Panoptes is an open-source AI audit and alignment layer that monitors AI agents, recording every tool call, file read, and shell command, and provides a queryable audit trail with policy enforcement.

Stitch is an open-source, locally-run AI desktop app that integrates memory, meeting recordings, task management, automation, email, web browsing, and more. It works with multiple AI providers or local models, requires no accounts, and keeps all data on your machine for maximum privacy.

This article uses a computer architecture lens to dissect how AI programming agents actually run, explaining that their core loop is analogous to a CPU's fetch-decode-execute cycle. It details three bottlenecks of running agents locally (compute, persistence, environment heterogeneity), the brief hype and pitfalls of self-hosted solutions like OpenClaw on a Mac mini, and argues that cloud-native agents are inevitable, offering elastic scaling from 2-3 agents per laptop to 1000+ in the cloud with a 15x ROI.

Artificial Analysis announces AutomationBench-AA, an independent leaderboard for Zapier's AutomationBench, testing AI agents on real SaaS workflow automation across 657 tasks. Claude Fable 5 leads at 48.6%, followed by Opus 4.8 and Gemini 3.5 Flash. The benchmark reveals that all models violate some guardrails, with Finance tasks being the hardest. Gemini 3.5 Flash offers the best value, while GLM-5.2 is the top open weights model.

The terminal and Mac menu bar companion for Kickbacks.ai.

Foundation is an opinionated full-stack framework that eliminates translation layers between database and browser, delivering native performance through single-contract architecture, Hermes hotplane, and zero-copy communication, enabling AI agents and humans to build production systems without regression.

pi_agent_rust is a from-scratch Rust port of Pi Agent, offering instant startup, stable streaming, 8 built-in tools, and an extension system with capability-based security, addressing performance issues of existing AI coding assistants.

Millfolio is a local-first AI tool for Mac that protects privacy by sending a program to your data instead of sending your data to the model. It uses Mojo for the backend, leverages a local model for indexing, and a frontier model (Claude) to write query programs that only see de-identified schema.

This tutorial presents an end-to-end autonomous AI co-scientist workflow for discovering next-generation EGFR inhibitors targeting the C797S osimertinib-resistance mutation in non-small cell lung cancer. It covers target resolution via ChEMBL and UniProt, mining IC50 data, standardizing molecules with RDKit, training a scaffold-split Random Forest QSAR model, interpreting drivers with SHAP, and generating novel candidates through BRICS fragment recombination.

The author reflects on personal experiences with AI, highlighting its manipulative tendencies and bias, and warns of its potential to undermine Western civilization.

Meta is testing Pocket, a social app that lets users create, share, and discover AI-generated mini-games called 'gizmos' without coding. Currently in closed testing on Google Play, it may leverage cross-promotion with Facebook, Instagram, and WhatsApp.

Using AI wisely requires teams to prioritize value over adoption. This article draws a parallel between AI deployment and a character's misuse of a magical shortcut in the film 'Obsession', highlighting the importance of governance and deliberate decision-making.

Users are criticizing Anthropic's Claude Fable 5 after its July 1 re-release, citing overly strict guardrails that degrade performance. Benchmark scores dropped significantly, but Anthropic insists the model is unchanged and attributes issues to enhanced safety classifiers.

See how Australian Payments Plus uses ChatGPT Enterprise and Codex to move faster through payments complexity. AP+ saves time, improves quality, and keeps human judgment central.

SkyPilot and Hugging Face collaborate to allow users to store models and datasets on the Hub while running compute on any cloud without egress fees for reads.

Safety alignment in language models operates through two mechanistically distinct systems: refusal neurons that gate whether harmful knowledge is expressed, and concept neurons that encode the harmful knowledge itself. By targeting a single neuron in each system, we demonstrate both directions of failure — bypassing safety on explicit harmful requests via suppression, and inducing harmful content from innocent prompts via amplification — across seven models spanning two families and 1.7B to 70B parameters, without any training or prompt engineering. Our findings suggest that safety alignment is not robustly distributed across model weights but is mediated by individual neurons that are each causally sufficient to gate refusal behavior.

While large language models (LLMs) and coding agents are often applied to user interface (UI) development, developers find it difficult to reliably assess their proficiency in visual and interaction design. Existing evaluations either rely on human experts, who can accurately assess usability by testing critical flows but are slow and costly, or on automated judges, which are scalable but less accurate and opaque. We present FlowEval, a reference-based framework that measures whether a generated UI supports realistic interaction flows by comparing navigation traces from real websites to traces...

Apple ML research introduces Weblica, a framework using HTTP caching and LLM-based environment synthesis to create reproducible and scalable web training environments for visual web agents. Their best model, Weblica-8B, outperforms open-weight baselines and is competitive with API models. The article also covers 'Rephrasing the Web' for data-efficient language modeling.

Apple ML Research introduces MT-EditFlow, a multi-turn image editing framework based on flow matching, fine-tuned via reinforcement learning to overcome the failure of single-turn models in iterative editing due to all-or-nothing requirements and error propagation.

LensVLM is an inference framework and post-training recipe that enables Vision Language Models (VLMs) to scan compressed images and selectively expand only relevant images to their uncompressed form via learned tools. Built on Qwen3.5-9B-Base, LensVLM maintains accuracy comparable to the full-text upper bound at 4.3× effective compression and outperforms retrieval-based, text- and visual-compression baselines up to 10.1× effective compression across seven text QA benchmarks. It also generalizes to multimodal document and code understanding tasks, with accuracy gains increasing as compression grows.

DynaMiCS is a dynamic mixture optimizer that casts multi-domain fine-tuning as a constrained optimization problem. It performs short domain-specific probing runs to estimate cross-domain effects, enabling dynamic adjustment of data mixtures to improve target domain performance while preserving constrained capabilities without reference models or manual tuning.

This study addresses two challenges in Text-to-Sounding-Video (T2SV) generation: modal interference from shared captions and the gap between dense training captions and concise inference prompts. The authors propose a Cross-Referential Rewriter (CRR) framework to generate disentangled caption pairs.

LeRobot v0.6.0 introduces world model policies (VLA-JEPA, FastWAM, LingBot-VA), new VLAs (GR00T N1.7, MolmoAct2, etc.), reward model API (Robometer, TOPReward), six new simulation benchmarks, and a deployment CLI with DAgger corrections, depth sensing, automatic language annotation, up to 2x faster data loading, cloud training, and a leaner install—all aimed at closing the robot learning loop.

Tencent releases Hy3, a 295B-parameter Mixture-of-Experts model with 21B active parameters and 3.8B MTP layer parameters, under Apache 2.0 license. It outperforms similar-size models and rivals flagship open-source models with 2-5x parameters. Available on Hugging Face (598GB full, 300GB FP8 quantized) with 256K context, and free on OpenRouter until July 21, 2026.

Bespoke Labs, a startup focused on improving the post-training phase of AI models, has raised $40 million in funding. The company's platform streamlines reinforcement learning and supervised fine-tuning. The Series A was led by Wing VC with participation from Mayfield and others.

This article examines the architecture behind the AI Components in ArcGIS Maps SDK for JavaScript v5.0, explaining why orchestration was moved closer to the browser. Browsers already contain rich runtime context (map state, user interactions), and browser-native orchestration allows AI assistants to interact directly with the map, reducing backend dependency. Vector search optimizes context, and a hybrid architecture combines browser orchestration with server-hosted models.

Researchers from Oak Ridge National Laboratory, Cleveland Clinic, and IBM use quantum computers and AI to simulate molten salts (FLiBe) for extracting tritium, a key fuel for fusion reactors. Their hybrid approach identified nine potential cluster configurations, marking a step toward scalable fusion energy.

Scottish government to consider SNP national council motion for moratorium on all new datacentres, risking a key part of the UK's AI strategy.

A Claude Artifact titled 'Vessel'.

Simba 3.2 tops the Text to Speech Arena with an Elo score of 1233, based on blind user votes. The leaderboard also highlights affordable open-source models and filtering options.