Large Language Model News

The Kimi Team published a new attention architecture, Kimi Linear, on arXiv in 2025. The architecture uses Kimi Delta Attention interleaved 3:1 with full attention, becoming the first linear hybrid to outperform softmax …

Researchers Zhao et al. published a paper on arXiv 2025 introducing Absolute Zero, a method where a model proposes its own tasks and a code executor grades them, enabling reasoning reinforcement learning with zero human …

OpenRouter launched Fusion API, a compound AI system that orchestrates up to eight LLMs in parallel to improve reasoning accuracy, achieving 69% on the DRACO benchmark. Microsoft released seven new MAI models built from …

Industrial digital twins are evolving from visualization tools into machine information systems that integrate live data, documentation, and operational knowledge, driven by AI demand for structured data and new EU regul…

OpenAI's IPO asks public investors to bet that the company can own the layer where AI does real work—harnesses, interfaces, and workflows—rather than just the underlying GPT models. The company restructured as a Public B…

MiniMax introduced a sparse attention mechanism that selects per-group blocks to enable cost-effective inference on million-token sequences. The technique reduces computational overhead while maintaining model quality, p…

A new AI workflow enables the generation of a complete YouTube video from a single prompt, using Claude Fable 5 to write the script, clone the voice, render the avatar, build motion graphics, and edit the video autonomou…

GitHub Copilot switched to token-based billing, causing one user's monthly bill to jump from $28 to a projected $700. The shift reflects the AI industry's move from flat subscriptions to usage-based pricing as heavy user…

A new paper titled 'When AI Says It Feels' warns against mistaking eloquent AI language for evidence of consciousness, urging a multi-angle approach to evaluating machine sentience. Researchers argue that claims of machi…

AI agents are moving beyond capability concerns to trust bottlenecks, with Anthropic reporting over 80% of code merged at the company now written by Claude. As agentic output outpaces human review, explainability must be…

Anthropic released Claude Fable 5 on June 29, 2026, but temporarily suspended access shortly after launch due to unforeseen issues. The model, which had been anticipated with an average release guess of July 2, 2026, is …

The memory needed to store one token of context in large language models has fallen by roughly 100x since 2017, driven by advances in KV cache compression techniques such as Multi-Query Attention (MQA) and Grouped Query …

Anthropic was forced to disable its frontier AI models Fable 5 and Mythos 5 after the Trump administration imposed new export controls on June 12, citing national security risks. The directive from Commerce Secretary How…

Data from New York's WARN Act filings shows no companies attributed layoffs to AI in the first year of mandatory disclosure, contradicting predictions of mass unemployment. Software engineers remain essential due to bott…

A developer argues that the most underrated feature in a website AI assistant is the ability to say 'I don't know,' rather than optimizing for fluency. The post describes a retrieval-based system with strict refusal poli…

A developer released Raidho, an open-source coding agent that uses a hybrid architecture with separate models for reasoning and execution, plus a durable algebraic memory (VSA) instead of RAG. The agent reduces costs by …

Senior Anthropic technical staff are in Washington to meet with White House officials over an export control dispute that forced the company to take its two most advanced AI models, Fable 5 and Mythos 5, offline. The dis…

Anthropic, the AI startup behind the Claude chatbot, is negotiating with the US Commerce Department after an emergency export-control directive forced it to shut down its Fable 5 and Mythos 5 models globally. The directi…

A developer built Quorum, a multi-region database application, by directing Claude Code, an AI coding agent, under the governance of an append-only architecture decision log. The log, containing numbered decisions from D…

Thirty mathematicians at Harvard blind-graded AI solutions to 10 original, unpublished research-level math problems, finding that leading AI systems passed on 7 of the 10 problems. The results, released June 10, 2026, sh…