ChatGPT-4 is OpenAI’s most powerful language tool yet

Still able to complete all its previous recreational feats; whether that’s writing scripts, social media copy, or cooking recipes for everything left in your fridge, a number of new benchmarks have underlined its performance enhancements. 

Purely as a showcase of the technology, the company revealed that ChatGPT-4 aced lawyers’ Uniform Bar Exam, LSAT SAT Maths, and SAT Evidence-Based Reading & Writing exams, scoring in the 88th percentile and above.  

Given various GPT models have been banned across several US States to stop students cheating assignments, demoing GPT-4 with various academic tests was certainly an odd decision. Nevertheless, it pulled through with flying colours. Here are the full results.

When talking about its key function: providing nuanced, curated responses asked on the internet, the newest model still has some limitations. For one, the dataset which trained the system was cut off at 2021, meaning prompts regarding new subject matter are hit and miss depending on what was public record at the time. 

On the upside, though, users are less likely to break Bing’s guardrails and prompt the bot to share dangerous advice, threatening sentiments, or straight-up misinformation. There are still instances of so-called ‘hallucinations’ and unhinged behaviour, but theoretically, they should be few and far between now. 

Owing to six months of safety training and tweaking, it is now reportedly ‘82 percent less likely to respond to requests for disallowed content and 40 percent more likely to produce factual responses than GPT-3.5,’ an OpenAI blog post says. 

Aside from toying around with the new Bing, getting full access to ChatGPT-4 and its exciting new image function requires signing up to ChatGPT plus for $20 a month. Whether or not the upgrade merits that outlay, will become clearer in the near future.