Written by
Steve Roberts
Article
•
3 min
OpenAI's Generative Pre-trained Transformer 4 (GPT-4) is an artificial intelligence model that has achieved new benchmarks in natural language processing capabilities compared to previous models like GPT-3. In this blog, we’ll explore GPT-4 and how it aims to improve upon previous GPT models.
What Is GPT-4?
GPT-4 is the next-generation language model created by OpenAI's Alignment Research Center. It’s the largest AI model created so far by OpenAI, building upon the capabilities of the company's previous models like GPT-3. GPT-4 offers:
Significantly larger model size compared to previous GPT versions—up to 100 trillion parameters
Ability to ingest and generate multiple data modalities beyond just text, including capabilities like image and video
Improved reasoning, logic, and common sense capabilities compared to prior models
State-of-the-art performance on academic benchmarks for natural language and multimodal tasks. GPT-4 achieves human-level performance on the Uniform Bar Examination, LSAT, and SAT.
Training on even larger datasets of text, images, videos, and other data for broader world knowledge
Essentially, GPT-4 aims to achieve new breakthroughs in deep learning model performance on a variety of AI capabilities. While GPT-3 impressed with its natural language prowess, GPT-4 expands into additional modalities like image inputs while also continuing to enhance text understanding.
How GPT-4 Builds Upon Previous GPT Models
To understand GPT-4, it helps to first look at the previous GPT models developed by OpenAI and how each iteration has expanded capabilities.
GPT-1—The Beginning
This first Generative Pre-trained Transformer model was released in 2018 and had 110 million parameters. It achieved state-of-the-art performance on language modeling benchmarks at the time.
GPT-2—The Sequel
This 2019 upgrade dramatically increased model size to 1.5 billion parameters. It performed even better at language tasks like text generation and machine translation.
GPT-3—The Leap
Considered a huge leap forward in 2020, GPT-3 grew to 175 billion parameters. It reached new benchmarks in few-shot learning for natural language and could perform a wide range of language tasks at human-level performance or better with little training data. This was also the first model most consumers became familiar with during the so-called, “AI boom.” While the max request value is 2,049 tokens and on the GPT-3.5, this limit is 4,096 tokens or about three pages of single-lined English text.
GPT-4—The Innovation
Now GPT-4 aims to continue pushing these capabilities even further. Training it on huge datasets ranging from text corpora to image collections, the model absorbs vast general knowledge about language and visual concepts. Leveraging this broad knowledge, GPT-4 achieves stronger reasoning, logic, and common sense. GPT-4-8K offers 8,192 tokens and the variant GPT-4-32K can provide up to 32,768 tokens, which is about 50 pages of text.
GPT-4 has also improved upon GPT-3's correctness. GPT-4 scores 40% higher on factual performance than GPT-3 and produces fewer "hallucinations" or errors.
The Path Forward After GPT-4
GPT-4 aims to be OpenAI's most capable AI model to date, but it is unlikely to be the final milestone. The arc of progress in deep learning and massive models has been swift, with companies like Google and Meta also pursuing huge models. It’s likely that GPT 5, 6, or beyond will follow in the coming years as computing power grows.
There is still significant research required to overcome challenges like bias in training data, risks, and improving alignment with human values. Initiatives like OpenAI's alignment research division are doing important work here. And just like the models that came before, GPT-4 lacks knowledge of events that occurred after September 2021. It can still generate inaccurate advice, bad code, or harmful information, and because of those risks, it should be used with careful judgment.
As models continue to grow in size and capability, integrating more data and modalities, they will unlock new frontiers in replicating and advancing natural and artificial intelligence. GPT-4 represents the next step forward, but not the finish line.