Theo Jaffee and Sophia Dew discuss a joint industry call for stronger cyber defence and reports about an AI security incident. They distinguish model alignment from external controls and monitoring, consider unresolved liability questions, and emphasize extending defensive tools to under-resourced critical infrastructure. Their discussion also covers reported model releases and a preview of interfaces for agents to operate laboratory equipment.
Sergey Edunov explains that AI drug discovery combines specialized physics and chemistry models, experimental data and language-model orchestration. Alex Zhavoronkov describes licensing drug candidates and collaborations as distinct from selling software, while warning that a promising molecule still faces lengthy testing and approval. Company pipeline and performance figures are presented as the speakers’ claims, not proof of clinical effectiveness.
Andrey Styskin argues for fresh search benchmarks that resist memorized answers and shortcuts, alongside more affordable retrieval for agents. Psyho discusses the difference between unaided model scores and systems equipped with memory and harnesses, and how AI changes competitive programming. Sayash Kapoor emphasizes consistency, calibration and safety beyond benchmark accuracy, and argues that reliability and oversight still constrain deployment.
Varun Sivaram describes shifting, slowing or buffering AI workloads to relieve grid peaks while respecting service constraints. Jen Zhu Scott discusses cooling efficiency and the strategic value of open-model distribution, treating the reported Nvidia-Hugging Face acquisition as an uncompleted deal. Both interviews address infrastructure economics, with commercial forecasts attributed to their proponents.
Pablo Palafox explains why enterprise operations combine deterministic processes with agents handling unstructured communication, and why organizational context and deployment work matter. Madeline Minshew and Tony Conrad discuss trust, shared context and model costs, suggesting that expensive frontier capability should be matched to tasks that justify it. The programme spans complementary interviews rather than a single independently verified performance study.
Watch on YouTube
