We tested every major AI reasoning system. There is no clear winner.

We tested every major AI reasoning system. There is no clear winner.

Reflecting on Frontier AI Reasoning Systems

It's been six months since OpenAI achieved a breakthrough high score on ARC-AGI-1 with . That moment marked the end of AGI scaling exclusively through pre-training. It's hard to overstate how strongly the industry led the public to believe that moment would never come.

As a concrete reminder, California contended with the SB-1047 AI safety senate bill last summer, predicated on urgent action necessary to curtail imminent danger from pre-training scaling. In a coicidence of history, OpenAI announced one week before Governor Newsom vetoed the bill.

Fortuantely for AGI progress, the "vibe" has shifted to exploring new ideas. The leading idea, adopted now by every major lab, gives AI systems additional "time to think" at test-time. All frontier labs are using chain-of-thought (CoT) methods to accomplish this — DeepMind, OpenAI, xAI, DeepSeek, and even Anthropic who is a staunch proponent of pre-training scaling.

This has given rise to a new narrative: scaling up test-time compute (TTC) will get us to AGI. Let's examine this.

Test-Time Adaptation

Modern CoT usage leverages three distinct techniques.
Long-Running Inference

The first is simply letting the model generate tokens in a long-running continuous stream. These are sometimes a dedicated process model, other times a fine-tuned base model, but always trained explicitly to emulate human-reasoning traces via supervised fine-tuning. During inference, output tokens are fed back into the model. This gives rise to model's capability to pattern match against degenerate reasoning (e.g., noticing when CoT is caught in a loop) and to adopt explore vs exploit techniques all via a single long-running CoT.
Knowledge Recomposition

The second is to remix or recombine knowledge from the pre-trained or process model. Generally the goal here is to create a "better" CoT. One complete CoT can be thought of…