Ex-Google CEO on the Consequences of an A.I. (Eric Schmidt) | EP #7
Recorded in April 2022 at Peter Diamandis's Abundance360 summit (pre-ChatGPT), this conversation with former Google/Alphabet CEO Eric Schmidt ranges from AI-accelerated biology (Broad Institute, Colossal's woolly mammoth project) to quantum computing timelines, the explosive scale of large language models like Google's PaLM, and Schmidt's central worry: that 'compression of time' from AI is outpacing human, legal, and political systems' ability to respond, using hypersonic-missile defense and nuclear-launch doctrine as examples. Schmidt lays out a detailed AI roadmap (multimodal within 5 years, industrial-strength conversational AI within 5-10 years) and makes a specific, dated prediction that AGI will arrive around April 2042, plus a forecast that only five to ten AGI-scale systems will exist, controlled by nation-states. He also argues AI proliferation is fundamentally harder to regulate than nuclear weapons, explicitly says he does not want AI to repeat social media's ungoverned rollout, and frames US-China AI competition as the defining rivalry of his remaining lifetime, predicting rough parity within five years. The episode closes with reflections on working with brilliant, difficult founders like Steve Jobs, Larry Page, and Sergey Brin.
Social media's ungoverned AI-feed rollout as a cautionary taleAI▶ 0:00:00
Schmidt opens (and later revisits at 34:20) with regret that the shift from linear social feeds to engagement-maximizing, outrage-driving AI feeds was a decision made by technologists, not society, and says he does not want AI's rollout to repeat that mistake.
AI as the key to unlocking biology (Broad Institute, digital cell models)Biotech▶ 0:09:59
Schmidt describes funding the Broad Institute because biology needed to go from 'squishy' to 'digital,' notes there was still no digital model of the cell in 2022, and argues AI's strength is estimating non-computable biological functions well enough for human scientists to spot causal mechanisms.
Colossal Biosciences and woolly mammoth de-extinctionBiotech▶ 0:09:52
Brief mention of Ben Lamm (Colossal CEO) partnering with George Church to bring back the woolly mammoth, cited as an example of the squishy-to-digital biology trend.
Quantum computing timeline and SandboxAQCompute▶ 0:12:13
Schmidt, chairman of SandboxAQ (spun out of Google, led by Jack Hidary), explains that fault-tolerant quantum computers are roughly 8-10 years out due to error-correction needs (100-1,000 physical qubits per reliable qubit, cooling to ~0.01 Kelvin), and that quantum simulation on classical computers can already help chemists improve compounds ahead of that.
Large language model scaling and Google's PaLMAI▶ 0:16:34
Schmidt cites Google's newly announced PaLM model as evidence of emergent capability (translating between programming languages without paired training examples) and notes language models are scaling roughly 10x in size every 10 months.
Compression of time and human/institutional decision limitsAI▶ 0:18:44
Schmidt's central worry: AI-driven acceleration is outpacing the speed at which human minds, legal systems, and governments can process decisions, illustrated by a hypersonic-missile defense scenario requiring a 23-second human decision and Cold War nuclear-launch doctrine timings.
Defining AGI and the nation-state AGI scenarioAI▶ 0:28:24
Schmidt defines AGI as 'computer intelligence that looks human but is not human intelligence,' predicts its arrival around 2042 (median of expert estimates), and speculates only five to ten such systems will exist two decades out, each controlled by a nation-state absent a unified world government.
Why AI is harder to regulate than nuclear weaponsGeopolitics▶ 0:35:51
Using Kissinger's Kremlin negotiating stories and the Asilomar recombinant-DNA precedent, Schmidt argues software proliferation defeats nuclear-style regulation: hidden research can't be verified, open-source safety checks get stripped by adversaries, and hardware limits are trivially reproducible.
Schmidt discusses China's publicly stated 2030 AI dominance plan, estimates the US is ahead by only one to two years, predicts rough parity within five years built on differing values, and calls the US-China rivalry the defining competition of his remaining lifetime, framing it as a 'rivalry partnership' where only information-related technology (not steel, farming, plastics) will be fiercely contested.
Working with brilliant, difficult foundersOther▶ 0:41:23
Responding to Diamandis's question about whether success requires becoming desensitized, Schmidt reflects on Steve Jobs's rare charismatic brilliance, Larry Page and Sergey Brin's extreme self-confidence, and his own strategy of finding the smartest, most difficult person in the room and becoming their essential helper.
Predictions made
openEric Schmidt: Real, fault-tolerant quantum computers (the kind that function like a full computer rather than a limited-algorithm device) will arrive.
EP #? · · due: ~2030-2032 (8-10 years from April 2022) · ▶ watch
“real quantum computers of the kind that look like a computer or like half a computer... is probably 8 to 10 years away”
⚖️ Tracking roughly on schedule but not yet arrived: Google's Willow chip (Dec 2024) achieved below-threshold error correction, a key fault-tolerance milestone, and Google itself now targets a useful error-corrected machine 'around the end of the decade' (~2029-2030) -- consistent with Schmidt's window, but no full fault-tolerant general-purpose quantum computer exists as of August 2026.
Your call:
openEric Schmidt: Current digital communications/encryption will become breakable, initially by foreign powers, once sufficiently mature quantum computers exist.
EP #? · · due: ~2030-2032 (8-10 years from April 2022) · ▶ watch
“the simple rule is that in 8 to 10 years all of your digital Communications will be breakable, probably initially by foreign powers”
⚖️ Not yet true as of August 2026 -- no quantum computer has broken standard encryption -- and most current expert estimates (post-2024 error-correction advances) put cryptographically relevant quantum computers around 2030-2035, slightly later than Schmidt's 2030-2032 window; NIST finalized post-quantum cryptography standards in August 2024 partly in anticipation, and 'harvest now, decrypt later' data theft is already an acknowledged active threat.
Your call:
hitEric Schmidt: Industrial-strength AI will be achieved.
EP #? · · due: ~2032 (next 10 years from April 2022) · ▶ watch
“I think it's extremely clear that in the next 10 years we're going to get industrial strength AI”
⚖️ Achieved well ahead of the 10-year deadline: by 2025-2026, AI is embedded in production software (GitHub Copilot, Claude Code, Cursor), enterprise operations (Gartner projected 40% of enterprise apps would feature task-specific AI agents by 2026, up from under 5% in 2025), and industrial/manufacturing/logistics/defense use cases -- genuinely 'industrial strength' deployment, not a lab demo.
hitEric Schmidt: AI systems will become genuinely multimodal, handling text, speech, and video together.
EP #? · · due: ~2027 (next 5 years from April 2022) · ▶ watch
“for the next 5 years you should expect multimodal which means you're going to have text and speech and video”
⚖️ Arrived years ahead of the 5-year (2027) deadline: GPT-4V (Mar 2023) and native-multimodal Gemini (Dec 2023) combined text/image, GPT-4o (May 2024) unified text/audio/video/image in one model, and Google's Veo (May 2024) added high-fidelity text-to-video -- text, speech, and video are now standard in flagship models.
partialEric Schmidt: Usable, explainable, conversational AI systems (where you can ask the system why it made a decision) will become part of everyday global life.
EP #? · · due: ~2027 (next 5 years from April 2022) · ▶ watch
“think of it as the next 5 years industrial strength usable Global conversational systems you can say to the system why did you do this... it will become part of our lives”
⚖️ The 'part of everyday global life' half landed hard -- ChatGPT alone reached hundreds of millions of weekly users by 2025-2026 and conversational AI (Claude, Gemini, Copilot) is mainstream -- but the 'ask the system why it did this' explainability half has not: 2026 AI Safety Index and interpretability research still describe frontier LLMs as unreliable black boxes, with mechanistic interpretability an active, unsolved research frontier rather than a shipped consumer feature.
openEric Schmidt: Artificial general intelligence (computer intelligence that looks human but is not human intelligence) will arrive.
“I predict right now April 2042 is the arrival of AGI and I'm statistically going to be correct when when that occurs”
⚖️ Unjudgeable until the deadline: AGI (by most definitions, including Schmidt's own 'looks human but isn't') has not arrived as of August 2026, though expert median timeline surveys and prediction markets have shortened considerably since 2022 following GPT-4/o1/o3-class reasoning models -- meaning the field-wide consensus Schmidt cited has moved earlier than his own 2042 marker, not later.
Your call:
hitEric Schmidt: The share of software code written by AI (versus humans) -- roughly a third in 2022's Microsoft Codex -- will grow to a much larger percentage.
EP #? · · due: within a few years of 2022 · ▶ watch
“imagine that that number will get a much larger percentage written by computer over the next few years”
⚖️ Confirmed: Microsoft CEO Satya Nadella said in April 2025 that AI generates ~30% of Microsoft's code, Google reported a similar ~25-30% figure, industry-wide estimates put roughly 41% of all new code as AI-generated by early 2026, and Anthropic reports far higher internal figures (spokesperson cited 70-90% company-wide, with some engineers like Claude Code lead Boris Cherny saying 100% of their own code is now AI-written) -- all well above the 2022 one-third Codex baseline.
openEric Schmidt: Only five to ten AGI-scale systems will exist because they remain so computationally expensive, and they will be controlled by nation-states absent a single unified world government.
EP #? · · due: ~2042 (20 years from April 2022) · ▶ watch
“I think there will be five to 10 of these things because these things are so computationally expensive even 20 years from now... they'll be controlled by nation states”
⚖️ Deadline (~2042) is too far out to judge, but the interim trend as of 2026 cuts against it: rather than 5-10 nation-state-controlled systems, the frontier landscape has broadened to at least 8 competing corporate labs (OpenAI, Anthropic, Google DeepMind, xAI, Meta, Mistral, DeepSeek, Moonshot AI), privately/hyperscaler-owned rather than state-run, though increasingly entangled with national export-control and industrial policy.
“China 2030 plan which included a statement in Chinese that they would be dominant by 2030 that they would catch up by 2025”
⚖️ The 2025 checkpoint has passed and shows narrowing, not full catch-up: DeepSeek-R1 (Jan 2025) matched OpenAI's o1 within 0.4% on Chatbot Arena and shocked Silicon Valley, and by mid-2026 the leading Chinese lab trails the US frontier by roughly 6-8 months (down from Schmidt's own 1-2 year estimate in 2022) -- real convergence on specific benchmarks (coding, math) but US labs (Anthropic, OpenAI, Google) still generally lead overall as of August 2026; 2030 dominance remains open/unresolved.
partialEric Schmidt: The US and China will be roughly equal on all significant AI capabilities, though their systems will be built on differing values.
EP #? · · due: ~2027 (next 5 years from April 2022) · ▶ watch
“in the next five years the two countries will be roughly equal on everything interesting but the systems will be built with different values”
⚖️ With the 5-year deadline (2027) about to elapse, the trend strongly supports near-parity but not full equality yet: the US lead shrank from Schmidt's cited 1-2 years (2022) to roughly 6-8 months by 2026 per multiple trackers, and China leads or matches on some tasks (coding, math) while trailing more on novel reasoning (e.g., ARC-AGI-2) -- 'roughly equal on everything interesting' is closer to true than false but not fully realized, and the values divergence (open US models vs. censored/state-aligned Chinese models) did play out as predicted.
hitEric Schmidt: The US-China competition will be the single most important competition faced during the remainder of Schmidt's life, framed as rivalry-plus-partnership rather than war.
EP #? · · due: open-ended (rest of Schmidt's lifetime) · ▶ watch
“a prediction I will make is that the competition between the US and China is the most important competition that we're going to face during the rest of my life”
⚖️ This has clearly played out: US-China AI/tech rivalry has dominated geopolitics since 2022 -- export controls and entity-list bans on chips (Nvidia H20/H100 restrictions), the CHIPS Act, an AI diffusion framework, and the January 2025 'DeepSeek shock' to US markets and policy -- while staying short of open conflict, matching Schmidt's 'rivalry partnership, not war' framing.
Numbers that matter
Schmidt Futures co-founded in 2017Philanthropic vehicle for backing exceptional talent on hard problems.
$1 billion philanthropic commitment announced in 2019Eric and Wendy Schmidt's commitment to identify and support talent globally.
100 to 1,000 physical qubits needed per reliably accurate logical qubitExplains why fault-tolerant quantum computers are still years away -- replication is required to overcome measurement error.
Quantum computers must be cooled to roughly 0.01 KelvinIllustrates the extreme refrigeration required for current quantum computing approaches.
Google's PaLM model was trained using four TPU clusters, costing millions of dollars, over six weeksCited as evidence of the immense cost and scale behind large language model training.
Language models are scaling roughly 10x in size every 10 monthsSchmidt cites this figure to illustrate the extraordinary pace of LLM growth.
A ship's AI gives its captain 23 seconds to decide whether to launch a hypersonic counterattackUsed as a stark example of AI-driven compression of decision time in national security.
A nuclear missile from Russia takes about 30 minutes to reach a US target; Cold War doctrine allowed roughly 3 minutes to wake the president, 2 minutes for cognition, and 5 minutes for conversation before ordering a responseContrasts historical nuclear response timing with the far shorter windows AI-era threats allow.
Fountain Life (sponsor) measures 40 different biomarkers every quarterAd segment describing Diamandis's home-phlebotomy health-monitoring company.
About one-third of code in Microsoft's Codex-assisted development was AI-written versus two-thirds human, as of 2022Used as a baseline for predicting AI's growing share of code authorship.
$125 million research program announced by Eric Schmidt and James ManyikaFunds research into the hard societal/governance problems posed by advanced AI.
China announced its 'China 2030' AI plan roughly two and a half years before this April 2022 conversationChina's plan stated intent to catch up to the US in AI by 2025 and be dominant by 2030.
Schmidt estimates the US AI lead over China at roughly one to two years, as of 2022Based on his work chairing the National Security Commission on Artificial Intelligence.
Worth digging into
🕳️ Schmidt's dated AGI prediction: April 2042
One of the rare specific, dated AGI predictions on record from a major tech leader, made in the pre-ChatGPT era -- ideal for scoring against how AGI-timeline consensus shifted after GPT-4, o1/o3-class reasoning models, and 2024-2026 frontier releases.
🕳️ AI writing an ever-growing share of software code
Schmidt's 2022 baseline (~1/3 AI-written via Microsoft Codex) is a concrete, checkable number against the explosive rise of Copilot, Cursor, and agentic coding tools by 2025-2026.
🕳️ US-China AI parity within five years (by ~2027)
Directly testable against the real trajectory of Chinese frontier labs (e.g., DeepSeek, Qwen, Moonshot AI) versus US labs since 2022, and against Schmidt's own 1-2 year gap estimate at the time.
🕳️ 'Five to ten AGI systems controlled by nation-states'
This oligopoly/nation-state-control scenario contrasts sharply with what actually emerged by 2023-2026: a competitive multi-company landscape (OpenAI, Google, Anthropic, Meta, xAI, plus multiple Chinese labs) rather than a handful of state-controlled systems.
🕳️ Quantum computing '8 to 10 years away' (~2030-2032)
A specific, falsifiable technical timeline made before major 2023-2025 error-correction milestones (e.g., Google's Willow chip); worth checking against actual progress toward fault-tolerant quantum computing.
🕳️ Whether AI got its own 'Asilomar moment'
Schmidt explicitly said he didn't want AI to repeat social media's ungoverned, technologist-only rollout, and pointed to the 1975 Asilomar recombinant-DNA conference as the model to emulate.