"The first AGI" refers to the model that causes this market to resolve (unless it resolves to 2049 because of no AGI by then, in which case this market will resolve N/A.)

"A transformer model" refers to any neural network whose primary mechanism for integrating information across sequence positions is self-attention, as introduced in Vaswani et al. (2017), "Attention Is All You Need."

A model qualifies as a transformer if the majority of its inter-token information mixing—the process by which the model allows different parts of its input to influence each other—is performed by attention layers (including standard softmax attention, linear attention, or other direct variants of the attention mechanism). Mixture-of-Experts architectures remain transformers if their underlying experts use attention-based processing. Retrieval-augmented systems are judged solely by the architecture of the core model, not the retrieval mechanism. A model does not qualify as a transformer if its primary inter-token mixing mechanism is recurrence (RNNs, LSTMs), convolution, or state-space operations (e.g., Mamba, RWKV, S4). Hybrid architectures that combine attention with these alternatives qualify as transformers only if attention accounts for more than half of the model's inter-token mixing, measured by parameter count in the relevant layers. If this is ambiguous or cannot be determined from public information, the default assumption is that the architecture is a transformer if it is marketed or described as such by its developers, and not a transformer if it is marketed as a post-transformer or alternative architecture. If the qualifying AGI system is an ensemble or multi-model system, this market resolves based on the architecture of the model responsible for the majority of the system's reasoning and planning capabilities, as judged by the market creator in consultation with credible technical sources.