GPT-4: Everything about the OpenAI's newly introduced large language model

GPT-4 accepts texts and images input, generates text outputs responding to text and image inputs, and provide descriptions and suggestions for the image

OpenAI on Tuesday announced GPT-4, a large multi-modal language model that is said to scale up efforts in deep learning and provide a more customised human-level interaction on professional and academic benchmarks.

Parallel to the ability to churn out creative essays for users in text-only settings, the new language model accepts texts and images input, generates text outputs responding to text and image inputs, and provides descriptions and suggestions for the image.

OpenAI said that GPT-4 can better handle meticulous instructions and generate more nuanced and reliable answers compared to GPT-3.5. While the two versions can create a similar tone of conversations, “the difference comes out when the complexity of the task reaches a sufficient threshold” the research report released by OpenAI read.

GPT-4 could produce around 25,000 words, about eight times as many as ChatGPT 2. Testing on academic benchmarks, GPT-4 outperformed GPT 3.5. For example, GPT-4 scored in the top 10 percentile while GPT 3.5 was ranked in the bottom 10 percentile for the 2022-23 editions of the simulated Uniform Bar Exam, according to the technical report released by the company.

In comparison with other such models, GPT-4 ranks higher on traditional benchmarks than other existing language models. While the dominant language for machine learning benchmarks still remains to be English, GPT-4 fared well in 24 out of 26 other languages chosen, in contrast with GPT-3.5 and models like Chinchilla, and PaLM.

The users can now steer the user experience in the ChatGPT new version, by prescribing customised system messages to answer in a tone they would prefer, contrary to the fixed verbosity of ChatGPT. While scoring 40 per cent higher on factual evaluation than the antecedent version, the model still is not fully reliable as it  “hallucinates facts and makes reasoning errors”.

Microsoft continues to thrive from its partnership with OpenAI in January and stands to benefit from GPT-4 adoption, an analyst at RBC Capital Markets commented, as reported by Reuters. The new robust model is helping power Bing’s search engine as well, as revealed by Jordi Ribas, Corporate Vice President at Bing.

“Good news, we've increased our turn limits to 15/150. Also confirming that the next-gen model Bing uses in Prometheus is indeed OpenAI's GPT-4 which they just announced today”, tweeted the Microsoft executive.

Ahead of the GPT-4 launch, Google announced its new generative AI lineup, as it released a list of new generative AI capabilities and features for developers through the release of PaLM API and Google Cloud.

“Today marks the next chapter of our AI journey with new tools and experiences across Google Workspace and Google Cloud that allow users, developers and businesses to harness the power of generative AI”, tweeted Thomas Kurian, GoogleCloud CEO, while introducing plus PaLM API and MakerSuite for developers. This is the most recent development since the search engine giant unveiled its chatbot ‘Bard’ in February.

Currently, to have GPT-4 access, users require a paid subscription to the premium version ChatGPT+, and to figure out how to access GPT 4 subscription, users will have to upgrade by selecting the ‘Upgrade to Plus’ button, where the pop-up appears to upgrade plan.