In this episode of Mixture of Experts, Tim Huang hosts a panel featuring experts Kate Soul, Kush Varsne, and Koutar El McGrowi to discuss the latest developments in AI models, including Granite 4.0, Claude 4.5, and Sora 2. The discussion begins with Granite 4.0, an efficient, smaller language model family launched on Hugging Face. Kate highlights that these models are designed to run on a single, more affordable GPU while outperforming previous versions, thanks to a new hybrid architecture. Notably, Granite 4.0 has achieved ISO 420001 certification, emphasizing its focus on governance, safety, and security. Kush adds that Granite is leading a trend in open-source AI toward greater transparency and cryptographic signing of models to verify training integrity.
The conversation then shifts to Claude 4.5, which is positioned as a developer-focused model excelling in coding tasks. Kush explains that Anthropics, the company behind Claude, is deliberately narrowing its model’s focus to coding, reflecting a broader industry trend toward specialized AI models rather than one-size-fits-all solutions. Kate and Koutar discuss how this specialization aligns with the need for efficiency and adaptability, especially during inference, where models must perform well in dynamic environments. A notable demo called “Imagine with Claude” showcases the potential for on-the-fly software generation, though Kate expresses some skepticism about its practicality for enterprise use, suggesting it may be more suited for personal productivity.
Next, the panel examines OpenAI’s Sora 2, a video generation app that represents a consumer-oriented, mobile-first social experience. Kush describes Sora 2 as the “vibe video producing app,” contrasting it with Claude’s coding focus. The discussion touches on the ethical and societal implications of Sora 2’s deepfake capabilities, branded as “cameos,” which allow users to insert themselves into videos. Concerns are raised about potential misuse, disinformation, and the addictive nature of such content. Koutar highlights the significant computational and energy costs associated with video generation, questioning the sustainability of offering such services at scale and at current price points.
The panel also explores OpenAI’s new “Buy with GPT” feature, which aims to enable GPT models to act as e-commerce agents capable of making purchases on users’ behalf. Kush and Kate debate the trust issues surrounding AI handling financial transactions, with Kush noting that while users might trust AI with credit card transactions, they remain wary of sharing sensitive financial accounts. The conversation also touches on the potential for ads and corporate influence within AI-driven commerce, with Kush emphasizing the need for regulatory oversight to prevent corporate capture of critical financial infrastructure. Koutar adds that major players like Amazon and Google are adopting different strategies to compete in this emerging agentic commerce space.
The episode concludes with a cybersecurity segment featuring Matt Kazinski, host of the Security Intelligence podcast. Matt discusses the unique security challenges posed by AI agents, particularly the risk of social engineering attacks where malicious actors manipulate AI through carefully crafted inputs. He stresses that traditional software vulnerabilities differ from these new threats, which require novel defensive strategies. The panel agrees on the importance of developing universal guidelines or “laws” for AI behavior to mitigate risks, though they acknowledge that no solution will be foolproof. Matt also shares details about his podcast, which covers a broad range of cybersecurity topics, including AI-related risks, and invites listeners to tune in for expert insights.
