OpenAI has officially announced GPT-4, the successor to its existing GPT-3.5 model for ChatGPT. Like with previous iterations, the improved large multimodal model is capable of emitting text outputs via image and text inputs. While less capable than humans in real-world scenarios, it exhibits human-level performance on various professional and academic benchmarks.
Through taking standardized exams, OpenAI is able to gauge the level of improvement to a degree. While GPT-3.5 scored around the bottom 10% of test takers for the bar exam, GPT-4 passed with a score in the 90th percentile of test takers. Other results include an improvement from 4/5 to 5/5 on the AP Biology exam and 590/800 to an estimated 700/800 on SAT Math.
OpenAI reports that after rebuilding the entire deep learning stack over the past two years, the new training runs for GPT-4 have been “unprecedentedly stable” with the team able to accurately predict performance ahead of time which is crucial for safe scaling. There has also been an increase in the chatbot’s steerability, with users able to better prescribe the desired behavior of the AI in specific styles or tones, as well as a decrease in the model’s tendency to hallucinate facts, make reasoning errors, and respond to requests for disallowed content.
More information can be found on the OpenAI website.
Elsewhere in tech, Daniel Arsham and CASETiFY launch Nebula 928 accessories line.