Qwen3.6-35B-A3B: Agentic coding power, now open to all

Qwen3.6-35B-A3B is an open-source MoE model featuring 35B total and 3B active parameters, specifically optimized for agentic coding and multimodal reasoning. Despite its efficient active parameter count, it rivals larger dense models like Gemma4-31B in language tasks and matches Claude Sonnet 4.5 across several vision-language benchmarks. The model supports native multimodality, thinking modes, and is accessible via open weights or an API compatible with both OpenAI and Anthropic protocols for integration with tools like Claude Code and OpenClaw.

Read more

Claude Opus 4.7

Claude Opus 4.7 is now generally available, delivering significant improvements in software engineering, long-horizon autonomy, and multimodal resolution (up to 3.75 megapixels). The update introduces an xhigh effort level for granular reasoning control and task budgets to manage token spend in complex agentic workflows. While an updated tokenizer increases token density by 1.0–1.35x, the model outperforms Opus 4.6 across coding, finance, and legal benchmarks with enhanced instruction following and self-verification capabilities.

Read more

The local LLM ecosystem doesn’t need Ollama

Ollama is a popular LLM runner criticized for being an inefficient wrapper of llama.cpp that has historically obscured its dependencies and failed to comply with MIT licensing. Technical issues include a 30-70% performance overhead compared to upstream llama.cpp, broken GGUF chat template handling via its proprietary Modelfile system, and misleading model labeling that misrepresents distilled models as full-parameter versions. The project's shift toward closed-source components, VC-driven cloud monetization, and a history of slow responses to security vulnerabilities like CVE-2025-51471 have led many in the community to recommend more transparent alternatives like LM Studio, Jan, or direct llama.cpp implementations.

Read more

€54k spike in 13h from unrestricted Firebase browser key accessing Gemini APIs

A developer reported a €54,000 Gemini API billing spike occurring within hours of enabling Firebase AI Logic, attributed to automated traffic and delayed budget alerts. Google Cloud initially denied a billing adjustment, but Logan Kilpatrick responded by highlighting new safeguards including Tier 1 spend caps, project-level spend limits, and the deprecation of unrestricted API keys. Google is also rolling out prepaid billing globally to provide developers with better cost control and prevent similar unexpected overages.

Read more

Cloudflare's AI Platform: an inference layer designed for agents

Cloudflare has introduced a unified inference layer that allows developers to access over 70 models from 12+ providers through a single API and credit system. The platform optimizes agentic workflows by providing automatic failover, centralized cost tracking via custom metadata, and low-latency execution on Cloudflare’s global network. Additionally, the integration of Replicate’s Cog technology enables users to containerize and deploy custom fine-tuned models directly onto Workers AI.

Read more

LLM risk spreading misinformation to humans who are least able to identify it

This research investigates how LLM response quality, including accuracy, truthfulness, and refusals, varies based on user English proficiency, education level, and country of origin. Experiments on SOTA LLMs and datasets reveal that undesirable behaviors occur disproportionately more for users with lower English proficiency, lower education, and non-US origins, rendering these models unreliable for their most vulnerable users.

ArXiv

Sparser, Faster, Lighter Transformer Language Models

This research introduces a novel sparse packing format and optimized CUDA kernels to exploit unstructured sparsity in LLM feedforward layers. By utilizing L1 regularization to achieve over 99% sparsity with minimal performance loss, the approach significantly improves throughput, energy efficiency, and memory utilization during both training and inference.

ArXiv

LLM Personalization Breaks Down in High-Stakes Finance

The paper identifies four fundamental limitations of standard LLM personalization when applied to individual portfolio management: behavioral memory complexity, long-term thesis consistency, style-signal tension, and alignment without immediate ground truth. To address these, the authors propose architectural responses and research directions for deploying personalized LLMs in high-stakes, temporally extended decision domains where user preferences evolve and outcomes are stochastic.

ArXiv

What hackers talk about when they talk about AI

This study analyzes over 160 cybercrime forum discussions to explore how threat actors leverage AI, ranging from the misuse of legitimate tools to the development of bespoke criminal models. While criminals are actively investigating AI to scale and sophisticate attacks, they also express significant concerns regarding its impact on operational security and business models. The research provides a thematic analysis of these emerging AI-enabled threats to assist law enforcement and policymakers.

ArXiv

Sir-Bench – benchmark for security incident response agents

SIR-Bench is a benchmark for autonomous security incident response agents that uses the OUAT framework to replay real-world incident patterns in cloud environments. It evaluates agents across 794 test cases using metrics for triage accuracy, novel finding discovery, and tool usage, verified by an adversarial LLM-as-Judge to ensure genuine forensic investigation over alert parroting. Baseline evaluations demonstrate high TP detection and effective evidence discovery, providing a standardized measure for future LLM-driven security agents.

ArXiv

Guy builds AI driven hardware hacker arm from duct tape, old cam and CNC machine

AutoProber is an automation stack that enables agents to perform flying probe hardware analysis by integrating GRBL CNC machines, USB microscopes, and oscilloscopes. It uses computer vision to identify PCB features, stitch annotated maps, and execute precise probing of pins and pads under a safety-critical monitoring system. The framework provides a Python-based control layer and web dashboard for agentic or manual hardware interaction.

Repo

MacMind – A transformer neural network in HyperCard on a 1989 Macintosh

MacMind is a 1,216-parameter, single-layer transformer implemented entirely in HyperTalk for the Macintosh SE/30. It utilizes standard components like self-attention, cross-entropy loss, and backpropagation to learn the bit-reversal permutation for FFTs. The project serves as a pedagogical tool to demonstrate that the mathematical foundations of modern LLMs are hardware-agnostic and fully inspectable.

Repo

SDL bans AI-written commits

SDL is a cross-platform library designed for developing multimedia software, including games and emulators. It is distributed under the zlib license, with further information and installation instructions available on its official website.

Repo

Agent Armor, a Rust runtime for enforcing policies on AI agent actions

Agent Armor is a zero-trust governance runtime for AI agent actions, utilizing an 8-layer deterministic pipeline to allow, block, or flag tool calls for human review. It features MCP-aware inspection, response scanning for PII and secrets, behavioral fingerprinting, and rate limiting. The Rust-based system provides a full audit trail with risk scoring and supports both MCP proxy and server modes to secure agentic access to shells, databases, and APIs.

Repo

Sudomake Friends, personalized AI personas in a Telegram group chat

Sudomake Friends is a framework for deploying autonomous AI agents into a Telegram group chat, featuring persistent memory, timezone-aware schedules, and distinct personalities. The system uses a CLI wizard to scrape a user's digital footprint across platforms like GitHub and Hacker News to generate tailored agent profiles and shared backstories. Technically, it leverages Claude for inference, incorporates RSS feeds for real-time context, and manages long-term state through periodic chat summarization and Docker-based deployment.

Repo