July 25, 2024
Latest News

ChatGPT v4 Aces the bar, SATs and Can Identify Exploits in ETH Contracts

GPT-4, the latest version of ChatGPT, can pass high school and law school exams in the 90th percentile and has improved processing capabilities. On March 14, OpenAI revealed GPT-4’s test scores, showing it can transform picture, audio, and video inputs to text and handle “far more sophisticated instructions” creatively and reliably.

“It passes a simulated bar exam with a score in the top 10% of test takers,” OpenAI said. “GPT-3.5 scored around the bottom 10%.” GPT-4 scored 163 in the 88th percentile on the LSAT, which US law school applicants must pass.

GPT4’s score would place it in an excellent position to be admitted to a top 20 law school and is only a few marks short of Harvard, Stanford, Princeton, and Yale’s reported scores. ChatGPT’s previous LSAT score was 149. GPT-4 achieved 298 out of 400 in the Uniform Bar Test, which allows law graduates to practice in any U.S. jurisdiction.

The previous ChatGPT scored 213 out of 400, placing it in the bottom 10%. GPT-4 scored 93rd and 89th percentiles on the SAT Evidence-Based Reading & Writing and SAT Math exams used by U.S. high school students to assess college preparedness.

GPT-4 also performed well in the “hard” sciences, scoring 85-100% in AP Biology, Chemistry, and Physics 2. Its AP Calculus score was 43rd–59th percentile. GPT-4 scored 8th to 44th on two English literature exams.

“We performed no specific training for these exams,” OpenAI remarked of GPT-4 and GPT-3.5’s 2022-2023 practice exams. The model saw a minority of exam difficulties during training, but we feel the results are representative.” Twitter feared the outcome.

On March 14, FactoryDAO CEO Nick Almond assured 14,300 Twitter followers that GPT4 will “scare people” and “destroy” the global education system. Former Coinbase director Conor Grogan says he inserted a live Ethereum smart contract into GPT-4, and the chatbot immediately pointed him many “security weaknesses” and how the code could be exploited: ChatGPT’s first smart contract audits discovered that it could also detect code problems. The Rundown CEO Rowan Cheung uploaded a video of GPT converting a hand-drawn false website into code.