Moonshots EP 174: Why Google Might Win The AI Race

Summary

A WTF episode with Diamandis, Blundin, and Ismail covering a packed week of AI announcements dominated by Google I/O, Claude 4’s launch, and OpenAI’s $6.5B acquisition of Jony Ive’s AI device startup. Dave frames OpenAI’s Ive acquisition as the move of all moves — Sam Altman is building a vertically integrated consumer AI company (search competitor to Google, plus hardware) rather than just a foundation model provider. The Google I/O breakdown is extensive: Gemini 2.5 Pro briefly claimed number one across all LM Arena benchmarks before Claude 4 reclaimed the coding crown within 24 hours. Dave calls it “The Empire Strikes Back” — Google had these capabilities in the lab but needed competitive pressure to release them. Key Google announcements include Deep Think reasoning mode, Google Beam (3D video communications), real-time speech translation in Meet, Project Mariner agentic web browsing in Chrome, Veo 3 video generation, and the $250/month Gemini Ultra tier. Dave argues this pricing proves foundation models are not commoditizing — the “race to the bottom” thesis was wrong. The algorithmic efficiency discussion is notable: Dave claims published estimates of 30-240x improvement are “way understated,” citing three multiplicative dimensions (quantization 20-40x, chain of reasoning improvements, and model subsetting ~100x) that together suggest 1,000-10,000x actual improvement, supporting Eric Schmidt’s “AI is massively underhyped” position. Jevons paradox applied to AI: more efficient models won’t reduce GPU demand, they’ll increase it because better capabilities drive more usage. Salim’s observation about language learning rewiring the brain and South Indian math competition dominance being linked to language structure is an interesting cultural data point.

Key Segments

• [00:00-09:00] OpenAI acquires Jony Ive startup for $6.5B, AI device strategy, always-listening privacy implications
• [09:00-18:00] Google Gemini 2.5 Pro benchmarks, Claude 4 leapfrogs within 24 hours, SWEBench 80% tipping point for functional code
• [18:00-24:00] Algorithmic efficiency improvements 1,000-10,000x (understated), Jevons paradox applied to AI compute
• [24:00-38:00] Google I/O recap: Deep Think, Google Beam, real-time translation, $250/mo pricing, education disruption
• [38:00+] Project Mariner agentic browsing, Veo 3 video generation, media industry disruption

Notable Claims

• OpenAI paid $6.5B for Jony Ive’s startup (no revenue) — largest acqui-hire in tech history for AI device play
• SWEBench scores above 80% represent a tipping point where AI can build entire functional products overnight; below 60% too many bugs
• Algorithmic efficiency gains are multiplicative across 3+ dimensions, suggesting 1,000-10,000x total improvement (not the published 30-240x)
• AI chips will be sold out for at least the next 5 years due to Jevons paradox
• Google Gemini Ultra at $250/month signals foundation models are definitively not commoditizing

Bias/Sponsor Notes

Blundin references Link Ventures portfolio company Blitzy (long-form coding) when discussing SWEBench benchmarks. Standard Diamandis newsletter/summit plugs. Panel has no disclosed Google or OpenAI positions but Blundin’s fund invests in AI companies that benefit from model improvements.