HSBC's CEO warns not every AI data center bet will pay off, as Microsoft brands agents the new insider threat. Plus DeepSeek's cheaper agentic model, Anthropic's AI safety bill rewrite, Stripe's voice-payment agents, a look inside Anthropic's interpretability lab, Sean Goedecke on requirements, Perplexity's open-sourced agent sandbox, Derek Thompson on math grief, and a wildcard: the first known green star.

CNBC · Oct 20

HSBC's CEO says investors should be selective about AI data center financing, warning not every deal merits support.

Julie Bort, TechCrunch · Oct 19

Microsoft rolled out tools to monitor and restrict AI agents in enterprise software, treating them as insider-threat risks.

Matthias Bastian, The Decoder · Oct 17

DeepSeek released V4.1-Flash, a cheaper open-weight model that outperforms larger rivals on agentic coding tasks.

Ana-Maria Stanciuc, The Next Web · Oct 20

Anthropic says it will revise part of its draft federal AI safety bill after safety-advocate criticism.

Unite.AI · Oct 20

Sierra and Stripe now let AI voice agents take payments and send receipts within a single phone call.

Lenny Rachitsky, Lenny's Newsletter · Oct 18

An Anthropic interpretability researcher discusses what looking inside language models reveals about planning and internal structure.

Sean Goedecke · Oct 19

Argues that identifying which requirements you don't actually have is key to simpler software design.

Asif Razzaq, MarkTechPost · Oct 16

Perplexity open-sourced a Firecracker microVM sandbox for running AI agent code in isolation at low latency.

Derek Thompson · Oct 15

Explores why mathematicians are grieving as AI systems solve problems once worked out through human collaboration.