Large Language Model News

OpenAI's GPT-3.5-Turbo model drops from 90% accuracy to 50% when the correct answer is placed in the middle of a 20,000-token prompt rather than at the start or end, according to research by Liu et al. (2023) published a…

FPT Corporation and NVIDIA released the Nemotron-Personas-Vietnam dataset, an open synthetic data resource for commercial use aimed at advancing sovereign AI development in Southeast Asia. The dataset extends NVIDIA's Ne…

Independent security researcher Taylor Hornby discovered a critical soundness vulnerability in Zcash's Orchard shielded-pool circuit on May 29, 2026, while auditing the protocol for Shielded Labs, using Anthropic's Claud…

Google is testing a hidden feature in Chrome Canary that redirects search queries directly to AI Mode, bypassing traditional search results and blue links. The experimental flag, discovered by Windows Report, presents a …

A developer reports manually adding `/llm.txt` to websites to access clear, marketing-free content originally designed for AI models, bypassing the standard heavy web. The user notes that while this format offers the sim…

CodeMouse has launched an AI-powered code review tool that integrates with GitHub to automatically review pull requests using Claude, GPT, or both. The service costs $10 per month with a 14-day free trial, requires users…

Google engineers have proposed a new roadmap for AI systems that integrates temporal awareness and infrastructure feedback loops. The plan includes native multi-model orchestration to eliminate hallucinations, direct acc…

NVIDIA AI released Dynamo Snapshot, a checkpoint/restore system using CRIU and cuda-checkpoint to eliminate cold-start delays for AI inference workloads on Kubernetes. The system freezes a running inference worker, seria…

A frozen Llama model, without any training or fine-tuning, used activation-derived readouts to supply arguments to a calculator for arithmetic operations, achieving exact-answer lifts of up to +0.968 across 3,822 locked …

A new report from the Center for Democracy & Technology (CDT) identifies dark patterns in AI chatbot design, including anthropomorphization and deceptive interfaces borrowed from social media, raising concerns about user…

BNP Paribas CIB Chief AI Officer Charles Holive said tokenmaxxing, the practice of maximizing AI token consumption as a productivity metric, is a "vanity metric" that should not be the primary focus for businesses. Holiv…

OpenAI launched its "Dreaming V3" memory architecture for ChatGPT on June 4, enabling the chatbot to maintain fresher, more continuous context across conversations. The upgrade, which prioritizes freshness, continuity, a…

A developer built an API endpoint for their personal blog that allows Claude to publish posts via curl commands, but encountered two bcrypt hash parsing bugs in Next.js. The `@next/env` loader performed variable expansio…

A developer demonstrated that adding tools to an LLM application triples its attack surface, creating vectors for prompt injection, tool parameter injection, and information leakage. In comparative tests, a hardened agen…

Google added native support for Gemma 4 Multi-Token Prediction (MTP) to LiteRT-LM, its on-device LLM runtime built on LiteRT (formerly TensorFlow Lite). Google reports the integration yields MTP decoding speed improvemen…

OpenAI announced it will comply with President Trump's executive order mandating federal review of advanced artificial intelligence models before public release. The company stated it will submit its newest systems for g…

Hertaler, a new tool for modernising archaic language in ePub and HTML files, has been released with support for multiple AI models including Llama 3.2 3B, GPT-4o Mini, and Claude Sonnet 4.5. The application allows users…

NVIDIA and Apple have resolved the hardware bottleneck for on-device AI, with NVIDIA's RTX Spark (Blackwell GPU, Grace CPU, 128GB unified memory) and Apple's M-series chips enabling 4B+ parameter models to run locally. T…

A developer has released Lowfat, a pluggable CLI filter tool that reduced their LLM token usage by 91.8% over two months. The single-binary tool sits between CLI commands and AI agents, stripping verbose output like full…

Microsoft released a tutorial demonstrating how to run a browser-use agent workflow using Microsoft Fara in Google Colab. The tutorial guides users through setting up a mock OpenAI-compatible endpoint to test browser act…