AI infrastructure gains shift the bottleneck back to ideas over execution

A growing number of top AI researchers in industry believe their own jobs will be automated within a few years. The reason is not a sudden leap in model intelligence, but a massive acceleration in the engineering and infrastructure that supports AI development. This shift will make experimentation far easier, though it may not change the fundamental nature of today's models.

The core prediction is that models will become superhuman distributed GPU engineers, optimizing the entire training and inference stack. This will trigger a sharp decline in the effective cost of model intelligence, potentially dropping near-exponentially. For professionals across research, IT, and strategy roles, the bottleneck is moving from execution to ideas.

The engineering acceleration and Jevons paradox

Many parts of the AI stack are highly verifiable. Training speed is measured in tokens per second per GPU. Inference efficiency comes down to cost per answer and FLOPs per token. These metrics are optimizable through established sub-problems and architecture trade-offs. AI agents are expected to handle this end-to-end optimization within a few years, pushing inference capabilities close to the physical limits of accelerators like GPUs.

Companies have already captured giant efficiency gains, sometimes shaving 10-30% off the cost to serve a model after its initial pricing. As this compounds, demand for agentic models will only increase-a classic Jevons paradox. The industry remains bottlenecked by figuring out better ways to orient and deliver agents, not by raw model performance. Meta's Muse agent is an early indicator of experiences built around understanding how agents work.

Pretraining research and the value of good ideas

Research itself is becoming easier with coding agents, signaling an era where good ideas outweigh good execution in software. A reasonable prediction is that pretraining research-at least in architecture and data selection for current model classes-will be automated in two to three years. The flexibility of GPUs will support this exploration phase before longer-term co-design of accelerators and models delivers further orders of magnitude in efficiency.

This echoes an earlier transition before deep learning took off, when AI was far more of a research endeavor. Today's best researchers are judged by their ability to implement and scale ideas in complex infrastructure. That engineering barrier is about to drop. Some call this recursive self-improvement, but a more grounded description is "parallelized, AI-assisted language modeling."

Low-hanging fruit in RL and scientific discovery

Another industrial-scale opportunity sits in the quality of reinforcement learning environments. Multiple RL data companies have crossed $100 million or $1 billion in revenue, yet researchers widely agree that much of what they buy is low-quality. The leading labs still see clear return on investment from purchasing the data, and the quality problems are clearly fixable.

In the broader scientific literature, AI models are already superhuman at crawling research and making connections across sparse networks of scientists who rarely interact. This could herald a new era of discovery-potentially including cures for most cancers-or simply accelerate the arc science was already on. The line between those outcomes is fine.

Why this matters for executives and strategy

The coming efficiency wave will be a massive boon for diffusing AI into the economy, even without economically valuable superhuman traits outside math and coding. Near-term progress relies on scaling inference-time compute with current tools, not dramatic step-changes in model capability. For leaders planning budgets and timelines, the signal is clear: the cost of intelligence will keep falling fast, and the strategic advantage will shift to those who can identify and act on good ideas before engineering constraints vanish.