OpenAI has created GPT-4, a powerful new AI GPT model for comprehending images and text, which the company describes as “the latest breakthrough in its efforts in upgrading deep learning.”
GPT-4 is accessible now to OpenAI’s paying customers through ChatGPT Plus (with a use limit), and developers may join up on a waiting to access the API.
YOU MAY ALSO LIKE: ChatGPT passes the MBA exam given by a Wharton professor and here is why we are all concerned
Cost of GPT-4
Unlike its predecessor, OpenAI’s GPT-4 is available at a cost. The cost is $0.03 per every 1,000 “prompt” tokens (equivalent to around 750 words) and $0.06 for every 1,000 “completion” tokens (again, about 750 words). Tokens stand in for unsanitized text; the word “fantastic,” for instance, would be broken down into the tokens “fan,” “tas,” and “tic.” Prompt tokens are the components of words given into GPT-4, whereas completion tokens are the material created by GPT-4.
Companies That Offer GPT-4
The company said that Bing Chat, Microsoft’s chatbot technology created in collaboration with OpenAI, is officially operating on GPT-4 now.
Stripe is another early user that uses GPT-4 to crawl company websites and report findings to their support teams. Duolingo included GPT-4 into their paid membership tier to expand their language-learning product line. Morgan Stanley is developing a system powered by GPT-4 to extract data from corporate papers and provide it to investment bankers. Moreover, Khan Academy is using GPT-4 to develop an automated instructor.
GPT-4 can create text and take a picture and text inputs — an advance over GPT-3.5, its successor, which only acknowledged text — and operates at a “human level” on many academic and professional benchmarks. For instance, GPT-4 scores in the top 10% of test takers on a mock bar exam, while GPT-3.5 lands in the bottom 10%.
The Training of GPT-4
OpenAI claims that after six months of “iteratively aligning” GPT-4 with learning from an internally adversarial testing program and ChatGPT, the system achieved “best-ever outcomes” regarding factuality, steerability, and refusal to step outside of guardrails. Like earlier GPT models, GPT-4 was trained to utilize publicly accessible data, including those from public websites and additional data that OpenAI licensed.
To train GPT-4, OpenAI collaborated with Microsoft to build a cloud-based “supercomputer” on the Azure platform from the ground up.
One of the most intriguing features of GPT-4 is its dual-mode text-and-image comprehension. Captioning and even understanding pictures as complicated as an iPhone with a Lightning Cable converter connected into it is well within GPT-4’s capabilities.
To begin with, only Be My Eyes, one of OpenAI’s partners, will have access to the picture understanding feature for testing purposes. The new Virtual Volunteer function in Be My Eyes is powered by GPT-4 and can respond to queries regarding photographs that users send it.
Perhaps more useful in practice than the steerability as mentioned above toolset is another GPT-4 enhancement. OpenAI is expanding its API to include “system” messages in GPT-4, enabling programmers to provide detailed instructions for a desired aesthetic and functional outcome. Future iterations of ChatGPT will include system messages and effective instructions that define the mood and parameters for the AI’s subsequent interactions.
Improvements In ChatGPT 4
OpenAI acknowledges that it has improved in certain respects; for example, GPT-4 is less likely to reject queries on synthesising harmful substances. According to OpenAI, GPT-4 is 29% more likely to reply appropriately to sensitive queries like those seeking medical advice or information on self-harm and is 82% less likely to respond to requests for “disallowed” material in general compared to OpenAI’s GPT 3.3
GPT-4 has several layers of meaning. OpenAI is charging forward headfirst, seemingly at ease with its introduced improvements.
OpenAI said that they anticipate GPT-4 to become a vital tool that powers numerous apps that help improve people’s lives. There’s still a lot to accomplish, but the community’s efforts to build on top of, explore, and contribute to this model will help make it better in the long run.
YOU MAY ALSO LIKE: Microsoft invests billions to strengthen its partnership with ChatGPT’s maker OpenAI
Continue reading iTMunch for the latest AI news, trends, and stories.
Feature Image Source: redgreystock on Freepik