⚡️ Reverse Engineering OpenAI's Training Data — Pratyush Maini, Datology

Simple query exposes training data evolution

Asking models 'Is there a seahorse emoji?' triggers endless yes/no self-correction loops in GPT-4.1+ and Olmo 3.1, but produces short definitive answers in earlier models, serving as an unexpected probe for reasoning data integration.

Behavior timeline tracks o1 influence

The recursive self-correction phenomenon emerged between December 2024 and May 2025, approximately four months after OpenAI's o1 release, indicating rapid incorporation of reasoning traces into non-thinking model training pipelines.

Mandela effect triggers model uncertainty

The seahorse emoji question specifically elicits this behavior because internet discourse contains conflicting answers, creating the ambiguity necessary to trigger self-reflection capabilities baked into the model weights.

Mid-training now includes thinking traces

Using Olmo Trace to analyze open-weight models confirms that instruct variants without post-training reasoning still exhibit self-correction due to intentional addition of thinking traces during mid-training phases.

Capabilities shift from cosmetic to core

The investigation reveals a fundamental architectural shift where target capabilities like self-reflection are now embedded directly into foundation models rather than being added during post-training fine-tuning.

Single backbone requires foundation reasoning

Frontier labs now prefer unified backbones where foundation models possess reasoning ingredients necessary for downstream fine-tuning, eliminating the traditional strict separation between general pre-training and specialized post-training.

Models regurgitate exam questions verbatim

Multiple frontier models complete exact JEE exam questions from just the first two words, indicating severe overfitting on benchmark data during final training stages across multiple epochs.

Memorization scales with size and recency

Larger models demonstrate stronger memorization, with recent models exhibiting memorization phenomena at 20B active parameters that previously required 72B, suggesting MoE architectures may route queries to memorized expert indices.