By Roman Aglaganov
ChatGPT launched in November 2022 running on GPT-3.5. Four months later, on 14 March 2023, OpenAI released GPT-4, and the conversation about AI shifted from “fun toy” to “wait, it can do that?”
The exam results
OpenAI’s technical report listed GPT-4’s scores on human exams. The one that made headlines: on a simulated US bar exam, it scored around the top 10% of test takers, where GPT-3.5 had been around the bottom 10%. It also did well on SAT, GRE and many Advanced Placement exams.
Later research argued that the bar exam percentile was flattering, because it was compared against a pool of test takers that included many repeat candidates; against first-time takers, the estimate came out lower. Still impressive for software, slightly less “AI is now a lawyer”.
Seeing, not just reading
GPT-4 could also accept images as input: a photo of a fridge’s contents, a hand-drawn website sketch or a chart. Multimodal models, which handle text, images and later audio together, became the standard.
The secrecy
The report was notable for what it left out. Citing competition and safety, OpenAI declined to share the model’s size, architecture, training data or compute, a break from the field’s habit of publishing details, and a sign of how commercial AI had become.
Legacy: GPT-4 set the bar that rivals chased for the next year, and its successors, starting with GPT-4o in 2024, made voice conversation and multimodal input everyday features.
