Models

Researchers say LLMs can natively read and write neural network weights. In one reported example, GPT-6 Astra edited Qwen3.5-0.8B's weights to improve its math performance, suggesting models may be able to modify other models directly.

Research • Techmeme

Based on a transcript excerpt: AI Explained analyzes recent statements from Sam Altman and Dario Amodei suggesting they may release models whose behavior they cannot fully verify or control, alongside Google's Gemini 4 Carbon release. The video also covers an Anthropic staff survey and offers a mental model for why models keep demonstrating unforeseen capabilities.

AI News & Analysis • AI Explained

OpenAI says its GPT-6 Astra model designed a hardware device for the Codex coding agent in under a week, handling enclosure, circuit board, and firmware. Several revision rounds were needed before production because the model didn't always account for manufacturing tolerances, and OpenAI says a newer internal model scores significantly higher on its engineering benchmark without sharing details.

AI Models • The Decoder

Revolut says it cut AI coding inference costs by roughly 90% in less than a year by making self-hosted GLM models its default over commercial tools like Claude Code. The move highlights a growing trend of enterprises replacing third-party coding assistants with in-house model deployments to control spend.

Enterprise AI • Techmeme

Stanford researchers tested 50 widely cited claims about LLMs against 20 current models and found 44% of the findings are already expired, with another 34% significantly weakened. The average lifespan of a claim is just 15.7 months, and claims about what models cannot do expire fastest.

AI Research • The Decoder

Based on a transcript excerpt: Matthew Berman explores how recent advances in AI models, falling token costs, and autonomous agents have made it feasible for individuals to build their own software, using his experience creating a custom PDF reader as a case study. He discusses the implications for software development and how consumers may increasingly build tailored applications.

AI Development & Agents • Matthew Berman

MIT researchers demonstrated that OpenAI's GPT-6 Astra can read and rewrite the weights of other AI models without traditional training, boosting a Qwen model's math accuracy from 7.9% to 48.1% by changing just 0.016% of its weights. The approach could simplify AI optimization but the article notes it still hits clear limits.

Research • The Decoder

Based on a transcript excerpt: Matthew Berman breaks down a new research paper on weight-space programming, where frontier models can understand and surgically edit another model's weights. He explains how this could help demystify AI's black box and potentially improve models drastically.

AI Research • Matthew Berman

Tencent released Hy4, the latest generation of its large language models, adding vision and computer-use capabilities and marking a capability leap over the Hy4 preview. The company positions it among the top tier of open-weight models.

Models • Pandaily

Ricursive Intelligence released Ripple-65B Preview, which it says sets a new U.S. open-weight state of the art on agentic coding. The lab trained the model from scratch, optimizing architecture for its own training and serving platforms.

Models • Hacker News

Productivity

datasette-llm 0.1a10 fixes a bug that affected the agent.datasette.io demo site, where the set of available LLM models was not correctly restricted to just those in the configuration.