MiniMax has launched M2.7, an AI model capable of autonomous self-improvement through self-training loops. The model achieves performance comparable to top Western models in agentic engineering tasks and matches coding benchmarks near Opus-level, while being more cost-efficient.
Sources: 4 Covered by Latent Space, The Code, The Neuron, The Rundown AI
The UK government has backtracked on its plan to allow AI companies to train on copyrighted works with an opt-out mechanism, reverting to existing law that requires permission. This reversal follows significant outcry from artists and sector groups, emphasizing the ongoing debate around intellectual property rights and AI development.
Sources: 2 Covered by The Neuron, The Rundown AI
Astral, the company behind popular Python tools like Ruff and uv, has been acquired by OpenAI and will join its Codex team. This acquisition aims to combine Astral's expertise in developer productivity with OpenAI's AI advancements to revolutionize software development.
Sources: 1 Covered by Hacker News
The International Conference on Machine Learning (ICML) 2026 desk-rejected 497 submissions after 506 reciprocal reviewers violated policies regarding the use of Large Language Models (LLMs) in peer review. The conference used PDF watermarking to detect LLM-generated reviews, underscoring its commitment to maintaining scientific integrity.
Sources: via Hacker News
NVIDIA has released NemoClaw, an open-source stack designed to simplify the secure deployment of OpenClaw always-on AI assistants. It provides a sandboxed environment using NVIDIA OpenShell runtime and routes inference through NVIDIA cloud, enhancing security and control for autonomous agents.
Sources: via Hacker News
NVIDIA researchers have developed SPEED-Bench, a new benchmark designed to provide a unified and diverse evaluation framework for speculative decoding (SD) techniques in large language models (LLMs). The benchmark addresses the limitations of existing evaluations by using diverse semantic domains and realistic serving conditions to accurately measure SD performance.
Sources: via Hugging Face Blog
A new arXiv paper argues that transformers are fundamentally Bayesian networks, with each layer implementing a round of weighted loopy belief propagation. This research provides a theoretical framework for understanding transformer functionality and suggests that issues like hallucination are structural consequences of operation.
Sources: via arXiv AI
A new research paper proposes L2A (Learning To Attend), a novel layer that enables large language models to efficiently process longer contexts by conditionally invoking global attention. This extends models like Qwen 2.5/3 from 32K to 128K tokens while significantly improving training throughput and reducing KV cache memory by up to 50%.
Sources: via arXiv AI