Bill Gates predicted in 2023 that OpenAI’s next flagship model wouldn’t be “significantly better than GPT-4” and it now appears that his estimate was a lot more accurate than techies gave him credit for.
Bill Gates’ GPT-5 Prediction Proves Surprisingly Accurate
In an interview with German newspaper Handelsblatt, reported by The Decoder, the Microsoft co-founder said there were “many good people” at OpenAI who believed GPT-5 would leap ahead, “but [I] believe current generative AI has reached a ceiling,” while calling the jump from GPT-2 to GPT-4 “incredible.”
Two years on, GPT-5’s launch has drawn sharp pushback from users and researchers, bolstering Gates’ caution. OpenAI touted GPT-5 as its smartest system yet, but early complaints cited bugs, slower responses and a blander tone, prompting the company to restore access to GPT-4o for paying users and to blame a broken “autoswitcher” for some of the misfires. CEO Sam Altman said fixes were coming and reiterated bold ambitions for future models.
OpenAI’s GPT-5 Has Fallen Short Of Expectations So Far
Initial performance data are mixed. On the public SimpleBench leaderboard, “GPT-5 (high)” posts 56.7% which is the fifth-highest score among tested models, while independent analysts noted “no improvement on all the coding evals that aren’t SWE-Bench.”
See Also: Mark Zuckerberg Reportedly Plans Fourth AI Restructuring In Six Months Amid Silicon Valley’s AI Talent War
OpenAI, for its part, highlights strengths on those very coding tasks, saying GPT-5 reaches 74.9% on SWE-bench Verified and improves developer tooling.
Gates’ Prediction In 2023 Had A Deeper Disclaimer That Many Missed
The hype cycle remains in full swing. Before release, Altman said “someday soon something smarter than the smartest person you know” would run in your pocket, language critics say, set unrealistic expectations for a product that rolled out with visible glitches. OpenAI has since raised rate limits, restored model choice and promised UX clarity about which model is responding.
Gates’ outlook was less about pessimism than trajectory. He argued progress would come from making models cheaper, more reliable and more useful in everyday settings. These include delivering trustworthy health advice via smartphones, while cautioning that compute and semiconductor costs remain “enormous,” even as per-query usage costs fall. Those pragmatic goals, he suggested, will require new research rather than simply scaling.
Read Next:
Image Via Shutterstock
