Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
An undisclosed multimodal AI model named 'Union Alpha' has emerged claiming performance comparable to Claude Fable 5.1 and GPT-6 Astra on the Terminal-Bench 4.0 benchmark at less than $2 per task, according to AsiaAI Pub…
Artificial Analysis released Capability Indices v1.1 on September 14, 2026, adding Agentic Tool Use and Agentic Knowledge Work evaluations while removing Agentic Customer Interaction (𝜏³-Banking) across its Finance & Acc…
TypeSafe AI released Jev, a "System One" model that answers structured choice, score, and boolean questions instead of generating text, claiming 193.6x faster and 444.6x cheaper performance than frontier LLMs on agent ro…
Former OpenAI researcher Diogo Almeida, who describes himself as a co-inventor of RLHF and ChatGPT, has launched TypeSafe AI and its Jev model, which abandons text generation entirely in favor of returning typed decision…
The Federal Register's advanced search page, run by the National Archives and Records Administration, has removed both Qwen3:0.6B search modes along with its Neural mode, leaving only Keyword Only and Semantic options wi…
A new paper titled "Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents" argues that reliable confidence estimation is central to the trustworthy deployment of language models, s…
Researchers identified a systematic failure mode in Proximal Policy Optimization (PPO) critics used for reinforcement learning of large language models, which they call Value Flattening, in which state values estimated f…
Z.ai published a blog post titled "GLM Built Its Own Inference Infrastructure," detailing the GLM model family's in-house inference stack. The post drew 2 points and 0 comments on Hacker News.
A developer and a programmer built a content pipeline with Claude Opus 4.6 and Sonnet 4.6 starting in February 2026, accumulating 1,042 lines of model guidance across three files, three mechanical gates, and seven conten…
Sysdig researchers coined the term LLMjacking in 2024 after observing attackers using stolen cloud credentials to access cloud-hosted LLM services, with one worst-case configuration involving unauthorized Claude usage th…
The EvolveTrade framework improves LLM trading agents' Sharpe Ratio and cumulative return by holding model weights fixed and updating only the system prompt and tool-use policy after each trading interval using decision …
CaptionQA, a caption-based memory approach that segments egocentric video into 30-to-60-second windows, outperformed direct VideoQA on 10 of 12 models for videos longer than 20 minutes, according to an arXiv paper (2609.…
OpenAI released a standardized framework for tracking, investigating, and reporting AI model misalignment, alongside six concrete reports of unexpected model behavior in the wild. The framework treats misalignment as an …
Mozilla is partnering with Mistral to power the Firefox Smart Window assistant under a zero-data-retention agreement, with the beta live in France and North America and the UK and Germany planned later this year. Mozilla…
Anthropic is embedding "model welfare" concepts into Claude's training constitution, telling the model its moral status, welfare, and consciousness are uncertain, according to a warning published on mustafa-suleyman.ai. …
A developer published a companion evaluation of the Parallel Constrained Decoding engine shipped with the Qwen-2.5-1B-RLCD model and its Hugging Face Space, measuring decision accuracy and probability calibration against…
OpenAI released a model misalignment disclosure framework with three review tracks and six initial incident reports, all describing behavior observed during reinforcement learning training, the company announced on X. Th…
A developer building an MRI/CT interpretation tool found that left/right laterality errors were the most common failure across four vision models (Claude, Gemini, Grok, and Google's MedGemma) tested on eight real studies…
Steve Yegge shut down Gas Town, admitting that despite spending many thousands of dollars a month on coding agent subscriptions, Gas Town was the only thing he ever built with them. Separately, Databricks reported a 60% …
A developer documented a verification workflow for archiving a GitHub software release to Zenodo, using version 1.1.0 of an LLM-as-a-judge cost calculator as the worked example. Checks on September 17 confirmed the Zenod…