AI War

AI Wars

2006 -
CUDA - Parallel processing

NVIDIA GPUs were originally designed for graphics rendering, but their architecture turned out to be perfectly suited for HPC (High-Performance Computing). Until the early 2000s, the computing industry's basic assumption was: performance improvement = clock speed increase (frequency scaling). Most programs were based on single-threaded serial computation. However, between 2003–2005, two things hit simultaneously: CPU heat and power limitations (the power wall). Around 4GHz, systems reached the thermal wall. CUDA was announced in 2006, introduced in 2007, and experienced explosive growth after AlexNet in 2012.

2012
AlexNet

Geoffrey Hinton,

Ilya Sutskever, and Alex Krizhevsky trained AlexNet using GPUs. Using 2 GPUs, they successfully trained a large-scale CNN on ImageNet (1.2 million images) for the first time, revolutionizing computer vision performance.

2016 GPU supercomputer

2017
Transformer Model

To use the increased compute more efficiently, the Transformer architecture was proposed, enabling parallel token learning instead of time-dependent RNN structures. This marked the beginning of serious exploration into meta-learning and general intelligence.

2020
AI Scaling

OpenAI, with sufficient compute resources, was the first to release GPT-1, 2, and 3 with autoregressive insights and scaling, observing induced intelligence. They focused on scaling at the pretraining stage.

2022
ChatGPT

By applying RLHF, AI became truly useful for users for the first time, dramatically expanding the user base. Massive amounts of inference responses began accumulating, which became the foundation for rough test-time RL in the form of feedback.

2025
Reasoning Model

With NVIDIA's Blackwell deployment delays reducing chip compute scaling, test-time scaling similar to GRPO began with DeepSeek. Inference scaling compensated for the missing scaling dimension. Simultaneously, OpenAI, which had dominated pretraining hegemony, had to slow down, and xAI, Google, and Anthropic rose to form a four-way competition. With inference-time reasoning becoming critical for agents, companies with abundant inference resources like Google and Amazon-backed Anthropic established a competitive structure.

Measuring AI Ability to Complete Long Tasks

We propose measuring AI performance in terms of the *length* of tasks AI agents can complete. We show that this metric has been consistently exponentially increasing over the past 6 years, with a doubling time of around 7 months. Extrapolating this trend predicts that, in under a decade, we will see AI agents that can independently complete a large fraction of software tasks that currently take humans days or weeks.