AI in 24 Hours: Biosafety, Strategy Games, and Practical Analyses

In today’s roundup: where AI is already changing biological research and why it’s still too early to trust it with strategic decisions. Plus, recent papers on speeding up and correcting language models, and an analysis of 152-FZ.

🔥 Hot:

🔹 Researchers analyze threats and defenses related to the use of AI in biology — The paper describes how language and biological foundation models affect the digital and physical dimensions of biological research. 🔹 Researchers believe AI is not ready for strategic conflicts — Language models can simulate adversaries and crises, but their scenarios remain fragile and poorly suited for high-stakes decisions.

➡️ Useful materials:

🔹 A new paper studies request routing between GPU pools for LLMs — The author proposes estimating the additional processing time for each instance to more effectively divide resources between prefill and decode. 🔹 CRN v2 corrects errors in Gemma 4 E2B without fine-tuning the base model — The lightweight correction module operates on top of a fully frozen model and is intended to preserve its original capabilities. 🔹 A study proposes neural network pruning based on Fisher distances — Parameters are removed based on a geometric estimate of how much the model will change when they are zeroed out. 🔹 An analysis explains what a company may face for sending a contract to ChatGPT under 152-FZ — The author clarifies the current penalty amounts and examines Russian court practice concerning the use of neural networks.

➡️ Discussions and case studies:

🔹 The community discusses whether an old X399 computer is suitable for a build with multiple GPUs — The question concerns motherboards with four PCIe 3.0 x16 slots and second-generation Threadripper processors.

📝 If you’d like to add other news and materials to the list, leave a comment.