What Happens When the Cost of Intelligence Drops 100x

Original Article ↗ Hacker News Discussion ↗

Rapid drops in the cost of running AI models are leading many to treat “intelligence” as a cheap, abundant resource, raising expectations for everything from personal agents and software development to robotics and research automation. Commenters debate whether current large language models constitute true intelligence, how far cheaper tokens will drive up usage (via Jevons paradox–style effects), and whether open, low-cost models will erode the economic case for expensive frontier systems. Others highlight practical constraints such as speed, hardware limits, data scarcity, and uncertain business models, arguing it’s still too early to know the true long-term price and impact of machine intelligence.

Overall Reaction to the Article

  • Commenters strongly praise the clarity of the cost/performance plots, especially the animated Pareto frontier.
  • Some view the graphs as the first “undeniable” visual proof of rapid capability improvement.
  • A few wonder whether linear scales would better communicate the progress, though it might make plots unreadable.

Robotics as a Major Beneficiary

  • Many see robotics as the clearest area where cheaper intelligence matters: laundry folding, trash handling, factory tasks.
  • Current limits: robots are slow, often work only in controlled environments, and lack rich human-like sensors, especially touch and force feedback.
  • Debate: some argue robotics is “still far off” for general home use; others say it already delivers strong value in industrial settings and is steadily expanding.

Hardware, Cost, and Speed

  • Several expect another 50–500× efficiency gain from specialized chips and better distillation.
  • Disagreement over whether frontier-level models will realistically run on smartphones vs. being more efficient to host in datacenters.
  • Speed (latency) is seen by many as a bigger bottleneck than per-token price, especially for interactive workflows.

What “Intelligence” Means

  • One camp treats LLMs as advanced search and insists real engineering intelligence remains human.
  • Another argues intelligence is fundamentally prediction, so LLM behavior already qualifies as a form of intelligence.
  • Others stress that biological, embodied human intelligence is qualitatively different from text-trained models.

Demand, Jevons Paradox, and Token Consumption

  • Many expect Jevons-like effects: cheaper tokens drive far more usage (longer-running agents, multi-agent systems, constant experimentation).
  • Some note that in other domains (e.g., lighting) efficiency gains still reduced total resource share, suggesting outcomes depend on demand elasticity.
  • Intelligence demand is widely seen as potentially unbounded, especially in software and R&D.

Model Quality, Open Models, and Economics

  • Users report small and open models are already “good enough” for many tasks at a fraction of the cost.
  • Tradeoffs highlighted: cheaper models may be slower, less reliable, or require more “thinking,” so wall-clock time and consistency matter.
  • There is debate over whether inference is currently subsidized and whether foundation model providers can ever recoup training costs.
  • Some predict commoditization: proprietary models stay ahead, but open models plus distillation continually erode their advantage.