Blog
Blog on technology essentials
Once a week we review fresh research in AI and agentic systems: what the authors showed, where the limits are and what it means for the industry
LLMs have learned to read and edit neural network weights
A frontier model takes apart the raw weights of small networks, repairs them and writes them from scratch — without gradient descent. What the authors showed and where the limits are.
Based on the paper by Takehi, Teoh, Zhang, Khattab · alphaXiv, November 2, 2026
Read the article7 min read
Memory tax: why agent memory can work worse than plain searchA UIUC and Tsinghua study: many memory systems lose to RAG over the raw history. We look at where the facts get lost.Search without an index: the model walks the document corpus itselfCornell's DecisionRAG beats dense retrieval plus a reranker by 14–17 points and stays within a couple of seconds per query.An agent should deliver work, not answersA new framework for work agents: model plus harness, a verifiable result and two axes — duration of responsibility and openness of the environment.When an agent should hand control to a specialized modelGPT-6 Astra controls a robot on its own and paired with a VLA policy. The pair is stronger, but the handover of control is a separate engineering problem.Two softmax bugs found in almost every frameworkDeepMind found how rounding zeroes out the gradient and breaks distillation, and proposed a simple fix.Heat by the weather: how the system decides how much heat a building needsProject breakdown: polling meters and automatically controlling regulators in thousands of residential buildings in Moscow.
What we write about
We read research because we build such systems ourselves. We write about what can be verified and applied in production.
Accredited IT company · Russian Ministry of Digital Development registry No. 20142