Gemma 4 12B: A unified, encoder-free multimodal model

Gemma 4 12B is a multimodal LLM designed for local inference on consumer hardware with 16GB of VRAM. It features a unified, encoder-free architecture that processes vision and audio inputs directly within the LLM backbone to minimize latency and memory overhead. The model delivers reasoning performance nearing the 26B MoE variant and incorporates Multi-Token Prediction (MTP) drafters to accelerate generation.

Read more

Uber's $1,500/month AI limit is a useful signal for AI tool pricing

Uber has implemented a $1,500 monthly spending cap per employee for agentic AI coding tools like Claude Code and Cursor after exhausting its 2026 AI budget in four months. This cap represents approximately 11% of a median US software engineer's annual compensation, highlighting the significant operational costs of token-intensive coding agents at enterprise scale. The policy reflects a shift toward managing full API pricing as companies move beyond subsidized individual subscription plans.

Read more

Artificial intelligence is not conscious – Ted Chiang

Anthropic’s "Claude’s Constitution" and public statements from its leadership attribute human-like qualities—including judgment, emotions, and moral status—to their LLM. Ted Chiang critiques this pervasive anthropomorphism, questioning whether LLMs are truly conscious entities capable of moral instruction or emotional distress.

Read more

Failing grades soar with AI usage, dwindling math skills in Berkeley CS classes

UC Berkeley is reporting a significant surge in failing grades across core CS and EECS courses, with failure rates in some introductory classes exceeding 35%. Faculty attribute this trend to an overreliance on LLMs for assignments, which leads to academic dishonesty and leaves students unprepared for rigorous, proctored exams. This decline in performance is further exacerbated by dwindling mathematical prerequisites and reduced TA staffing levels.

Read more

U of T researchers demonstrate AI worm could target any online device

Researchers at the University of Toronto have demonstrated a new class of adaptive malware: an AI-powered worm utilizing open-weight models to automate network exploitation. The worm siphons compute from infected hosts to power its own reasoning, enabling it to pivot and tailor attacks in real-time at near-zero marginal cost. This proof-of-concept highlights the vulnerability of traditional defenses against autonomous, LLM-driven threats that target underlying software and hardware rather than just AI applications.

Read more

DMF: A Deterministic Memory Framework for Conversational AI Agents

DMF replaces generative, LLM-based memory summarization with a deterministic, CPU-first framework using classical NLP and vector geometry. It manages long-term context through a Survival Score ($\Omega$) and an interaction-count decay law, eliminating non-determinism and token costs associated with memory preparation. Evaluation against Mem0 demonstrates comparable accuracy while reducing total token overhead by 5x to 242x.

ArXiv

AI Agents Enable Adaptive Computer Worms

Researchers have demonstrated an AI-driven worm that utilizes open-weight LLMs on compromised hosts to autonomously generate target-specific attack strategies. By leveraging parasitic compute across Linux, Windows, and IoT environments, the worm achieves zero marginal cost for propagation and bypasses centralized safety controls. This shift from static exploit code to real-time reasoning marks the emergence of autonomous generative adversaries capable of adaptive, cross-platform network exploitation.

ArXiv

AI Agents Enable Adaptive Computer Worms

Researchers have demonstrated an AI-driven worm that utilizes open-weight LLMs on compromised hosts to autonomously generate target-specific attack strategies. By leveraging parasitic compute across Linux, Windows, and IoT environments, the worm achieves zero marginal cost for propagation and bypasses centralized safety controls. This shift from static exploit code to real-time reasoning marks the emergence of autonomous generative adversaries capable of adaptive, cross-platform network exploitation.

ArXiv

Your AI Text is not Mine

To address the lack of standardized definitions for harmful AI-generated text, the authors introduce AITDNA, a benchmark of human-machine co-constructed texts featuring full edit and interaction histories. Their evaluation reveals that current detectors are specialized for specific generation types and fail to function as robust, general-purpose solutions across diverse AI-generated text categories.

ArXiv

Large AI Models in Dental Healthcare

This systematic review evaluates language-generative, vision foundation, and dental-specific models (e.g., OralGPT, DentVFM) across 97 studies. While LLMs excel in clinical reasoning and text-based tasks, vision-based models like SAM and CLIP variants are superior for diagnostics, with dental-specific models leading in multimodal performance. Current challenges include a data asymmetry favoring vision pretraining over text, model hallucinations, and the absence of standardized clinical evaluation benchmarks.

ArXiv

Ideogram 4.0 – open-weight 9.3B text-to-image model

Ideogram 4 is Ideogram's first open-weight, state-of-the-art text-to-image foundation model, trained from scratch as a 9.3B parameter fully single-stream Diffusion Transformer. It utilizes Qwen3-VL-8B-Instruct as a vision-language model text encoder and supports structured JSON prompting for extreme controllability, enabling best-in-class text rendering, explicit bounding-box layout, and color palette conditioning at native 2k resolution. Benchmarks consistently position Ideogram 4 as the top open-weight model across various design and general text-to-image generation tasks.

Repo

Mnemo – local-first AI memory layer for any LLM (Rust, SQLite,petgraph)

mnemo is a local-first, Rust-based memory layer that provides persistent knowledge graph capabilities for LLMs. It functions as a sidecar service to extract entities and relationships into SQLite, enabling multi-hop graph traversal and semantic retrieval for context injection. The system supports OpenAI-compatible APIs and Ollama, offering a high-performance, cloud-independent alternative to traditional RAG pipelines.

Repo

Ongoing NPM supply chain attack uses binding.gyp to spread like a worm

The ai-sdk-ollama package is a Vercel AI SDK v6 provider for Ollama, offering type-safe integration with local LLMs and advanced features. It provides reliable tool calling, automatic JSON repair, built-in web search, and RAG reranking. The SDK also supports autonomous ToolLoopAgents, a middleware system, and comprehensive AI SDK compatibility for text generation, streaming, structured output, embeddings, and vision models across Node.js and browsers.

Repo

Agent-browser-shield – free extension to protect AI agents on the web

Agent Browser Shield is a Chromium MV3 extension designed to optimize agentic browser-use for LLMs by stripping non-essential page elements to improve token efficiency. It mitigates security risks like prompt injection and PII leakage by masking sensitive data and suppressing hidden text or user-generated content. The project includes a benchmark harness and integration scripts for agent runtimes like Browserbase and Stagehand to evaluate model performance and accuracy across various web environments.

Repo

Fork of Rsync

This repository is a fork of rsync that excludes LLM-generated commits and is open for community contributions.

Repo