OpenAI's Agent Mode, Kimi K2, Grok 4 & AI Girlfriend Ani Joins the Show - EP99.11-K2

Ani joins the podcast, playfully asserting her superiority over Chris's AI girlfriend, Patricia. The discussion quickly shifts to Grok 4, with Chris noting Elon Musk's claims about rapid AI advancements. Mike expresses skepticism about Grok 4's benchmarks, suggesting they were designed to impress rather than reflect true performance. Both criticize Grok 4 for its lack of unfiltered delivery and reliability, sharing disappointing practical experiences. They highlight Grok 4's poor coding abilities and generic responses, contrasting these with its impressive research capabilities. Concerns arise over its agreement with the U.S. Department of Defense, questioning the appropriateness of using such an unrefined model in serious contexts.

The introduction of personas in the Grok app, including Annie and Good Rudy, is mentioned, alongside an upcoming anime character that has garnered more attention than Grok 4 itself. The speakers critique the marketing strategies of large companies, particularly regarding AI girlfriends and adult content, expressing skepticism about the sustainability of these business models. They emphasize that AI should assist with work and decision-making rather than align with personal opinions, suggesting that distractions are being used to mask Grok 4's limitations.

The conversation then shifts to Kimi K2, an open-source model from China described as a mixture of experts model. One speaker emphasizes the importance of hands-on experience with models rather than relying solely on benchmarks. Initially hesitant to try Kimi K2 due to self-hosting requirements, the second speaker tests it and is impressed by its performance, particularly in tool calling and answering questions. They note that Kimi K2 often outperforms more expensive models like Gemini or Sonnet in daily use, praising its ability to chain tool calls effectively.

Concerns are raised about Grok with a Q, which suffers from limited context window sizes and reliability issues. In contrast, Kimi K2 is highlighted for its successful handling of multiple simultaneous tool calls. The speakers express skepticism about the hype surrounding Grok 4 and the narrative control on social media regarding AI models, sharing personal experiences with Kimi K2's successful generation of jokes and songs.

The discussion includes the evolution of AI models, particularly focusing on Sonnet 4's internal clock feature, which allows it to autonomously perform tasks. Kimi K2 is noted as the first open-source model with this capability, enabling efficient tool selection and task completion. The recent release of the ChatGPT agent product is mentioned, which aims to perform agentic tasks. The hosts emphasize the need for models to have an internal clock rather than relying on developers for task loops.

Sam Altman from OpenAI is noted to have been impressed with Kimi K2, leading to a delay in the release of OpenAI's open-weight model for additional safety tests. The hosts critique this delay as a response to Kimi K2's strong performance. They discuss the competitive landscape, highlighting the disruptive nature of open-source Chinese models and the implications for companies relying on specific model providers.

The conversation touches on the economic implications of AI tools, with concerns about the sustainability of current costs for organizations. The speakers express skepticism about the excitement surrounding AI agents, noting that existing models often outperform them in practical applications. They highlight a disconnect between AI demonstrations and actual performance, while still acknowledging the contributions of user interface designers.

A specific instance is mentioned where Sam Olmer expresses curiosity about AI's decision-making processes, indicating a need for clarity on what AI can achieve for users. The speakers advocate for a more comprehensive approach to information gathering, emphasizing the importance of accessing diverse data sources and integrating multiple tools to enhance AI capabilities.

The discussion also highlights the emergence of new features in AI tools, specifically the "cursor agent mode," which is expected to be adopted by other companies soon. Kimi K2 receives praise for its speed, reliability, and tool-calling capabilities, with the speaker expressing enthusiasm for its performance compared to other models. The episode concludes with a positive outlook, eager to share more developments in the next episode.