2025-05-23

Why Google Might Win The AI Race w/ Salim & Dave | MOONSHOTS

A whiplash week-in-review: OpenAI buys Jony Ive's device startup io for $6.5B to build an AI-first consumer device, then Google I/O drops a torrent of announcements (Gemini 2.5 Pro Deep Think, Google Beam, AI Mode in search, Project Mariner agent mode, Imagen 4, Veo 3, Android XR glasses, Project Astra's 'Jarvis' demo) only for Anthropic to grab the coding crown back within 24 hours with Claude 4. The panel digs into why algorithmic efficiency gains will increase rather than decrease AI's compute and energy appetite (Jevons paradox), debates the death of digital privacy, and flags national compute access as an emerging geopolitical fault line. The episode closes on Anthropic's aggressive 'disciplines solved by AI' timeline (math by 2028, materials science by 2030) and Bitcoin's surge past Amazon and Google's market cap.

▶ Watch on YouTube

Topics

OpenAI buys io (Jony Ive) for $6.5B AI ▶ 2:33

OpenAI's acquisition of Jony Ive's no-revenue device startup is framed as a bet that an AI-first consumer device can bypass the iPhone's advantages entirely, forcing every major AI lab to think about owning the consumer front end.

Gemini 2.5 Pro Deep Think sweeps the leaderboard AI ▶ 25:37

Google unveils Gemini 2.5 Pro with a new 'Deep Think' reasoning mode, claiming #1 across image generation, hard prompts, coding, math, and long-context benchmarks at Google I/O.

Claude 4 retakes the coding crown in 24 hours AI ▶ 15:39

Anthropic launches Claude 4 the day after Google's announcement, reclaiming the #1 SWE-bench coding spot; the panel notes crossing ~80% on SWE-bench is the tipping point where AI-written apps become reliably functional.

Algorithmic efficiency gains and Jevons paradox Compute ▶ 18:08

A slide projecting 30x-240x algorithmic efficiency gains sparks a discussion of Jevons paradox: more efficient AI models don't reduce compute/energy demand, they increase it, because capability growth drives usage up faster than cost falls.

Google Beam and the $250/month Ultra tier AI ▶ 29:11

Google Beam turns 2D video calls into realistic 3D experiences; access to the newest capabilities requires a $250/month Google AI Ultra subscription, a bold price point the panel debates as the new cost of staying competitive.

Real-time speech translation in Google Meet AI ▶ 32:56

Google demos live speech translation that preserves the speaker's own voice, reopening a debate on whether kids still need to learn foreign languages or coding once AI can do both instantly.

Project Mariner 'agent mode' and personalized smart replies AI ▶ 37:50

Google brings agentic browsing (booking apartments, running errands) into the Gemini app and search, plus smart replies that mimic the user's tone by reading their emails, raising both excitement and unease about AI acting on someone's behalf.

AI surveillance and the death of privacy Other ▶ 43:08

The panel argues there's effectively no regulation around AI listening on calls and meetings, and that convenience will make people accept AI access to their data even as traditional privacy disappears.

Veo 3 native-audio video generation AI ▶ 56:58

Google's Veo 3 generates video with synchronized sound effects, ambient audio, and dialogue; the panel calls it a 'grand slam home run' and predicts it will hollow out traditional Hollywood production.

Android XR glasses and Project Astra's 'Jarvis' demo AI ▶ 1:03:24

Google partners with Warby Parker and Gentle Monster on Android XR smart glasses, and a Project Astra demo shows an assistant handling a multi-step bike-repair task across apps, email, and phone calls in real time.

National compute access as a geopolitical resource Geopolitics ▶ 1:12:51

Peter and Salim argue roughly 180 of 200 countries have no national compute strategy, meaning most of the world's population won't get access to capabilities like the Astra demo unless governments treat compute as critical infrastructure now.

Anthropic's 'disciplines solved by AI' timeline AI ▶ 1:18:52

A chart attributed to Anthropic projects AI will 'solve' pure mathematics by 2028, computational chemistry by 2029, materials science by 2030, and climate/earth modeling by 2033, which the panel finds both exciting and under-scrutinized.

Predictions made

open Dario Amodei: The first $1 billion company run by a single human employee, with AI agents doing the rest of the work, will exist.

· due: 2026 · ▶ watch

“2026 we'll see the first $1 billion company with one human employee.”

Your call:

open Dave Blundin: Compounded algorithmic efficiency gains (quantization, reasoning-chain pruning, parameter specialization) will push AI model efficiency improvements to roughly 1,000x-10,000x, far beyond the already-aggressive 30-240x shown in the slide being discussed.

· due: unspecified · ▶ watch

“no doubt in my mind that this is way understated both in time and in impact.”

Your call:

open Dave Blundin: AI chip demand will keep chips sold out for at least five years, and possibly indefinitely, because Jevons-paradox dynamics mean more efficient models increase total compute demand rather than reducing it.

· due: 2030 · ▶ watch

“chips will be sold out for at least the next 5 years and maybe infinitely in the future.”

Your call:

open Anthropic (slide shown on podcast): Pure mathematics will be 'solved' by AI, meaning all currently unsolved math challenges get resolved by AI systems.

· due: 2028 · ▶ watch

“pure mathematics being solved by 2028.”

Your call:

open Anthropic (slide shown on podcast): Most diseases will become curable, per the inflection point marked on the same disciplines-solved forecast chart.

· due: 2027 · ▶ watch

“most diseaseable curable uh by in two and a half years.”

Your call:

open Demis Hassabis (reported by Peter Diamandis): All diseases will be cured within the next decade using AI-driven biotech breakthroughs.

· due: 2035 · ▶ watch

“they're going to have all diseases uh you know cured within the next decade.”

Your call:

open Peter Diamandis: Hollywood studios in their current form won't survive as AI video generation (Veo 3 and successors) makes professional-quality filmmaking cheap and on-demand.

· due: 2028 · ▶ watch

“how do they continue to exist in two or three years from now?”

Your call:

open Dave Blundin: Bitcoin will surpass Microsoft's market cap, and eventually gold's, though reaching gold's valuation will take much longer.

· due: unspecified · ▶ watch

“we'll see Bitcoin surpass uh Microsoft as well.”

Your call:

Numbers that matter

OpenAI acquires Jony Ive's device startup io for $6.5 billion A no-revenue acquisition paid mainly for Jony Ive and his team, likened to a Steve Jobs-style talent acquisition.
AI-coded apps become reliably functional once SWE-bench scores cross ~80%; below 50-60% they're too buggy to ship overnight Why Claude 4's benchmark jump matters more than it looks on the chart.
Lithium-ion battery price-performance has dropped 90% over the last decade Used as an analogy for how fast AI model efficiency can improve.
Fuel needed to fly across the Atlantic has dropped by two-thirds over the last few decades Real-world example of efficiency gains outpacing naive extrapolation, applied to AI compute/energy forecasts.
A slide projects 30x-240x algorithmic efficiency gains for AI models; Dave Blundin argues the real multiplicative total is closer to 1,000x-10,000x Quantization alone contributes roughly 20-40x, with reasoning-chain pruning and parameter specialization multiplying further.
Google AI Ultra (Gemini Ultra) subscription costs $250/month for full access to the newest I/O capabilities Framed as likely to become mandatory for professionals despite the high price point.
The financial-services industry spends about $100 billion a year on back-office operations Dave Blundin says are AI-automatable Based on sampling done at his firm's own portfolio companies.
A synthetic, AI-generated pharmaceutical TV ad was produced for about $500 versus a typical $500,000 production budget Illustrates generative video's cost collapse for professional-grade ad content.
Polymarket odds for leading AI lab by end of May 2025: Google 80%, Anthropic 19%, XAI 1% Snapshot of market sentiment right after Google I/O and Claude 4's launch.
Polymarket odds for leading AI lab by end of 2025: Google 38%, OpenAI 26%, XAI 23% Framed as effectively a four-horse race, including Anthropic, that could still flip repeatedly.
Bitcoin's market capitalization surpassed Amazon's and Google's this week Cited by Salim as proof of democratized ownership versus hard-to-access assets like gold.
Bitcoin crossed $110,000, an all-time high, with about $50 billion traded per day Panel cites Bitcoin's liquidity as remarkable relative to its market cap.

Worth digging into

🕳️ Jevons paradox applied to AI compute demand

The panel argues AI efficiency gains will increase total compute and energy demand rather than shrink it, upending the naive take that cheaper models mean less power needed — worth rigorous economic modeling.

🕳️ Shane Longpre's parameter-relevance pruning research at MIT

Claims roughly 100x efficiency gains are available just by excluding irrelevant knowledge (e.g., Swahili, quantum physics) from a given query's active parameters — a concrete, near-term efficiency lever distinct from quantization.

🕳️ Anthropic's 'disciplines solved by AI' timeline chart

Dates as specific as pure math by 2028 and material science by 2030 are aggressive and worth stress-testing against the actual methodology behind them, which the panel admits is gated by simulation and synthetic data.

🕳️ National compute access as the new geopolitical divide

Peter and Salim argue roughly 180 of 200 countries have no national compute strategy and won't be able to give citizens access to frontier-AI capabilities like the Astra 'Jarvis' demo, creating a new form of global inequality.

🕳️ Dave Blundin's synthetic-celebrity-ad startup, tested with LeBron James at Everquote

Early tests show AI-generated celebrity ads timed to peak relevance windows (like playoffs) drive strong click-through, hinting at a new business model where celebrities license their likeness without ever showing up to shoot anything.