Introducing GPT-4: The next generation AI that can ‘see’ – National

Just four months after ChatGPT’s release, OpenAI has announced its next-generation artificial intelligence (AI) model. GPT-4 not only beats ChatGPT on various standardized tests like the SAT and bar exam, but it also adds significant new features.

GPT-4 stands for “Large multimodal modelThis means that you can analyze images as well as text. The exciting addition of “computer vision” will allow AI users to input photographs and drawings, which the model can then analyze and understand at first glance.

and Promotional video published by OpenAIthe company was given a picture of a helium balloon tied to an anvil, with questions such as “What happens if the string breaks?” — GPT-4 can logically guess that the balloon will fly away.

GPT-4 can’t generate images (OpenAI’s DALL-E has it covered), but its computer vision applications are amazing.

In a live demonstration of AI’s capabilities, OpenAI president and co-founder Greg Brockman showed that AI can create entire websites based solely on handwritten notes.

GPT-4 can generate and ingest up to 25,000 words, an 8x improvement over ChatGPT. The company says it will help them “compose songs, write scripts, and learn how to write for users.”

in a research paper OpenAI was released with the launch of GPT-4 and shared GPT-4 scores on various academic tests such as the AP exam, bar, Graduate Record Examination (GRE), and even the Sommelier Certification exam.

This result shows how much the AI ​​has improved since its ChatGPT-enhanced predecessor, GPT-3.5. GPT-3.5 scored in the 10th percentile, just above 50% on the Unified Bar Exam, while GPT-4 scored in the 90th percentile to square off the court.

Similar large jumps in test scores were seen for the LSAT, quantitative and oral GRE, and medical knowledge self-assessment programs.

Test scores on the GPT-4 suggest good basic reasoning and comprehension skills, but low scores on the AP English Literature and Language exam and the GRE Writing exam suggest that creative I’m still struggling with my thoughts. Scores on these tests haven’t changed from his GPT-3.5 performance, so there seems to be little improvement in this area.

OpenAI claims that its new model “exhibits human-level performance” while “underperforming humans in many real-world scenarios.” The company added that it finished training the new AI last August, but put the training on hold to make the chatbot safer for users.

“GPT-4 is 82% less likely to respond to requests for unauthorized content and 40% more likely to generate factual responses than GPT-3.5,” the company said.

Tempering expectations, the company says that GPT-4, like its predecessors, is a “hallucination”, explaining when the AI ​​produces sentences that are coherent but not factual or not based in reality. ,” he added. OpenAI CEO Sam Altman says the model is significantly less hallucinogenic than his GPT-3.5, but the company is looking forward to “feedback on its shortcomings.”

Another improvement that GPT-4 has over its predecessor is that it now works in dozens of languages. During testing, GPT-4 performed better than his GPT-3.5 in English on multiple-choice questions such as Latvian, Welsh and Swahili.

GPT-4 uses a transformer-style architecture for neural networks. This method uses attention mechanisms loosely inspired by human cognition to enable neural networks to parse more relevant data than others, improving model accuracy and reducing training time. Shorten.

trendy now

• Paris Hilton says she had an ‘inappropriate’ relationship with her 8th grade teacher

• ‘The Last of Us’: Every Location in Alberta Has Transformed Dramatically

The idea of ​​Transformers was revolutionized and popularized in a 2017 paper by researchers at Google Brain. attention is all you needOne of the principal investigators on the paper was Aidan Gomez, now CEO of Cohere in Toronto. Cohere is a Canadian natural language processing company that operates in the same space as OpenAI.

Many of the world’s most important advances in machine learning that underpin ChatGPT were pioneered by Canadian scientists.

Three men have been hailed as the godfathers of AI, two of whom are Canadian. Yoshua Bengio from the University of Montreal and Geoffrey Hinton from the University of Toronto (U of T). The third, his Yann LeCun, is French, but some of his most groundbreaking research was done at Bell Labs and his U of T.

In fact, Ilya Sutskever, chief scientific officer and co-founder of OpenAI, was educated at T University and was a PhD student at Hinton.

As for Bengio, he The world’s most cited computer scientist.When asked if he could draw a direct line to ChatGPT from his work, he bluntly said, “Yeah, definitely.”

Bengio warns that AI could disrupt the social and economic fabric of the world through unemployment, misinformation campaigns and potential AI-powered weapons. He and other scientists are calling for greater regulation of AI to ensure that its benefits are available to all.

He also points out that ChatGPT is far from being able to reason like humans, and that such technology is still a long way off. But he is confident that one day humans will be able to create artificial general intelligence with human-level perception.

“What is inevitable is that scientific progress will get there.

OpenAI acknowledged in its GPT-4 research paper that its tools can be used maliciously, writing: We are collaborating with external researchers to improve how we understand and assess potential impacts and build assessments of potentially dangerous features that may emerge in future systems. ”

The company added that it will publish a follow-up paper with “recommendations on steps society can take to prepare for the impact of AI and initial ideas for predicting the likely economic impact of AI.” rice field.

© 2023 Global News, a division of Corus Entertainment Inc.