Timestamp Drift in Multimodal Models

On Monday’s show, Brian and Andy broke down several AI developments that surfaced over the weekend, focusing on tools and research that point toward more autonomous, long running AI systems. The discussion opened with hands on experience using ElevenLabs Scribe V2 for high accuracy transcription, including why timestamp drift remains a real problem for multimodal models. From there, the conversation shifted into DeepMind’s “Patchwork AGI” paper and what it implies about AGI emerging from orchestrated systems rather than a single frontier model. The second half of the show covered Claude Code’s growing influence, new restrictions around its usage, early experiences with ChatGPT Health, and broader implications of AI’s expansion into healthcare, energy, and platform ecosystems.

Key Points Discussed

ElevenLabs Scribe V2 delivers noticeably better transcription accuracy and timestamp reliability

Accurate transcripts remain critical for retrieval, clipping, and downstream AI workflows

Multimodal models still struggle with timestamp drift on long video inputs

DeepMind’s Patchwork AGI argues AGI will emerge from coordinated systems, not one model

Multi agent orchestration may accelerate AGI faster than expected

Claude Code feels like a set and forget inflection point for autonomous work

Claude Code adoption is growing even among competitor AI labs

Terminal based tools remain a barrier for non technical users, but UI gaps are closing

ChatGPT Health now allows direct querying of connected medical records

AI driven healthcare analysis may unlock earlier detection of disease through pattern recognition

X continues to dominate AI news distribution despite major platform drawbacks

Timestamps and Topics

00:00:00 👋 Monday kickoff and weekend framing

00:02:10 📝 ElevenLabs Scribe V2 and real world transcription testing

00:07:45 ⏱️ Timestamp drift and multimodal limitations

00:13:20 🧠 DeepMind Patchwork AGI and multi agent intelligence

00:20:30 🚀 AGI via orchestration vs single model breakthroughs

00:27:15 🧑‍💻 Claude Code as a fire and forget tool

00:35:40 🛑 Claude Code access restrictions and competitive tensions

00:42:10 🏥 ChatGPT Health first impressions and medical data access

00:50:30 🔬 AI, sleep studies, and predictive healthcare signals

00:58:20 ⚡ Energy, platforms, and ecosystem lock in

01:05:40 🌐 X as the default AI news hub, pros and cons

01:13:30 🏁 Wrap up and community updates

The Daily AI Show Co Hosts: Andy Halliday, Brian Maucere, and Carl Yeh