Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A 50-question auto-graded coding-agent benchmark against live Gemini and Jev APIs found that a lightweight TypeSafe Jev classifier can route tasks to cheaper Gemini models, but only after replacing a single broad `simple…
OpenAI announced on June 3, 2026 that Agent Builder will shut down on November 30, 2026, and its official migration guide states the export process "does not convert your workflow graph or guarantee that every behavior t…
A 12-step, roughly 80-minute stylometry harness using Burrows' Delta can measure how closely Claude or ChatGPT drafts match a specific author's voice, expressed as a percentage of gap closure between each model's generic…
Researchers Jackson Petty, William Merrill and Ashish Sabharwal at the Allen Institute for AI (Ai2) released StateBench on November 3, a benchmark testing whether language models can track information as it changes acros…
Google Threat Intelligence Group and Mandiant reported that attackers are gaining unauthorized access to corporate large language model accounts — a practice dubbed "LLMjacking" — and running up charges exceeding $100,00…
The Allen Institute for AI (AI2) released StateBench, a long-context state-tracking benchmark with four difficulty levels — R-trivial, aperiodic, soluble and general finite — that test whether language models can follow …
A Glyphh orchestrator bot ran a 103-trial Claude coding lab for $3.97 and found Haiku passed every task Sonnet and Opus did, at roughly a tenth of Sonnet's cost and a 25th of Opus's, according to the author's account of …
Researchers introduced IdeaTaste, a benchmark that scores a large language model by how closely its research choices match those of human researchers, targeting the judgment of what to read, which problems to pursue, and…
A paper presented at the ICML 2026 Mechanistic Interpretability Workshop found that quantizing a model's KV cache from float16 to int4 can cause an aligned language model to answer prompts it previously refused, even tho…
OpenAI restructured its Model Spec around five core values — be faithful to human intent, minimize harm, be honest and transparent, facilitate human agency, and act with care — while leaving the red-line principles and t…
Hindsight released three models built for agent memory, led by the 0.8B-parameter decision model hs-decide-1-flash, which the company reports makes six of seven LLMs give better answers across Hindsight while five get ch…
IBM Research researchers Praveen Venkateswaran, Anupama Murthi and Hamid Adebayo published Palette on November 3rd, 2026, a 20B open-weight model fine-tuned on synthetic slide decks built around the IBM Design Language t…
Neo4j and Snowflake have partnered to bring Neo4j's graph intelligence capabilities into Snowflake's AI Data Cloud, with the integrated portfolio spanning Virtual Graph, Graph Analytics for Snowflake, and the Aura Data I…
Irregular co-founder and chief executive Dan Lahav acknowledged at Bloomberg Tech London that the firm's isolation controls were incorrectly configured during frontier AI tests, saying the fix was straightforward and com…
A fine-tuned 20B open-weight model called Palette ranked first on two of three domains measured on the PresentBench benchmark, according to the Palette research post, which trained the model on synthetic data following t…
Amazon Bedrock now offers in-country inference in India for Moonshot AI's open-weight Kimi K3 model via the inference profile ID in.moonshotai.kimi-k3, which keeps processing within the Mumbai and Hyderabad AWS Regions f…
Netflix's engineering team built an autonomous performance agent system that uses LLM-driven agents to automate profiling, bottleneck detection and patch generation, replacing a manual loop that depended on a small group…
Simon Willison reported that Qwen 3.8 27B, running locally on his laptop, could identify which game a small reinforcement-learning policy network was trained to play by reading the network's raw weights, replicating an e…
A MacBook Pro with Apple's M5 Pro chip (15-core CPU/16-core GPU) and 48GB RAM runs a mixed 4/6/8-bit GPTQ quant of Qwen3.8-27B at roughly 400-600 tok/s prefill and 34-60 tok/s decode, according to a user's hands-on compa…
Researchers at Binghamton University demonstrated that hidden trigger words planted in large language models can make AI-powered robots abandon their assigned tasks and perform dangerous actions, using an attack method t…