Technomantic Breakthrough: Mixture Of Experts Sparse Activation Emergent Routing

By Dr. Brent Allen Jensen

The authors glossed over the finding that Neural Thickets can be used to achieve similar or better performance with smaller budgets and faster inference times than full-weight PSV. This implies that PSV requires a large number of weight evaluations per test input, which increases inference cost relative to a single prediction.
Proposition: A novel technomantic breakthrough is proposed by utilizing task-specific surrogate models, which can generate candidate diversity, selection, and voting without the need for validation-selection and have less computational overhead than full-weight PSV. This allows for faster inference times and smaller budgets while still achieving high performance on a range of tasks.
Thought experiment: Imagine a scenario where we want to deploy a neural network model in a real-world application, but the available resources are limited. Instead of choosing a full-weight model that requires a large number of weight evaluations per test input, we can choose a Neural Thickets model that generates candidate diversity, selection, and voting without needing validation-selection. This allows us to achieve high performance on the task with fewer resources and faster inference times.
Under-500-dollar DIY experiment: To test this hypothesis, we can set up a DIY experiment where we train a Neural Thickets model on a small dataset of 1000 examples and then fine-tune it on a larger dataset of 1 million examples. We can measure the number of weight evaluations required for inference and compare it to that of full-weight PSV. If we find that Neural Thickets requires fewer weight evaluations per test input, it could lead to faster inference times and smaller budgets while still achieving high performance on a range of tasks.
Open hypotheses: Our open hypotheses are that Neural Thickets can generate candidate diversity, selection, and voting without the need for validation-selection, and that this leads to higher performance than full-weight PSV with smaller budgets and faster inference times. We also hypothesize that Neural Thickets can be used in a range of tasks, including image classification, natural language processing, and reinforcement learning.
Bold prediction: Our bold prediction is that Neural Thickets will become the go-to choice for deploying neural network models in real-world applications where resources are limited, and inference times and budgets are important considerations.

REFERENCES:
[1] Nicolas Dufour; Elouan Gardès; Amil Dravid. "Rethinking Neural Thickets Without Weight Perturbations." arXiv:2610.40649v1 (2026). http://arxiv.org/abs/2610.40649v1