Unsupervised Learning Ep 76: Ari Morcos from Datalogy AI and Rob Toews from Radical VC on AI Talent Wars, xAI’s $200B Valuation, & Google’s Comeback
78 snips
Sep 24, 2025 
In this discussion, Ari Marcos, CEO of Datalogy AI and former DeepMind researcher, teams up with Rob Toews, a partner at Radical Ventures. They explore whether AI model progress is truly slowing or merely shifting focus, and debate the economics of reinforcement learning environments. The duo also analyzes Google's rejuvenation with its Gemini project, the turbulent job market for AI talent, and the inflated valuations surrounding major players like OpenAI and Anthropic. Finally, they ponder the future of AI's integration with robotics and brain-computer interfaces.
AI Snips
Chapters
Transcript
Episode notes
Progress Is Shifting Not Stopping
- Model progress isn't slowing; it's changing shape from raw scale to quality and cost-efficiency gains.
- Small models now match past giants and new axes like reasoning and data efficiency drive impact.
RL Gains May Not Sustain Exponential Growth
- Pre-training scale shows diminishing returns and RL-based test-time compute gives targeted capability boosts.
- But Rob warns this paradigm may not sustain the same pace of broad capability growth.
Monkey Lab Cheating Reveals Overfitting Risk
- Ari recounts a monkey lab where animals exploited tiny luminance cues to 'cheat' tasks.
- He uses this to warn models will overfit imperfect RL environments and fail to generalize.