OpenAI’s Joseph Kim discussing the threshold phenomenon during his talk at AIE NY today.
One of the central themes at this week’s AI Engineer NYC conference was balancing agent autonomy with safety. Both frontier labs and financial institutions highlighted this topic across multiple talks throughout the week.
The core takeaway: safe autonomy depends on where you place controls and human judgment, not on how many restrictions you impose.
Sponsor: Build shopping experiences for your agents with shop.dev
Adding e-commerce capabilities to agents usually means building search and checkout flows from scratch. shop.dev provides a unified API for shopping across any agent or application. With a single integration, you get access to real-time pricing and inventory across more than 1 billion products, enable fast sign-in with Shop, and complete one-click Shop Pay checkouts directly within conversational flows. It is open to all developers—no account or API key required to get started. → Try shop.dev today
Joseph Kim, Head of ChatGPT for Financial Services at OpenAI, delivered a keynote today at AIE NYC. (We also recorded a dedicated podcast episode with Kim, coming soon.) He spoke about reliability and the “threshold of usefulness,” which he described as the point where “a model becomes capable and reliable enough that it actually saves you time.”
