Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Security startup Hacktron used Anthropic's Claude Opus 5 to take over an OpenAI employee's ChatGPT account in July through a flaw in the third-party community forum platform Discourse, according to a Wall Street Journal …
Developer Jibs79 released SelMem, an early-stage experimental Rust project exploring whether LLM agents can develop persistent behavioural divergence through a reconstructive memory system that forgets, distorts, and con…
OpenAI's GPT-6 Astra solved between 97.6% and 98% of the 43 problems in Tier 4 of Epoch AI's FrontierMath benchmark, earning the first-ever "Major Advance" classification from Epoch for an AI-generated mathematical solut…
Newly unsealed court documents in a copyright suit brought by The New York Times, the New York Daily News and 10 other publications show Microsoft and OpenAI employees internally warned their AI models posed an "existent…
Tayoca has outlined a governed publication workflow that separates an AI system's access to internal evidence from its authority to publish, arguing the two controls must not collapse into one another. The approach stage…
OpenAI launched Astra for Law, a legal-work product that combines the GPT-6 Astra model with a dedicated legal search index and legal-specific instructions rather than a separately trained base model. OpenAI says the con…
The Acton-Boxborough Regional School District (ABRSD) published an AI accountability document, "AI Guidelines & Guardrails," several months before American Compass's Oren Cass argued in a September 10, 2026 New York Time…
Three security researchers used Anthropic's Claude models to break into OpenAI's internal systems through its community forum in less than 72 hours, according to The Decoder. The team said Claude Opus 5 succeeded where i…
A developer outlined a runtime prompt-assembly pipeline that builds LLM inputs from templates, retrieval, conversation memory, and caching, using a customer-support reply assistant as the example. The approach keeps a st…
Z.ai released GLM-5.3-Flash and GLM-5.3-FlashX, the first native multimodal models in the GLM-5 series, with GLM-5.3-FlashX delivering inference speeds of 200 tokens/s. The models accept video, image, text, and file inpu…
A research paper titled "Self-Evolving Search Index" argues that retrieval quality for LLM agents depends heavily on how effectively index keys expose the knowledge contained in each document, since retrieval relies on a…
Anthropic rolled out Projects within Claude Code, letting a single conversation spawn parallel cloud sessions that pass context between each other and keep running after the session closes, with Anthropic staff Cat Wu an…
Moonshot AI's Kimi K3 is now available on Amazon Bedrock, with AWS stating the model is the first open model to reach 2.8 trillion parameters and the first open-weight model on Bedrock to support explicit prompt caching.…
A RubyConf 2026 keynote and a widely-read essay by a veteran practitioner argue that AI code generation is eroding developers' understanding of the software they ship, invoking Giambattista Vico's Verum Factum to suggest…
A Hugging Face forum user reported that an Ollama pull of the model hf-co/empero-ai/Qwen3.8-35B-A3B-Distill-GGUF:Q8_0 failed with a "blocked redirect to a different host" error before resolving on its own. Forum user neg…
Google released Gemini 3.8 Live Extended Thinking in September 2026, a model that narrates its reasoning in real time via live audio, ranking first on the Artificial Analysis Speech-to-Speech Quality Index at 82.6, scori…
TypeSafe AI emerged from stealth on September 15 with a $40 million seed round led by DCVC for Jev, a model that returns bounded decisions instead of chat text. In an Ably-built Pong demo, Jev made 47 decisions in 12 sec…
Google announced the preview of a native BM25 index in AlloyDB and Cloud SQL for PostgreSQL 17+, built on the open-source pg_textsearch extension created by TigerData, eliminating the need for a separate full-text search…
Three independent cybersecurity researchers operating as Hacktron—Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini—used Anthropic's Claude Opus 5 to exploit a Discourse image-upload flaw and gain access to the ChatGPT and…
A developer built SoFar, a 425-byte gzipped, zero-dependency JavaScript library that parses partial JSON as an LLM streams tokens, returning the best complete value available instead of throwing. The library scans the bu…