2025-05-22

The AI Reset Is Here: Search, Jobs, and Everything Else w/ Anish Acharya, Dave Blundin, Salim Ismail

Peter, Dave, and Salim host a16z general partner Anish Acharya for a whirlwind recap of a single AI news week: Google I/O, GPT-5 rumors, Grok 3.5, and Nvidia's NVLink buildout all landing on top of each other. Anish frames AI as 'the most human technology ever built' and walks through why consumer AI moats hold up on adaptive/cultural products but not on static systems-of-record work, while the group debates chip sovereignty, Google's search-cannibalization dilemma, and Apple's stalled AI effort. The back half turns to a 2025-2045 future-history slide covering the end of white-collar work, robotaxis, brain-computer interfaces, and a coming wave of stablecoin-funded AI agent microtransactions.

▶ Watch on YouTube

Topics

Abundance as "the most human technology" Other ▶ 3:06

Anish Acharya argues market economies and technology are history's two biggest drivers of human flourishing, and that AI is the first technology to serve humanity's emotional/subjective side rather than just extending the intellect.

AI unlocking access for seniors AI ▶ 6:16

Dave Blundin's 80-something mother has a full conversation with ChatGPT's voice mode about her Ohio hometown; Anish argues seniors stand to benefit disproportionately from voice AI because it removes the interface barriers that locked them out of prior tech waves.

AI companionship and the challenge tradeoff AI ▶ 7:26

The group discusses AI companionship replacing parts of human relationships, and Anish's escape-room story about being tempted to ask ChatGPT for the answer illustrates the tension between abundance and the human need to struggle toward a goal.

Startup moats: adaptive vs. static AI systems AI ▶ 13:59

Anish argues moats built on adaptive, cultural systems (taste, network effects) hold up against AI commoditization, while moats built on static systems-of-record work are at serious risk, since LLMs predict averages well but can't invent genuinely new cultural shifts.

Consumer AI pricing rivaling rent and food Economy ▶ 20:26

ChatGPT's top tier is $200/month and Google just announced $250/month, dwarfing Spotify's $20 ceiling; Anish predicts software joins food and rent as a top-three household spending category.

The global chip and fab sovereignty race Geopolitics ▶ 29:57

India plans domestic chips by 2025 and its own GPUs within 3-5 years; Dave argues every serious country now needs its own fab strategy because GPU demand so outstrips supply that compute-poor nations will be permanently priced out.

Sovereign AI and a "Jones Act for AI" Geopolitics ▶ 35:13

Anish argues AI embeds the values of whoever trains it (citing DeepSeek vs. Western models) and predicts sensitive national use cases will eventually require domestically-trained AI, analogous to the Jones Act for US-flagged ships.

GPT-5 expectations and the reasoning-model wave AI ▶ 36:55

The panel debates what GPT-5 will actually add; Anish expects incremental multimodal gains (sharper image/video understanding) rather than a single dramatic leap, since reasoning models still need domain-specific correctness data built out field by field.

Grok 3.5's "truth-seeking" framing and bias risk AI ▶ 46:02

Elon Musk unveils Grok 3.5 as a physics-first, maximally truth-seeking model; the panel is skeptical that any model is bias-free, noting truth is contested across countries and models will fill training-data gaps differently.

Nvidia's NVLink spine and the Colossus 2 buildout Compute ▶ 51:56

An NVLink spine linking 72 GPUs moves more data per second than the entire internet's peak traffic; xAI's Colossus 2 adds 170 Tesla Megapacks to smooth the violent power spikes caused by massive synchronized training runs.

Google I/O and the threat to search's ad monopoly AI ▶ 55:59

Anish calls Google's AI Mode underwhelming ("a watered-down Perplexity") but notes VO video generation and the $250/month Ultra tier are real strengths; Google's stock dropped because any search-cannibalizing AI product threatens its 20-year ad-ecosystem commitments.

Apple's AI stumble and the visionary-integrator model AI ▶ 59:57

Siri's overhaul has repeatedly missed deadlines despite years of AI hires; Dave argues Apple and Google both lack a clear visionary-integrator pairing (like Jobs/Cook) and that founder-led, AI-first companies will out-execute committee-run incumbents.

The 2025-2045 future timeline: end of white-collar work to a robot economy Economy ▶ 1:06:10

A speculative future-history slide runs agentic AI (2025-26) through vertical AI and physics breakthroughs (2027-28) to the end of white-collar work (2029-31) and a trillion-dollar robot economy with biolongevity cures (2036-45); Dave warns CEOs are badly under-planning for how fast the disruption hits.

Robotaxis reshaping driving as a job and an experience Transportation ▶ 1:15:40

ARK Invest projects $34 trillion in robotaxi enterprise value by 2030; Waymo's 300 SF vehicles already out-ride 45,000 Lyft drivers, and the panel notes "driver" is the single most common job title on the planet.

Predictions made

open Peter Diamandis: The US will pass a stablecoin bill in 2025.

· due: 2025 · ▶ watch

“a 95% probability that the US will pass a stable coin bill in 2025”

Your call:

open Dave Blundin: Mass white-collar job dislocation from AI will hit hardest in 2026-2027, well before the commonly cited 2029-2031 'end of white collar work' endpoint.

· due: 2027 · ▶ watch

“the amount of job dislocation... in 2026, 2027 is going to be like nothing we've ever seen”

Your call:

open Dave Blundin: Robotaxis will be reliable enough to autonomously drive kids to practice within two to three years.

· due: 2027-2028 · ▶ watch

“it's within two, three years”

Your call:

open Dave Blundin: Ray Kurzweil's 30-year-old prediction of artificial superintelligence by 2029 will land almost exactly on schedule.

· due: 2029 · ▶ watch

“2029 artificial super intelligence predicted 30 years ago. It's going to land like on the exact moment.”

Your call:

open Dave Blundin: Brain-computer interfaces connecting broadly to the neocortex will become mainstream in the early 2030s.

· due: early 2030s · ▶ watch

“Ray predicted BCI our ability to connect everything to the neocortex in the early 2030s”

Your call:

open Dave Blundin: A trillion-dollar robot economy spanning robots of every shape and size, alongside biolongevity cures, will arrive between 2036 and 2045.

· due: 2036-2045 · ▶ watch

“it's trillion robot economy... we unleash human longevity”

Your call:

open Cathie Wood (cited): Robotaxis will generate $34 trillion in enterprise value by 2030.

· due: 2030 · ▶ watch

“Arc Invest... projects a $34 trillion enterprise value from robo taxis by 2030”

Your call:

open Dave Blundin: Agent-to-agent microtransactions enabled by stablecoins will become one of the largest segments of the economy by 2040.

· due: 2040 · ▶ watch

“that's going to be probably the biggest part of the economy by 2040 is those transactions”

Your call:

Numbers that matter

ChatGPT's top tier is $200/month and Google just announced $250/month, vs. Spotify's priciest plan at $20/month Illustrates how far consumer AI pricing has already exceeded every prior consumer software category.
The world will produce roughly 20 million new GPUs this year Dave Blundin notes that's still nowhere near enough to cover even basic AI call-center demand, let alone broader use cases.
Nvidia's NVLink spine moves 130 terabytes/sec across 72 GPUs - about 16% more than the internet's entire peak traffic Used to illustrate the sheer scale of interconnect bandwidth now required inside a single AI data center rack.
xAI's Colossus 2 cluster is scaling toward 200,000 GPUs (mostly H100s) in its next wave, with talk of a million-GPU class after that Context for xAI's 170 Tesla Megapack deployment to smooth training-run power spikes.
ARK Invest projects $34 trillion in enterprise value from robotaxis by 2030 Cathie Wood's firm's forecast, cited while discussing Waymo and Tesla's Cybercab.
300 Waymo vehicles in San Francisco now complete more rides than 45,000 Lyft drivers; each Waymo does the workload of about 150 human drivers running 24/7 Evidence for how fast robotaxis are displacing human driving at scale.
Google Trends shows Steve Jobs still gets roughly twice the search volume of Tim Cook, 14 years after Jobs's death Used to argue Apple lacks a modern 'visionary' figure the way Elon Musk or the late Steve Jobs played that role.
Bitcoin is trading near $106,500, and more Americans now own Bitcoin than gold Cited as evidence Bitcoin is becoming a mainstream American reserve asset.

Worth digging into

🕳️ A "Jones Act for AI" - sovereign training requirements

Anish's analogy to US maritime law predicts a coming wave of national AI-training mandates for sensitive use cases, which would reshape global AI competition and reinforce the chip/fab sovereignty race.

🕳️ Disposable, single-use AI-generated content

Dave's 'spawn it, use it once, throw it away' example - a 12-cent custom theme song generated for one podcast episode - implies an entire new category of ultra-cheap, compute-intensive, throwaway creation that could reshape inference demand and content economics.

🕳️ "Digital makeup" - real-time AI identity layers on video

Mentioned in passing as a real a16z-adjacent portfolio category (changing your appearance live on a call), hinting at a new consumer vertical around real-time AI identity and avatar control beyond novelty filters.

🕳️ Adaptive vs. static-system moats as a VC screening framework

Anish's hip-hop thought experiment - would a model trained on all music before hip-hop ever invent hip-hop - is a sharp, testable frame for which AI-era startups keep durable moats versus which get commoditized.

🕳️ The emerging $4 billion "right-sized" chip fab model

Dave references a new class of fabs priced around $4 billion versus the usual $20-40 billion, which could unclog the global chip shortage far faster than mega-fabs - a concrete, underreported lever in the sovereignty race.