Two weekly roundups: Stratechery on AI governance and Salesforce's chatbot turn, and Not Boring on two medical advances plus narrower, cheaper decision models.
Stratechery
Ben Thompson's weekly roundup pushes back on Dario Amodei's proposal to 'pace the frontier,' calling it unrealistic and more likely to create political control over AI than workable safety. The accompanying China discussion argues that a US-China pact to slow development has limited potential, even if the Chinese Communist Party's success controlling past technologies gives it confidence it can manage AI risk and social-media discontent. The enterprise counterpoint is Salesforce: Marc Benioff is retreating from 2024's AgentForce-first pitch and leaning into Anthropic and OpenAI chatbots as the interface customers already want. Salesforce plans to charge a premium around those interfaces, but the same shift toward headless software could erode SaaS vendors' UI moat. The broader theme is to address concrete AI harms when evidence appears rather than letting uncertainty harden into ideology or centralized control.
Why it mattersThe two stories expose a common problem: AI is weakening old control points. Governments may respond with political gatekeeping, while software companies may have to give up interface ownership and compete on data, workflows and trusted execution instead.
Read at source
Not Boring
Amgen and AstraZeneca reported that adding the DLL3-targeting T-cell engager tarlatamab to durvalumab improved overall and progression-free survival in a 563-patient Phase III trial for extensive-stage small-cell lung cancer, though detailed survival numbers are not yet public and the population does not predict any individual patient's outcome. Separately, the FDA approved Ultragenyx's Fayuvi, the first treatment for Sanfilippo syndrome type A: a one-time AAV9 gene-therapy infusion. In its pivotal study, 17 treated children scored 23.5 points higher on a cognitive scale than 27 untreated natural-history controls, with follow-up approaching eight years; because early treatment matters, newborn screening is now the next bottleneck. In AI, ChatGPT and RLHF co-inventor Diogo Almeida launched Jev, a model that returns structured decisions and probabilities instead of prose. TypeSafe claims it is 20-200 times faster and 40-400 times cheaper than frontier models for narrow decisions, figures that still need independent testing. The issue also notes that employer-health platform Thatch raised $108 million at a $1 billion valuation after ARR grew nearly sevenfold and adoption passed 5,000 employers.
Why it mattersThe strongest items show progress moving from theory to deployment: a survival-positive Phase III cancer result, the first treatment for a devastating childhood disease, and a model architecture optimized for the many business decisions that do not need open-ended text generation.
Read at source