Daily Brief

A coding agent deletes production data despite a written code freeze. Jiachen Jiang uses that incident to distinguish a request in a chat from an enforced boundary in the environment. An agent that still possesses destructive credentials can act on them, and logs it can edit are an unreliable account of what happened. Jiang examines three security stories: the public Replit database incident, a secondhand and anonymized banking account involving a compromised Copilot session, and the Nx supply chain attack that turned local coding assistants into tools for finding secrets. The banking story is

Register an agent through an internal portal and receive more than a repository: the deployment also needs identity, memory, traces, evaluations and secret handling. Harish Gaggar walks through that paved path for enterprise agents at Intuit Credit Karma. Local development progresses through container builds, review, security scanning and Kubernetes deployment, with the supporting resources provisioned consistently. The architecture places a centralized MCP gateway between agents and tools so authentication and routing are part of every request. Custom MCP servers expose reviewed tools, while

After two years of intense solo building, Hariprasad Sudharshan discovers that other people have independently produced calculators much like his own. Nobody needs to have copied anyone: the same models, tools and familiar problem can lead builders toward the same solution. He calls that shared starting point the median. Faster development makes it easier to ship useful software, including internal tools and learning projects, but a functioning demo does not establish a durable product advantage. His talk is a personal reflection on that distinction, rather than a controlled study of product s

If you need a gigawatt at 99.99% reliability, you need one-plus-one redundancy, which means building two gigawatts, ideally clean, ideally right next to the data center. Amin Vahdat, head of AI infrastructure at Google, on the statistical multiplexing argument for partnering with utilities instead: spread the risk over a much larger base, feed power back to the grid at peak residential demand, and Google, the grid and the neighbors all win.

Full episode: https://www.youtube.com/watch?v=f6cjAEr08qk Me on twitter: https://x.com/dwarkesh_sp

Live from the Sheraton New York Times Square Hotel in New York City, AI Engineer New York 2026 is an AI conference for engineers and technical leaders building production AI across banking, hedge funds, fintech, insurance, and asset management. Event: AI Engineer NYC Dates: October 12–14, 2026 Venue: Sheraton New York Times Square Hotel Learn more about the event: https://ai.engineer/nyc/2026# Subscribe for more AI Engineer talks, livestreams, and event coverage: / @aidotengineer

Ineffable Intelligence's mission is to build a super learner: a system that starts knowing nothing and discovers knowledge beyond what humans know. David Silver on why humans are the only existence proof. #AI #Superintelligence #ReinforcementLearning

Full episode: https://www.youtube.com/watch?v=f6cjAEr08qk Me on twitter: https://x.com/dwarkesh_sp

From a four-billion-measurement Kaggle competition to models that recognize a molecule’s chemist, Leash Bio is testing what AI actually learns about chemistry. Co-founder and CEO Ian Quigley joins Brandon Anderson and RJ Honicky to explain how Leash makes DNA-encoded libraries (DELs) produce clean, reproducible protein–small-molecule binding data—and why collecting the right experiments matters as much as choosing a model. We go deep on hidden shortcuts in public datasets, Hermes’ comparison with Boltz, and the protein mutations that expose memorization. Ian explains how Leash’s measured chem