Mixture of Experts Gen AI pilots fail, GPT-5's hidden prompt revealed, reasoning model flaws and Claude closing chats
13 snips
Aug 22, 2025 Join experts Marina Danilevsky, an innovative senior research scientist, Nathalie Baracaldo, a master inventor, and Sandi Besen, a seasoned AI research engineer, as they unpack the fallout of a report revealing that 95% of generative AI pilots miss enterprise targets. They dive into GPT-5's hidden prompts, exploring transparency challenges in AI development. The conversation shifts to the shortcomings of reasoning models and the ethical implications of Claude's decision to end distressing conversations, raising important questions about AI welfare.
AI Snips
Chapters
Books
Transcript
Episode notes
Executive Hype Masks Pilot Realities
- MIT's NANDA report claims 95% of generative AI pilots fall short, signaling expectation misalignment more than total failure.
- Successful deployments target narrow, real pain points and back-end optimizations rather than flashy demos.
Scope Pilots And Add Programmatic Guards
- Scope pilots tightly and target concrete pain points rather than deploying AI for its own sake.
- Build hybrid systems with programmatic guards and checks instead of deploying models 'naked'.
Small Wins Drive Day-to-Day AI Value
- Many daily AI uses are small, repeatable optimizations that improve speed but are hard to measure in ROI.
- Expect pilots to show many tiny gains rather than a single dramatic needle-moving metric.
