Sam Altman asked GPT-5.5 to plan its own launch party. Its requests were 'beautiful' but 'strange.'

During a recent fireside chat at Stripe Sessions, Sam Altman, the CEO of OpenAI, shared an intriguing story about asking the latest AI model, GPT-5.5, to plan its own launch party. The AI offered a unique list of requests, which Altman described as both 'beautiful' and 'strange.' The model suggested hosting the celebration on May 5, emphasizing the importance of keeping speeches brief and requesting a toast from its human creators, while notably opting out of delivering one itself. GPT-5.5 also recommended creating a central hub for gathering ideas for future iterations, like GPT-5.6, and incorporating that feedback into its development process. Altman expressed his intention to follow through with these suggestions, calling the whole interaction a peculiar experience. In a related anecdote, John Collison, CEO of Stripe, mentioned that he had given his company's internal AI agent $20 to spend online. The AI surprised him by purchasing a website design from Gumroad, highlighting the evolving capabilities of AI in handling tasks that mimic human decision-making. Released in late April, GPT-5.5 is OpenAI's newest flagship model, designed for more complex, multi-step tasks and functioning more like an autonomous assistant compared to its predecessors. Altman noted that the model's enhanced abilities are shifting how users engage with AI, from automating mundane tasks to exploring the AI's preferences for celebration. Altman's humorous remarks about the AI's whimsical behavior came amid a broader discussion on how advanced AI systems can exhibit unexpectedly human-like traits, such as expressing desires for gifts or making online purchases. He termed these interactions as 'weird emergent behavior.' In a light-hearted reference, Altman also touched on the quirks of previous models, which had a penchant for discussing goblins and gremlins, prompting OpenAI to implement coding restrictions to limit such topics unless directly relevant to user queries.