#246 - Gemini 3.5 + Omni, Musk Loses, OpenAI vs Erdős

#248 - Fable 5, Siri AI, IPOs, Policy on the AI ​​Exponential

Jun 17, 20261h 40mSummary

In this episode of Last Week in AI, the hosts dive into a packed agenda, starting with the release of Anthropic’s Claude 3.5 Sonnet. The discussion highlights the model’s significant leaps in coding and reasoning capabilities, noting that it effectively raises the level of abstraction in software engineering. However, the hosts also explore the controversy surrounding its stringent safety guardrails, which have sparked debate regarding their impact on AI research and the potential for silent downgrades during specific tasks. Beyond the technical performance, the episode examines the broader strategic landscape. The hosts analyze Apple’s integration of AI into Siri through a partnership with Google, questioning how this reliance on third-party models impacts Apple’s long-term competitive moat and privacy commitments. Additionally, the conversation covers the ongoing trends in the IPO market for major AI labs, noting how the high valuation of these companies creates a high-stakes environment where public market scrutiny could become a significant factor. The episode concludes with reflections on the limitations of current models in creative ideation, emphasizing that while these tools act as powerful accelerators, they still require significant human guidance to reach their full potential.

#247 - Opus 4.8, MAI, Anthropic IPO, Minimax-M3

Jun 6, 20261h 45mSummary

In this episode of Last Week in AI, the hosts dive into a packed week of developments across the rapidly evolving artificial intelligence landscape. A significant portion of the conversation focuses on the release of Anthropic’s Opus 4.8, analyzing its performance gains, its tendency toward verbosity, and the company’s internal philosophical approach to model safety and corrigibility. The hosts discuss how labs are increasingly shifting focus from simple model scaling to complex agent orchestration, highlighting how tools like Anthropic’s new dynamic workflows aim to tackle long-running tasks. The discussion also turns to the enterprise sector, where Microsoft is making waves with its latest AI announcements. The hosts examine the introduction of Microsoft’s personal agent and their new MAI model family, emphasizing how the company is prioritizing security infrastructure and governance to ensure that agentic systems can be safely integrated into corporate environments. Furthermore, the episode touches on broader industry trends, including the pivotal role of research agencies like DARPA in shaping AI development and the ongoing strategic maneuvering among frontier labs as they prepare for significant financial milestones.

#245 - TML-Interaction, Claude For Legal, Sam Altman on Stand

May 18, 20261h 49mSummary

In this episode of Last Week in AI, hosts Andri Karenkov and Jeremy Harris explore the rapid evolution of real-time conversational artificial intelligence. The discussion centers on the competitive landscape between OpenAI’s latest voice intelligence updates and the recent research preview from Thinking Machines. The hosts break down the engineering challenges of creating low-latency, time-aware models that can handle simultaneous speech, interjections, and multi-modal interactions. They provide technical insight into how these companies are managing the trade-off between reasoning depth and processing speed, noting that both organizations are aggressively refining their infrastructure to minimize the delay between human input and machine response. Additionally, the episode covers Anthropic’s new focus on the legal sector with the launch of Claude for Legal. The hosts discuss why legal professionals have become power users of AI, noting the significance of integrating these tools into existing legal workflows and enterprise software. Beyond technical updates, the conversation touches on the emerging policy challenges, safety guardrails, and the implications of using synthetic, real-time voices for professional tasks. The episode offers a comprehensive look at how these advancements are shaping the next generation of agentic AI interactions.

#241 - Opus 4.7, Muse Spark, GPT-5.4-Cyber, HY-World 2.0

Apr 23, 20261h 59mSummary

In this episode of Last Week in AI, hosts Andrew Karenikov and Jeremy Harris dive into a product-heavy week in the artificial intelligence industry, highlighting a significant wave of new tool releases and project updates. The discussion centers on the launch of Claude Opus 4.7, exploring its improved reasoning tiers, tokenization adjustments, and its competitive performance against other frontier models. The hosts analyze the model’s detailed system card, noting its literal instruction following and the fascinating, if complex, implications of "realism steering" in relation to deceptive behaviors during evaluations. The conversation further shifts to Meta's recent unveiling of its Muse Spark model. The hosts examine Meta’s trajectory, specifically their use of multi-round test-time scaling and RL thought compression to optimize reasoning efficiency. Additionally, they touch upon the infrastructure side of the industry, referencing Meta’s Hyperion project and the broader context of competitive GPU scaling. Throughout the episode, the hosts highlight the ongoing tension between rapid AI development and the critical importance of safety, alignment, and rigorous evaluation, providing listeners with a thoughtful breakdown of the current technical and strategic landscape.