GPT-5.2 Can't Identify a Serial Killer & Was The Year of Agents A Lie? EP99.28-5.2
Thursday, 11 December 2025 · 2 min read · Listen to the episode ↗
The discussion centers on GPT-5.2, highlighting its extended context window, pricing changes, and user frustrations regarding verbosity and decision-making reliability, especially in complex tasks like identifying trustworthiness. The "year of agents" is critiqued for overselling AI advancements without substantial outcomes, while a shift towards enterprise models indicates a growing focus on practical AI integration. Additionally, the episode notes Google's planned ads for Gemini and a partnership between Disney and OpenAI, suggesting transformative developments in the industry.
Chris discusses the release of GPT 5.2, noting its features like a 400k context window and a 128k output, alongside a price increase. While improvements in vision and tool calling are highlighted, a user expresses disappointment with GPT 5.2's verbosity and lack of conciseness, particularly in outputs for Create with Code. The conversation shifts to a comparison with Anthropic's models, which are perceived to manage tool calls more effectively. The internal clock concept of AI models is discussed, emphasizing the ability to anticipate future corrections.
A significant point of discussion is GPT 5.2's performance in making judgments about trustworthiness. An experiment involving a photo of Ivan Malat, Australia's worst serial killer, reveals GPT 5.2's non-committal response, contrasting with clearer judgments from other models like Gemini and Claude. This inconsistency raises concerns about GPT 5.2's reliability in more complex tasks.
The speakers critique the overall performance of AI models, calling for industry-based tuning to enhance reliability and decision-making. They argue that while safety is important, the refusal mechanism in GPT models may hinder performance. The broader theme of the "year of agents" is discussed, with one speaker expressing dissatisfaction with the hype surrounding new AI features and the lack of substantial promotional efforts from AI companies.
The hosts note that while many organizations are using AI, a significant percentage of workers conceal their usage due to limited access to advanced tools. They identify emerging themes, including the struggle of AI labs to integrate basic AI into enterprises and a shift towards enterprise models in response to competition. The importance of education for information workers is stressed as essential for the future of work.
Speaker 1 draws an analogy between AI and self-driving technology, advocating for more advanced tool usage and effective context transitions. They acknowledge existing automation services but note their lack of user-friendliness. Despite challenges, there is optimism about the future of AI and the importance of building necessary infrastructure.
The conversation shifts to Google's plans to introduce ads to Gemini in 2026, raising questions about the implications. Additionally, a partnership between The Walt Disney Company and OpenAI is announced, allowing the use of Disney characters in OpenAI's Sora platform, marking a notable shift in Disney's approach.
Speaker 1 critiques Microsoft and its new CEO, expressing disappointment in the lack of progress in AI development. The conversation concludes with a humorous invitation for listener feedback and a tease for the next episode, hinting at a diss track by GPT-5.2.
This summary was generated from the episode transcript and can contain mistakes.