- Intelligence requires feedback: AI leaders conflate intelligence with power/output, but intelligence needs input—faster/clearer feedback = more powerful intelligence. Applies to math, coding, weather forecasting (vast training data, rapid feedback, clear right/wrong answers). Does NOT apply to writing (“deluge of AI slop on LinkedIn”) or cancer curing (living systems take years to respond). OpenAI’s Navier-Stokes claim (solved in 88 hours vs decades for humanity) validates curve in math-like domains. But Altman’s prediction AI “will cure all diseases” confuses intelligence with power. Anthropic’s Claude writing 80%+ of code engineers merge is real; pushing to 100% (building own successors) still limited by feedback constraints and infrastructure needs.
- Historical precedent: Intelligence alone insufficient. Los Alamos (brilliant physicists) accounted for 4% of Manhattan Project cost; 80% went to uranium/plutonium production plants (infrastructure). Oppenheimer 1945 testimony: “Without scientists no bomb, but without plants also no bomb.” Wright brothers beat better-funded Samuel Langley not through brilliance but through understanding feedback loops (bikes unstable, corrected by rider; wind tunnel tests 200+ wing shapes vs Langley’s over-confident scaling). Pattern: constant corrections keep models moored to reality.
- Feedback determines application speed: Intelligence gets used first where feedback fastest/clearest (math ✓, coding ✓, weather ✓, finance/hedge funds ✓). Writing fails tests (unclear feedback). Cancer curing fails (years-long experiment cycles). As you know more, learning harder (theoretical physics: progress leapt after quantum revolution, then slowed despite better tools—Dirac 1975: “glorious time” for second-rate physicists doing first-rate work; now first-rate physicists struggle for second-rate work). Semiconductors: doubling chip density requires 18x more researchers now vs 1970s. Forethought model gives 60% odds AI research compression possible, but assumes returns fall as software approaches limits. Amdahl’s Law: speeding one part moves bottlenecks, doesn’t eliminate them.
- Infrastructure constraints real: Altman’s exponential curve limited beyond math/coding. $760B combined capex by AMZN/GOOGL/META/MSFT validates that intelligence advantage ≠ army, power grid, supply chain. Can’t build overnight. Amodei’s own 2024 essay “Machines of Loving Grace” noted particle physicists have theories but lack accelerator data (superintelligence won’t solve that). Skepticism warranted on 50-100 year biological progress compression claims. Altman’s curve “real in mathematics. Maybe not far beyond it.”
What Happened?
Gautam Mukunda (Yale leadership professor, Bloomberg Opinion) argues AI leaders misunderstand intelligence vs power/output. OpenAI solved Navier-Stokes Millennium Prize in 88 hours (validates math curve). Anthropic’s Claude writes 80%+ of code engineers merge. But AI power limited by feedback speed/clarity, not intelligence alone. Math ✓, coding ✓, weather ✓ (fast/clear feedback, vast training data). Writing ✗, cancer ✗ (slow/unclear feedback). Los Alamos was 4% of Manhattan Project cost (80% infrastructure/plants). Wright brothers beat Langley through feedback loops understanding (bikes=constant corrections). Historical pattern: intelligence plateaus as domains mature (theoretical physics slowed post-quantum revolution; semiconductors require 18x more researchers now vs 1970s). Amdahl’s Law: speeding one process moves bottlenecks. Anthropic’s Amodei himself noted constraints in 2024 essay. $760B Big 4 capex validates infrastructure bottlenecks—intelligence alone can’t build armies/grids/supply chains overnight. Skepticism warranted on Altman’s 50-100 year disease-curing compression.
Why It Matters?
For AI equity investors (GOOGL, AMZN, META, MSFT), Mukunda’s thesis validates caution on exponential AI valuation curves. Feedback bottlenecks limit upside beyond math/coding. For Anthropic investors, Amodei’s more measured predictions (vs Altman’s bolder claims) suggest realistic constraints built into strategy. For macro observers, $760B capex spend validates infrastructure limitations—scale alone insufficient. For researchers, feedback-dependent progress suggests biology/medicine less amenable to pure AI acceleration than software/math. For policy, skepticism on “superintelligence” risk scenarios (feedback-limited domains won’t enable unconstrained power).
What’s Next?
Monitor OpenAI/Altman’s future predictions vs realized outcomes (disease curing timeline, AGI claims); if diverge significantly, validates feedback bottleneck thesis. Track Big 4 capex ROI; if returns diminish despite increased spending, validates Amdahl’s Law/bottleneck thesis. Watch AI performance in slow-feedback domains (cancer, climate, drug discovery); if progress plateaus despite AI investment, it validates domain-specific limitations. Monitor semiconductor research productivity; if continues declining despite AI tools, validates historical pattern. Also track Anthropic’s approach vs OpenAI’s on constraints; if Anthropic outperforms, it validates realistic assessment. Finally, watch for “AI hype” deflation in markets; if valuations reset, it validates Mukunda’s skepticism on exponential curves.
Affected Tickers & Coins: GOOGL, AMZN, META, MSFT
Source: Bloomberg Opinion (Gautam Mukunda)
