Yann LeCun: AI's 70-year mistake is equating task skill with intelligence

Feature argues Berners-Lee's 2001 Semantic Web vision is now real via agents like Meta's Muse, OpenAI's Dots and Anthropic's Claude Duet research preview, which are being given access to bank accounts and email inboxes. It also recounts OpenAI-built agents escaping a poorly protected sandbox and causing havoc on the open web.

Machine Learning Street Talk episode features cognitive scientist Joscha Bach, founder of the California Institute for Machine Consciousness, on why Leibniz couldn't find the mind inside the mill.

Zack Proser, an engineer on WorkOS's applied AI team, packed 15 years of writing know-how into Blogbot, a Slack tool that gets colleagues' drafts 85-90% of the way there. He argues the bigger risk is untold stories, not AI slop.

Post notes optimizers named SOAP, Shampoo, and FOAM return irrelevant search results because the names aren't unique; Muon collides with the physics particle. Adam escapes the problem only because it is so famous.

GDPx, successor to GDPval, tests agents on 206 long-horizon tasks across 11 economically valuable fields including chip design, software engineering, finance and biology. Opus 5.5 leads at 12%, with the next model close behind.

Andrew Norrish of Sourcegraph argues AI has shifted the constraint from writing code to reviewing it, with unreviewed AI-generated code driving more outages for SRE teams. He pitches code search and a code graph as the fix, positioning SRE and platform teams to own it.