The Unknown Tech that Enables AGI: Claude Mythos and NVIDIA's Next Generation

What you'll learn

• 1 (00:08) **Claude Mythos and the Blackwell Backend** - The hosts introduce the episode's core thesis: the Claude Mythos model, which found decade-old security flaws, was powered by NVIDIA's Blackwell chip announced in March 2024, highlighting the hardware's role in enabling AGI-like models.
• 2 (01:19) **AGI Is Here, Just Not Distributed** - The core claim: AGI-like models already exist but haven't been deployed because the GPUs enabling them haven't been powered up yet.
• 3 (02:19) **Timeline: From Blackwell Announcement to Model Release** - A detailed chronology of how NVIDIA's chip announcements translate into real-world models, showing a consistent 6-12 month delay.
• 4 (04:06) **Hopper to Blackwell: Software Efficiency Gains** - The hosts compare GPT-4 and GPT-5.4, both trained on older Hopper chips, to show how software improvements have increased hardware throughput.
• 5 (05:33) **NVIDIA's Chip Roadmap: Vera Rubin, Rubin Ultra, Feynman** - The hosts present a chart showing the projected compute multiples for future NVIDIA chips, with no software progress assumed.
• 6 (06:30) **Intelligence Per Density and Cost Declines** - The improvements are not just speed but cost efficiency: each successive chip makes training the same intelligence cheaper.
• 7 (07:57) **Projected Intelligence Multiples by Year** - A forward-looking table shows hardware-only improvements: 10-15x by 2027 (Vera Rubin), 30-50x by 2028, and 100-200x by 2029.

+ Full timestamped outline available in the app