power of transformer models with their bot Khanmigo to help students answer their queries and doubts faster.
6. Government of Iceland: Iceland’s integration with the American and European cultures has radically affected the native language speakers of the country to which the Government of Iceland has tasked GPT-4 to help preserve and promote the language.
In all, the model is able to generate text better, hold human conversations better, support image-to-text better, and traverse massive textual datasets better.
Conclusion
We are getting closer to perfecting generalized AI, with a new outstanding model being released every month. It is taking baby steps, which are both costly and difficult. Nonetheless, it makes everyone wonder when the perfect Turing Machine will be developed, and whether it will be developed by OpenAI, DeepMind, or a wildcard lab.
Written By
Aryan Kargwal
Data Evangelist