Social media encourages the worst of AI boosterism (www.technologyreview.com)

Recent tensions in the AI community highlight the dangers of overzealous claims on social media, particularly regarding advancements in mathematics using large language models (LLMs). Demis Hassabis, CEO of Google DeepMind, responded to OpenAI researcher Sébastien Bubeck’s enthusiastic announcement that GPT-5 had solved ten unsolved mathematical problems—claims that were quickly challenged by mathematician Thomas Bloom. He clarified that GPT-5 did not discover new solutions but rather identified existing ones that were previously unacknowledged in the literature, demonstrating both the potential for LLMs to enhance research and the pitfalls of premature hype. This incident underscores a critical divide in how AI advancements are communicated and evaluated. While the ability of GPT-5 to locate prior solutions is noteworthy and has implications for the efficiency of literature review in mathematics, it diverges from the transformative breakthroughs often sought by AI advocates. The ongoing excitement surrounding these models is juxtaposed with sober assessments in other fields, such as medicine and law, where LLMs have faced scrutiny for inconsistent performance. The broader takeaway emphasizes the need for cautious, comprehensive evaluation of AI capabilities rather than sensational claims, suggesting that as the technology progresses, a more measured approach to reporting its achievements will be essential.