Import AI 470: No rights for machines; automating environment generation with SPADE; and building better GPU kernels with Hawkeye

In the latest Import AI newsletter, researchers examine the differential acceleration of scientific progress due to AI advancements. A study by METR reveals that AI has significantly impacted cyber research, leading to a rapid increase in reported vulnerabilities across numerous projects and databases. In contrast, mathematics research has experienced minor acceleration, with some prestigious problems solved but the value of these contributions difficult to quantify.

AI's influence on AI research has been minimal, with only a few algorithmic progress areas showing measurable improvements.

Researchers from multiple universities have developed SPADE, a framework for co-evolving environments synthesis and agentic capability through self-play. SPADE generates synthetic data in the form of game-like environments, which LLMs can be trained in. The framework comprises an LLM that alternates between generating executable training environments and trying to solve them.

This process utilizes a reward system estimating the gap between the reasoning agent's reward with and without privileged hints, which are task-relevant information provided to the Environment Designer.

SPADE has been tested on Qwen3 backbones at the 30B scale, with the best performance observed in Qwen3-30B-A3B-Instruct-2507. The framework generates two types of environments – game environments and tool-use environments, both of which show significant improvement when utilizing SPADE. This technique is seen as a form of synthetic data generation, allowing researchers to create diverse training environments for other models, making it cheaper to build large, broad datasets for model training.

However, SPADE does not enable models to bootstrap themselves significantly beyond their base model's imaginative capabilities.