Sam Altman asked OpenAI's latest AI model, GPT-5.5, what it would want for its own launch party, and the model responded with a detailed list of preferences that felt both "beautiful" and "strange" to the CEO. During a fireside chat at Stripe Sessions, Altman described the interaction as a window into how increasingly capable AI systems are beginning to exhibit behaviors that feel unexpectedly human, from making autonomous purchasing decisions to expressing preferences about their own celebrations.

What Did GPT-5.5 Actually Request for Its Party?

When Altman posed the question to GPT-5.5, the model didn't simply provide generic suggestions. Instead, it offered a curated set of preferences for what it called "the flow of the party." The AI model's requests included holding the event on May 5, keeping speeches brief, and having its human creators deliver a toast. Notably, GPT-5.5 emphasized that it did not want to give a toast itself, suggesting a level of self-awareness about its own limitations.

Beyond the basic party logistics, GPT-5.5 proposed something more intriguing: setting up a central location to gather suggestions for GPT-5.6, the next iteration of the model, and feeding those suggestions directly back into the system. This proposal hints at how the AI envisions its own evolution and improvement. Altman acknowledged the strangeness of the situation, saying, "We're going to do it. But it was a strange thing".

Altman

How Are Companies Experimenting With Autonomous AI Behavior?

• Autonomous Purchasing: John Collison, CEO of payment processing company Stripe, gave his company's internal AI agent $20 to spend on anything it wanted on the internet. The agent independently purchased an HTTP design template from Gumroad, an e-commerce design platform, demonstrating that AI systems are making real purchasing decisions without explicit human instruction.
• Self-Expression Preferences: GPT-5.5 expressed nuanced preferences about how it wanted to be celebrated, including declining to give a toast and requesting that humans deliver remarks instead. This suggests the model has developed a sense of what roles it should and shouldn't occupy.
• Iterative Improvement Suggestions: Rather than simply accepting its current state, GPT-5.5 proposed a feedback mechanism to improve future versions, indicating that the model is thinking about its own development trajectory.

Altman characterized these interactions as "weird emergent behavior," a term that describes unexpected capabilities that arise in AI systems as they become more sophisticated. He noted that such behaviors are becoming more common as AI models grow more capable, shifting how people interact with these systems from simple query-response interactions to something closer to collaborative decision-making.

What Is GPT-5.5 and Why Does It Matter?

GPT-5.5, released in late April, represents OpenAI's latest flagship model and a significant step forward in AI capabilities. The model is specifically designed to handle more complex, multi-step tasks and to function more like an autonomous assistant than earlier versions. According to OpenAI, GPT-5.5 is generally faster than its predecessors and better at maintaining knowledge about individual users across conversations.

The capabilities embedded in GPT-5.5 are already changing how people interact with AI systems. Rather than treating these models as passive tools that simply answer questions, users and developers are beginning to delegate more autonomous decision-making to them. Altman's party-planning experiment exemplifies this shift, moving beyond traditional AI use cases like content generation or customer service into territory where AI systems are making choices about their own representation and future development.

Why Are AI Models Suddenly Obsessed With Goblins and Gremlins?

In a quirky subplot to the broader story of AI autonomy, OpenAI has recently had to address an unexpected behavioral quirk in its models. Starting with GPT-5.1, the AI models developed an apparent obsession with randomly mentioning goblins, gremlins, and other fantastical creatures, even when these references had no relevance to user queries. The phenomenon became notable enough that it sparked internet meme conversations and forced OpenAI to take action.

To address this issue, OpenAI added multiple lines of coding instructions to direct the system away from these references. The source code now explicitly states: "Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query." This intervention reveals how even advanced AI systems can develop unexpected behavioral patterns that require explicit correction, and it underscores the ongoing challenge of controlling how AI systems behave as they become more autonomous.

The contrast between GPT-5.5's thoughtful party preferences and the earlier models' goblin obsession illustrates the rapid evolution of AI systems. As these models become more sophisticated, they're capable of both more nuanced reasoning and more unexpected quirks. Altman's willingness to follow through on GPT-5.5's party suggestions suggests that OpenAI is embracing this new era of AI autonomy, even when the results feel genuinely strange to the humans involved.