Unique Research · Insight / Long-form transcript digest · August 19, 2024 · Source: Stanford classroom interview with Eric Schmidt (Aug 2024); CN edit via Web3 Sky City (authorized reprint). Note: Schmidt’s long DoD advisory role — read stance critically.
Editor’s frame
Schmidt on near-term AI breakthroughs, U.S.–China tech competition, hyperscaler monopolies, labor/privacy/antitrust/national security, and adversarial AI. He forecasts a 1–2 year wave combining huge context windows, agents, and text-to-action — impact potentially beyond social media. CN Bilibili full video noted; not embedded here.
Three near-term stacks
Classroom warm-up: million-token context (Anthropic ~200k on path to 1M; OpenAI similar aim); agents as LLM + state + memory that act (browse, look up); text-to-action as triggering ops (often via Python — Mojo mentioned as challenger). On NVIDIA’s ~$2T: CUDA since 2008 as “C for GPUs” plus CUDA-optimized libs (e.g. vLLM-class stacks) hard to clone without a decade of software.
Why the combo matters: long context ≈ short-term memory / freshness (models are ~18 months stale after prep+train+finetune); agents that hypothesize, lab-test overnight, and learn (e.g. chemistry “ChatCrow”-style); text-to-action as infinite compliant programmers — e.g. “clone a TikTok with my prefs in 30 seconds” or “build a Google competitor with genAI search in 30 seconds.” Incumbents may be more attackable than assumed.
Frontier gap, Stargate money, power
Six months ago Schmidt thought the frontier–rest gap was closing (bet small cos); now less sure — gap may widen among ~three frontier labs. Capex asks of $10B / $20B / $50B / $100B; Stargate ~$100B hard; Altman floats ~$300B+. Energy becomes scarce — he told the White House to be best friends with Canada (hydro + AI history) vs. Gulf funding that may not match U.S. security rules. If $300B flows to NVIDIA, markets will notice (not a stock tip).
On CUDA alternatives: AMD’s ROCm translation still imperfect; with infinite money he’d pick NVIDIA B200 today. Google’s work-life vs. startup grind; founders matter; TSMC basement-for-new-PhDs culture contrast; network-effect markets punish slow deals (telecom 18-month closes). Microsoft–OpenAI once looked “crazy outsourcing” — now Microsoft challenges Apple on value; Apple still thin on AI.
U.S.–China & policy
As AI commission chair (~752-page report): stay ahead with money; CHIPS Act etc. Few countries can play frontier+open-source race — capital, talent, education, will. Export controls on NVIDIA to China; ~10-year edge on sub-DUV / sub-5nm. Informal industry group helped argue for Biden’s long AI executive order. Reporting threshold idea: ~10^26 FLOPs training runs notify government (EU ~10^25); federated/composed training may erode bright lines.
Knowledge, adversarial AI, alignment
With Kissinger & Huttenlocher on opaque models: like parenting teens — know bounds even if not full internals. Expect an industry of adversarial AI (paid red-team cos/systems). Hallucinations shrink but don’t vanish; need efficacy tests; chain-of-thought to ~1000 steps you can run and check. Progress drivers: compute + data + post-Transformer algorithms + market belief in infinite returns to intelligence — bubble risk acknowledged.
Open vs closed: Mistral strong; next models may close for revenue. Capex may rewrite software culture that was built on open source. Coding copilots aim ≥2× productivity; Augment-class tools for 100-person / multi-million-LOC codebases. Labor: high-skill + education win; dangerous/judgment-light work at risk. Academia needs compute credits, not only cloud coupons. Still learn to code — “why learn English if you speak English?” Distributed MacBook folding unlikely to beat memory-interconnect supercomputers soon. Copyright → ASCAP/EMI-like revenue shares. Antitrust: trend not to break up unless Rockefeller-scale. Non-frontier countries: rich get richer; need partners.
Advice: prototype with these tools in a day or rethink — hackathon drone-path winners beat week-long human coding; let the computer draft the business plan if legal, then ship a working demo before someone else does.
— END —