AI Scientist Bengio: Building Systems We Don't Know How to Control
THE SO WHAT
When Yoshua Bengio says we don’t have control of AI agents, he’s giving operators cover to slow deployment and demand governance budgets. If your board hasn’t asked for an agent risk register and escalation plan yet, this is the interview that will trigger it.
READ THE SOURCE
MORE FROM THE WIRE
Applied AII used Gemini in Google Sheets to create a weekly budget that updates itself — here’s how
End users wiring Gemini into Sheets is the quiet death of a thousand internal tools — every analyst is now a light automation engineer. If your product is just structured CRUD plus reporting, assume your users are already rebuilding 60–80% of it in their spreadsheet and plan accordingly.
Applied AIConsumer Tech Leaders Discuss the Industry Future & AI
Sonos, Oura, and Hatch talking AI is a tell — consumer hardware is shifting from “device + app” to “sensor + model + ongoing service.” If you ship hardware without a clear model lifecycle and data flywheel, you’re building a commodity accessory for someone else’s AI stack.
Applied AIOpenAI updates ChatGPT memory with a "more capable and compute-efficient" architecture and a summary page that lets users review and steer what it remembers (OpenAI)
Memory just became a first-class UX surface — users can now see and steer what the assistant remembers, which turns long-term personalization into something you can debug, not just hope for. For operators, this is the pattern: stateful assistants with user-governed memory will be the default, and your internal tools need the same transparency or they’ll lose trust fast.
Applied AIThinking Machines’ Murati on AI’s Next Chapter
A dedicated lab talking about the “future of human-AI interaction” is a reminder that the interface layer is still wide open — chat is a stopgap, not the end state. If you’re only experimenting with text boxes and not multimodal, continuous, or embodied interfaces, you’re already behind the next chapter they’re designing for.