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

#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.

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

May 25, 20261h 33mSummary

In this episode of Last Week in AI, the hosts dive into a packed week of developments, headlined by major announcements from Google I/O. The discussion centers on the introduction of Gemini 3.5, specifically the high-speed Gemini 3.5 Flash, and the unveiling of Gemini Spark—an agentic browser assistant designed to operate autonomously in the background via cloud-based virtual machines. The hosts analyze the architectural shift toward agentic workflows and note Google’s adoption of the Model Context Protocol (MCP), signaling a significant win for Anthropic’s open standard. The episode also covers the rollout of Gemini Omni, a unified multimodal model capable of reasoning across audio, video, and text. The hosts explore how Google is leveraging its massive proprietary data—such as Google Street View—to advance world modeling, a move that could provide a strategic advantage in training autonomous agents. Beyond Google, the hosts touch on the latest updates to Cursor’s Composer 2.5 and briefly discuss the conclusion of the OpenAI vs. Elon Musk trial. The conversation concludes with a thoughtful debate on the trend of recursive self-improvement and the challenges major labs face in converting massive infrastructure and data advantages into true frontier-leading products.

#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.