I expect rapid progress but not towards general superintelligence

Introduction

The pace of AI progress is accelerating, but the driver of that acceleration is largely engineering—better hardware, more efficient inference, and AI‑assisted tooling—rather than a fundamental leap toward general superintelligence. This distinction matters for expectations about what AI will achieve in the near term and how it will reshape work, research, and industry.

1. Engineering, Not Architecture, Is the Near‑Term Bottleneck

Recent gains in the AI stack are measurable: training speed (tokens per second per GPU) and inference cost (tokens per prompt, FLOPs per token) have been steadily optimized. Companies routinely shave 10‑30 % off serving costs, and the cumulative effect of these improvements is a near‑exponential decline in the cost of model intelligence. The next wave will be AI agents that automate the end‑to‑end optimization of inference pipelines, pushing utilization close to the theoretical limits of GPUs. Longer‑term, co‑design of specialized accelerators with model architectures promises additional orders of magnitude in efficiency.

2. The Shift From Execution to Idea Generation

As infrastructure becomes cheaper and faster, the value of AI work will tilt toward high‑level creativity. Coding agents already make model experimentation easier, turning the bottleneck from "can we build it?" to "what should we build?" Good ideas will outweigh raw engineering skill, echoing earlier transitions when deep learning moved AI from a research‑centric field to an engineering‑heavy one. This mirrors the historical pattern where breakthroughs first appear as research challenges, then become routine engineering tasks.

3. What This Won’t Fix

Efficiency gains alone do not create qualitatively new capabilities. Current models excel at narrow tasks—math, coding, literature mining—but they remain limited in general reasoning, long‑term planning, and robust understanding. Scaling compute and improving throughput will make these narrow superhuman abilities more affordable and widespread, yet they will not bridge the gap to a truly general, autonomous intelligence.

4. Economic and Scientific Ripple Effects

Lower costs will democratize access to powerful agents, spurring a Jevons‑paradox style surge in demand. Companies will focus on orienting and delivering agents rather than pushing raw performance. In science, models that can rapidly crawl literature and synthesize cross‑disciplinary insights could accelerate discovery in fields like biology and chemistry, though whether this leads to breakthroughs such as universal cancer cures remains uncertain. Similarly, the RL ecosystem suffers from low‑quality environments; investing in higher‑fidelity data pipelines promises substantial ROI for labs that can afford it.

Conclusion

The coming years will likely see AI become a ubiquitous, cost‑effective tool for specialized, superhuman tasks, driven by relentless engineering optimization. While this will transform industries and amplify scientific research, it does not inherently usher in general superintelligence. Understanding the limits of efficiency‑driven progress helps set realistic expectations and guides responsible investment in the AI future.