2024-05-02

Ex-Google CEO on Government AI Policy & Deepfakes w/ Eric Schmidt | EP #99

Recorded on stage after a multi-day 'great AI debate' event, Peter Diamandis interviews former Google CEO Eric Schmidt on where AI is headed next: the shift from language-generating AI to action-taking 'intentional AI', the thresholds that would signal dangerous recursive self-improvement, and the current US/UK/EU/China regulatory landscape. Schmidt covers US-China chip competition, deepfakes and 2024 election misinformation risk, how ubiquitous cheap drones are already ending tank warfare in Ukraine, and his prediction that a small number of heavily-regulated closed frontier models will coexist with many open-source 'middle-size' models. He closes on AI accelerating physics/chemistry/biology research via diffusion models and on Sandbox AQ's quantum-simulation work on drug design.

▶ Watch on YouTube

Topics

Language-to-action AI ('intentional AI') AI ▶ 0:12
Schmidt argues AI is shifting from language-to-language generation to language-to-action: plain verbal commands will be compiled into working programs that execute real tasks (e.g., running an entire conference's logistics), producing an explosion in per-person 'digital power'.
Recursive self-improvement and safety thresholds AI ▶ 7:04
Schmidt describes working with a ~20-person scientist group that judges current LLMs safe for now but identifies clear future danger thresholds: recursive self-improvement, agentic systems inventing their own private language, and AI performing advanced math autonomously.
US AI regulation and global policy landscape Geopolitics ▶ 9:11
Schmidt summarizes his involvement in the UK AI Act, the White House executive order, and new US-China track-two AI dialogues; notes the US approach is light-touch notification-based (10^26 FLOPs threshold) rather than mandatory disclosure, while Europe is 'its usual hopeless self.'
US-China AI and chip competition Geopolitics ▶ 10:14
Export controls on ASML and Nvidia H100/H800 chips have capped China at roughly A100-level/7nm hardware versus the West's 3nm-to-1.4nm roadmap; Schmidt expects China to close the gap by spending far more money on training rather than through better hardware.
Social media, deepfakes, and 2024 election risk AI ▶ 10:56
Schmidt discusses how most voters now get information from YouTube, Instagram, Twitter/X, Facebook, and TikTok (which he calls 'really television'), and argues platforms that fail to police election misinformation will trigger heavier regulation; he also describes deepfakes (citing the Taylor Swift incident) as a 'lock makers vs lock pickers' arms race that safety systems are currently losing.
AI and the changing nature of war (Ukraine drones) Geopolitics ▶ 17:36
Drawing on personal visits to the Ukraine front, Schmidt describes cheap ($5,000) drones overwhelming $5 million tanks, turning frontline areas into 24-hour drone 'death zones' and making him believe tanks, artillery, and mortars are becoming obsolete as weapons of war.
Open vs. closed AI models and compute economics Compute ▶ 21:32
Schmidt lays out the open-vs-closed model debate (open models like Llama 3 reaching ~80% of closed-model capability), rising training costs ($250M-$500M+ per run), and his prediction of a small number of heavily regulated closed AGI systems alongside many open-source 'middle-size' models.
AI red-teaming and guardrails vs. curated training data AI ▶ 23:58
Schmidt says attempts to selectively remove 'bad' data from training sets make models more brittle, not safer; the better current approach is training on everything and layering guardrails and red-teaming afterward, which he expects to become its own industry.
Industry self-regulation modeled on biotech's Asilomar era AI ▶ 29:34
Drawing on his own 1980s genetic-engineering lab experience, Schmidt compares today's AI safety meetings (a December meeting, a NeurIPS/AAAI-adjacent gathering, an upcoming Stanford meeting) to the Asilomar biotech conferences that let scientists self-regulate recombinant DNA before government (via the RAC, later folded into HHS) took over oversight.
AI accelerating physics, chemistry, and biology research AI ▶ 26:59
Schmidt describes attending physics/chemistry conferences where diffusion models and LLM variants are used to generate computationally tractable approximations to otherwise-incomputable physics/chemistry equations, with biology flagged as the field with the biggest expected upside given how vast and unmapped it is.
Quantum simulation and Sandbox AQ Compute ▶ 33:52
As Sandbox AQ chairman, Schmidt explains the company sidesteps the unsolved quantum error-correction problem by building classical simulations of quantum effects, which are already good enough to perturb drug molecules for better efficacy and shelf life — a near-term win that predates real quantum computers.

Predictions made

open Eric Schmidt: The world will change very quickly as AI shifts from language generation to taking real-world action ('intentional AI').
EP #? · · due: unspecified (near-term) · ▶ watch
“we're going to have a very different world and it's going to happen very quickly”
Your call:
open Eric Schmidt: His second book with Henry Kissinger, 'Genesis,' will be published later in 2024.
EP #? · · due: 2024 · ▶ watch
“Genesis coming out later this year”
Your call:
open Eric Schmidt: The dangerous recursive self-improvement threshold in AI will take longer than 5 years to arrive, contrary to some industry voices.
EP #? · · due: beyond 5 years from 2024 · ▶ watch
“there's a debate in the industry some people think five years I think it's going to be longer”
Your call:
open Eric Schmidt: The AI hardware gap between the US and China will keep increasing as US chip nodes advance to 2nm and 1.4nm while China stays capped near 7nm.
EP #? · · due: unspecified · ▶ watch
“it looks like the gap hardware gap is going to increase”
Your call:
open Eric Schmidt: China will overcome its hardware disadvantage in AI training by spending roughly five times more money than the US does per training run.
EP #? · · due: unspecified · ▶ watch
“can they pull it off absolutely how will they do it they'll spend more money”
Your call:
open Eric Schmidt: Social media platforms will face regulation in proportion to how badly they mishandle election-related misinformation around the 2024 US election.
EP #? · · due: 2024 · ▶ watch
“you should expect regulation of content because we regulate every country regulates television in one form or another for precisely this issue of election interference”
Your call:
open Eric Schmidt: Once countries build sufficient drone-based defenses, ubiquitous cheap drones will make tanks, artillery, and mortars obsolete and make invading a neighboring country effectively impossible.
EP #? · · due: unspecified · ▶ watch
“the ubiquity of drones means in my view that tanks and artillery and mortars go away as weapons of war”
Your call:
open Eric Schmidt: If the US approves the Ukraine aid package, it will buy Ukraine roughly one more year of runway for asymmetric drone-warfare innovation against Russia.
EP #? · · due: within about a year of the aid package (2024-2025) · ▶ watch
“my current phrase publicly is let's get another year here”
Your call:
open Eric Schmidt: The AI ecosystem will settle into a small number of extremely powerful, heavily regulated closed AGI systems alongside a much larger number of open-source 'middle-size' models.
EP #? · · due: unspecified · ▶ watch
“there'll be a small number incredibly powerful AGI systems which will be heavily regulated because they're so powerful ... and then be a much larger number of what I'm going to call middle size models which will be open source”
Your call:
open Eric Schmidt: AI red-teaming will grow into its own standalone business/industry.
EP #? · · due: unspecified · ▶ watch
“the consensus of the groups that I have been working with is that the red teaming will become its own business”
Your call:
open Eric Schmidt: The most powerful frontier AI models will ultimately be regulated because they are too capable and carry both enormous harm potential and enormous upside.
EP #? · · due: unspecified · ▶ watch
“these very large models are ultimately going to get regulated and the reason is they're just too powerful”
Your call:

Numbers that matter

Worth digging into

🕳️ Recursive self-improvement thresholds and detection
Schmidt describes concrete threshold signals (agents inventing private languages, autonomous advanced math) but no public methodology for detecting them exists yet.
🕳️ US executive order's 10^26 FLOP notification threshold
Schmidt calls it 'an arbitrary measure that we frankly just invented,' worth checking against the actual EO text and how it's being enforced/updated since 2024.
🕳️ Drone-cost economics ending tank warfare
Schmidt's $5K-drone-vs-$5M-tank claim is a strong, checkable structural claim about the future of land warfare that can be tracked against battlefield data.
🕳️ Chinese training runs starting from open-source releases
Schmidt claims 'every Chinese training run starts with an open-source event' -- a specific, checkable claim about Chinese frontier lab practices (relevant post-DeepSeek).
🕳️ Sandbox AQ's quantum-simulation drug design
Schmidt teases a specific claim -- quantum-effect simulation (without a real quantum computer) already improving drug efficacy/shelf-life -- as a live commercial result worth verifying.
🕳️ Asilomar-style AI self-regulation meetings (Dec 2023/2024 meeting series)
Schmidt references a December meeting, a AAAI-adjacent meeting, and a planned Stanford meeting explicitly modeled on the 1975 Asilomar biotech conference -- a concrete real-world governance effort to trace.