The Confluence Podcast for 5.24.2026
The episode examines a breakthrough in generative AI, highlighting OpenAI's new general model that solved the decades‑old planar unit distance problem and Anthropic's Mythos model uncovering major cybersecurity flaws. It contrasts these abstract achievements with the practical challenges of deploying AI in corporate settings, illustrated by Adam Kucharski’s experiment where default AI tools hallucinated differences in identical data sets. The hosts stress the need for friction—forcing AI to show its work—to avoid misleading outputs, and discuss a Carnegie Mellon study showing AI reviewers can outperform humans in academic peer review but suffer from context‑window limits and shared alignment biases.
Confluence for 5.8.26
The Confluence newsletter warns that AI‑driven meeting transcription can chill dialogue, create legal exposure, and distort context, urging teams to disable such agents unless explicitly needed. It highlights the rise of AI audio critique tools like NotebookLM that let creators...
Confluence for 5.3.26
The Confluence newsletter this week spotlights four themes: Amanda Askell’s practical view of LLM prompting as an empirical, high‑volume craft; the leadership dilemma of “uncertain uncertainties” where overconfidence can be fatal; new research exposing a gap between executive claims of...
Confluence for 4.19.26
Anthropic rolled out three major updates this week: the Opus 4.7 model with adaptive reasoning, Claude for Word in public beta, and Claude Design for AI‑generated visual work. Microsoft’s Copilot now offers Opus alongside its own models, tightening the Anthropic‑Microsoft partnership....