2025-11-20

What Everyone Missed About Gemini 3 w/ Salim, Dave & Alexander Wissner-Gross | EP#209

The panel dissects Google's Gemini 3 launch, arguing it is the biggest capability jump since OpenAI's o3 and that Google has decisively won the current AI race, evidenced by agentic Gemini, the Anti-Gravity coding IDE, and huge benchmark gains on ARC-AGI-2 and Humanity's Last Exam. They use the Vending-Bench Arena benchmark to argue AI agents are close to running zero-employee businesses, discuss GPT-5.1's inference-time compute economics, and cover Cursor's tripled valuation amid the coding-tool wars. A long segment covers AI biosecurity (Red Queen Bio, defensive co-scaling of safety with capability, the risks of dead open-source frontier AI in the US) before pivoting to Jeff Bezos's Project Prometheus and physical/world-model AI for manufacturing and space. The episode closes with a framework for how abundance becomes tangible: falling costs across housing, food, healthcare, transport and education driven by a roughly 40x/year drop in the cost per unit of intelligence.

▶ Watch on YouTube

Topics

Gemini 3 launch and Google ecosystem integration AI ▶ 4:07

Google shipped Gemini 3 with an agent mode that takes multi-step actions across Gmail, Calendar, YouTube and Workspace, plus generative UI that builds custom interactive widgets instead of text answers.

"Big model smell" and multimodal one-shotting AI ▶ 9:53

Alex Wissner-Gross describes qualitative jumps only visible in direct use: one-shot generating an interactive 3D voxel world from a photo, and the Anti-Gravity IDE (built partly by ex-Windsurf engineers) matching Cursor's coding experience.

Vending-Bench Arena: AI as autonomous economic actor Economy ▶ 14:04

Andon Labs' benchmark gives AI agents $500 and a simulated vending machine to run (emails, banking, restocking, pricing); Gemini 3 reportedly delivers ~3,000% more profit than GPT-5/Claude Sonnet, read as a proxy for zero-employee AI-run businesses.

One-shot game and app generation AI ▶ 20:42

Alex one-shots a playable cyberpunk FPS ('Neon Protocol') from a short prompt; the panel predicts a huge surge in user-generated games and apps now that building software requires only natural language.

Gemini Live voice and real-time translation AI ▶ 24:12

Gemini's voice mode leapfrogs its previous stilted quality (under competitive pressure from OpenAI's voice model) and now handles natural real-time translation, foreshadowing simultaneous-translation AR glasses.

AI agents calling businesses (Duplex revival) AI ▶ 28:47

New Gemini shopping agents can call stores to check stock/pricing, reviving Google Duplex seven years after its 2018 debut; discussion of disclosure norms, non-English callers, and AI enabling 'APIs for everything' via voice.

Gemini 3 benchmark results AI ▶ 32:29

Gemini 3 roughly doubles Claude 4.5 on Humanity's Last Exam and doubles GPT-5.1 on ARC-AGI-2, without appearing benchmaxed; framed as evidence AI is closing in on PhD-level research across math, science, engineering and medicine.

Scaling laws vs. Yann LeCun's anti-LLM stance AI ▶ 39:31

Responding to Yann LeCun's claim that LLMs are the wrong path to AGI, the panel argues scaling laws keep holding with no new paradigm needed, and that even if a new paradigm is required, an AI is more likely to find it than a human.

AI biosecurity and defensive co-scaling Biotech ▶ 46:16

OpenAI-backed Red Queen Bio ($15M investment) claims to have found scaling laws for biological-safety measures; the panel debates open-source AI's bioweapon-uplift risk now that only Chinese labs still release frontier open weights, and proposes airport-style biosensing/rapid sequencing as defense.

Coding tool wars: Cursor vs. Anti-Gravity Compute ▶ 58:52

Cursor tripled its valuation to ~$30B in six months ($2.3B raised) even as Google's Anti-Gravity IDE looks nearly identical; Dave argues Cursor's edge is multi-model access while its own frontier-model ambitions grow.

Project Prometheus: Bezos's physical/world-model AI Robotics ▶ 1:03:15

Jeff Bezos funded Project Prometheus with $6.2B and hired ~100 researchers from OpenAI, Google and Meta to build world models for manufacturing, framed as a pivot of capital from pure LLM 'superintelligence' toward physical AI for factories, robotics and space.

Pathway to abundance: cost deflation across life essentials Economy ▶ 1:12:00

Using 2023 US household spending (housing 33%, transport 17%, food 13%, health 8%, education 2.5%), the panel maps how AI/robotics-driven vertical farming, 3D-printed housing, autonomous EVs and AI diagnostics could push costs toward near-zero.

Predictions made

open Alex Wissner-Gross: Hard research problems in fields like math, science, and medicine will start to fall to AI models at Gemini 3's level.

EP #? · · due: end of 2026 · ▶ watch

“I'd be very surprised if by the end of say next year, we're not seeing hard research problems succumb to AI models like this one.”

Your call:

open Alex Wissner-Gross: Billions of user-generated games will be created because natural-language one-shot game generation is now trivially easy.

EP #? · · due: within a year (~2026) · ▶ watch

“I think we should expect to see billions of games created in the next year because it's now so easy.”

Your call:

open Dave Blundin: Humanoid robots will not exist in sufficient numbers to solve the global housing/manufacturing problem for about 15 years, even though individuals will have them much sooner, due to manufacturing ramp-up constraints.

EP #? · · due: ~2040 · ▶ watch

“we won't have enough of them to attack the global problem for about 15 years because of this, you know, just the manufacturing, you know, rate”

Your call:

open Alex Wissner-Gross / Dave Blundin: The dollar cost per unit of AI intelligence will continue to hyperdeflate by roughly another 40x within the next year.

EP #? · · due: 2026 · ▶ watch

“40x is what you should expect next year. 40x people really struggle 40x in anything.”

Your call:

open Peter Diamandis: Based on prediction-market pricing, Google's current dominant lead in AI (91% odds of being best model by year end) will erode to roughly a coin flip against the rest of the field.

EP #? · · due: next summer (2026) · ▶ watch

“Now it's 91% Google, but by next summer that's down to 60%. So it's sort of like 50/50 that someone else will take a lead by next summer.”

Your call:

open Elon Musk (clip): AI and humanoid robots will eliminate poverty, with Tesla pioneering humanoid robots alongside many other manufacturers.

EP #? · · due: unspecified · ▶ watch

“there is only basically one way to make everyone wealthy, and that is AI and robotics”

Your call:

Numbers that matter

Gemini 3 scores almost 50% on Humanity's Last Exam Cited by Dave Blundin as evidence of a step-function jump in model capability.
Gemini 3 delivers ~3,000% more profit than GPT-5 or Claude Sonnet on Vending-Bench Arena Simulated vending-machine business management benchmark run by Andon Labs.
Internet advertising is a $300 billion/year fully non-human industry Dave Blundin's baseline for estimating the size of the existing 'non-human economy.'
Duolingo down almost 50% over the past year Cited as an early casualty of AI-native translation/language tools like Gemini Live.
Google Duplex launched 7 years ago (2018) Time it took for AI store-calling to go from demo to a fully realized shopping feature in Gemini.
Gemini 3 is a ~7 trillion parameter class model, up from ~1 trillion a year earlier; 10x-40x more raw horsepower expected next year Used to rebut claims that AI progress is leveling off.
US biodefense/biosecurity market was $34 billion in 2024, projected to double by 2034-2035 Context for OpenAI's $15M investment in Red Queen Bio.
An extreme bioattack scenario could cause a multi-trillion dollar global loss Cited alongside the asymmetric risk that a bioweapon could be engineered for roughly $1,000.
Cursor's valuation roughly tripled from ~$10B to ~$30B in six months, raising $2.3B June-to-November 2025 valuation jump amid the AI coding tool race.
Jeff Bezos put $6.2 billion into Project Prometheus and hired nearly 100 researchers from OpenAI, Google, and Meta New physical-AI/manufacturing venture; panel notes this is a 'drop in the bucket' relative to the trillions needed to solve math/science/engineering/medicine.
Dave Blundin's foundation-model build time dropped from 5 years (original) to ~2 months (AI-assisted rebuild) Illustrates AI-accelerated AI development; expects another 5-10x reduction within a year.
2023 US average family expenditure was $77,000/year: 33% housing, 17% transport, 13% food, 12% insurance/pensions, 8% health, 5% entertainment, ~2.5% education Baseline used to argue where AI/robotics-driven cost deflation will hit hardest.
Vertical farming yields ~7x horizontal farming and saves 99% of fresh water; 35 Manhattan skyscrapers converted to vertical farms could feed the entire city Agriculture uses 70% of global fresh water; vertical farming framed as newly economically viable.
Cost per unit of AI intelligence is hyperdeflating at roughly 40x year-over-year Alex Wissner-Gross's headline metric, framed as upstream driver of all abundance milestones.
A comfortable life in Bali costs about $20/day (~$10 housing, ~$2 food); half an Ethereum (~$2,000) in DeFi pools earning ~1%/day covers that cost Salim Ismail's back-of-envelope calculation on crypto-funded low-cost living (explicitly not investment advice).

Worth digging into

🕳️ Vending-Bench Arena benchmark

Used as the episode's central proxy for AI becoming a first-class autonomous economic actor, with a striking ~3,000% profit claim over GPT-5/Claude Sonnet.

🕳️ Red Queen Bio's scaling laws for biological safety

Proposed as a template for 'defensive co-scaling' -- the idea that AI safety measures can scale proportionally with capability, which could generalize far beyond biosecurity.

🕳️ Project Prometheus's actual technical focus

Bezos's $6.2B, ~100-researcher venture is described as thin on public reporting; the panel speculates it may lean more toward biology than manufacturing.

🕳️ US open-source frontier AI shutdown and bioweapon uplift risk

Claim that Meta halting open weights leaves only Chinese labs open-sourcing frontier models, which the panel ties directly to bioweapon-uplift danger from uncontrollable local models.

🕳️ Cost-per-unit-of-intelligence hyperdeflation metric

Alex Wissner-Gross and Peter Diamandis reference an ongoing joint paper tracking this 40x/year figure as the master variable behind every abundance prediction in the episode.

🕳️ Vernor Vinge's 'Rainbow's End' as an AI-biosafety scenario

Recommended mid-discussion as a fictional but detailed depiction of what AI-enabled biosafety infrastructure could look like.