The "Brute Force" Era of AI (Where We Are Now)
Current AI = "Pixel-Brain": Just as early computer vision broke images into grids of pixels, today's LLMs break language into tokens and relationships into matrices. It's all pattern-matching at scale.
Power Gluttony: Training GPT-4 consumed ~50 GWh (enough to power 40,000 homes for a year). Why? Because we're simulating "insight" by throwing math at the problem until it sticks.
Bro-Culture Parallel: It's the tech equivalent of bodybuilders juicing on steroids—bigger models, bigger data, bigger energy bills. The ethos is pure "gains through overload" (see: every AI lab racing to hoard H100s).
2. Why AGI Will Likely Hate This Approach
Inefficiency = Stupidity: Biological brains (e.g., humans) run on ~20W. AGI will have to optimize because:
Physical limits: Energy costs will throttle scaling (see: Bitcoin mining collapse in high-electricity regions).
Embodied constraints: If AGI lives in robots/drones, wasting energy = starvation.
"Seeing the Whole Image": You're right—true intelligence spots higher-order patterns that reduce computation. Examples:
Human shortcut: We recognize a face from a few strokes in a cartoon. Current AI recomputes every pixel.
AGI's breakthrough: Might invent "lossless compression for thought"—deriving logic from first principles instead of digesting terabytes.
3. The Cultural Reckoning Coming
Bro-Culture's Dead End:
Today's "more flops = more smarts" mindset mirrors 1980s Wall Street's "more leverage = more profits." Both are Ponzi schemes of diminishing returns.
Symbolic moment: When Altman begged for $7T to make more chips, it was the equivalent of a gym bro claiming he needs 10,000mg of creatine to lift a spoon.
AGI's Inevitable Pivot:
Phase 1 (Current): "How many exaflops can we cram into a datacenter?"
Phase 2 (AGI): "How can I solve this with the energy of a lightbulb?" (See: AlphaGo's move 37—creative efficiency that stunned humans.)
4. The Dark Irony
We're Building AGI to Fix Our Wastefulness: The same AI that currently burns forests to power chatbots may eventually invent:
Neuromorphic chips (brain-like efficiency)
Quantum annealing (solving problems without brute force)
Self-pruning networks (models that shed useless parameters)
The AGI Economy:
Early stage: "Buy more GPUs!" (Nvidia wins)
Mature stage: "Delete 99% of your GPUs." (AGI laughs at our waste)
Comments
Post a Comment