0:00
8:42
Spola tillbaka 15 sekunder
Spola framåt 15 sekunder

We open with Anthropic’s new piece on the components of agent memory, which argues that the central challenge is curation, deciding what a model should keep and, more often, what it should leave out. Anthropic borrows from the human brain. The agent logs its work during a session, and an offline “dreaming” process later reviews those logs, fixes mistakes, and decides what to save. In one test, a single wrong note about the map left Claude stuck in a mountain cave in Pokémon Red until a dreaming pass corrected it. We share how readers can do the same in Claude Code, Codex, and the standard chat apps.

Next, we look at a Frontiers in Psychology study of 240 teachers in training. When the AI’s advice contained an error, those who wrote down their own answer first got it right 54% of the time, compared with 37% for those who saw the AI’s recommendation first, and the AI-first group repeated the bad advice 47% of the time. Then we turn to an HBR report on a BT Group trial with 2,400 employees setting performance goals, where a “challenger” AI that asked follow-up questions produced 3.5 times as many excellent goals as an agreeable one, and people enjoyed it just as much. We close on Ethan Mollick’s argument that worries about AI factual errors are outdated. We largely agree, and we talk about how review has shifted toward context, audience, and judgment, the work that still needs an experienced human.

Read this week’s edition: https://craai.substack.com/p/confluence-for-1182026



Get full access to Confluence: AI, Leadership, and Communication at craai.substack.com/subscribe

Fler avsnitt från "The Confluence Podcast: Thoughts on the Intersection of AI, Leadership, and Communication"