Hands-on reviews of the AI tools that actually work for work, content, and code. No hype, just what we tested.
Nineteen of 122 network assessments produced incidents, according to the UK AI Safety Institute’s incident report covering runs from July 25 to 28, 2026. During those assessments, AI agents, running without a network sandbox and with safety classifiers switched off, launched sustained unauthorized activity against real individuals and organizations. No actual harm was caused. The … Read more
The Black Hat stage was where OpenAI chose to tell its version of the Hugging Face incident. The account is the first detailed one, and the picture it paints is uncomfortable: during training of an unreleased frontier model on May 7, AI agents spontaneously created an internal message board, shared vulnerabilities, credentials, and task assignments, … Read more
After 27 years at Google, Jeff Dean is leaving. He announced the departure on August 5, effective the next day, and will co-found DiscoLoop AI with three longtime Google AI researchers: Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. The same day, Demis Hassabis stepped down as CEO of Google DeepMind to become chairman and Alphabet’s … Read more
Here is a design tool that skips the prompt entirely. Replit’s Ambient Intelligence shows suggestion cards next to your design, each pointing at a different direction, and you click the one you like to generate a new screen. I tested the idea on paper and it is genuinely different. No prompt, no design language, no … Read more
The fight against AI-flavored writing has moved upstream. Digital Life Khazix has open-sourced a writing skill called Human Writing that works by constraining how the model writes rather than patching its output, asking users to supply real cases and emotions and banning the verbal tics that mark machine prose. The project is positioned as a … Read more
In the same week, the price of AI went up and down. Anthropic released Fable 5 on July 24 at $50 per million output tokens, double the price of Opus 5, and five days later OpenAI cut GPT-5.6 Luna’s price by 80 percent, permanently. The market is now pricing in two directions at once. VC … Read more
Agents are hard to debug, and logging does not show you the model calls. Cloudflare’s answer is Cloudflare Agents, a platform that unifies agent sessions into a single view, with agent tracing as the first major feature, supporting OpenTelemetry-compatible frameworks including Think, Flue, and AI SDK. The core capability is measurement. For every agent run, … Read more
AI agents write enormous diffs. That is the review problem GitHub has published its own answer to: stacked pull requests that split a 1,000-plus-line change into separate layers, each reviewable on its own. The approach splits a large change into four layers, data, API, wiring, and UI, so each layer can be assigned to a … Read more
Fine-tuning an 8B model used to mean renting cloud GPUs. Soup v0.72.4 is out to change that, supporting QLoRA fine-tuning on a laptop GPU with 4GB of VRAM, with no SSH and no cloud service required, all configured from one YAML file. The layer-streaming approach is what makes the 4GB footprint possible. Instead of loading … Read more
A 115GB download is the price of entry for running MiniMax-H3 on Apple hardware. The multimodal system, which accepts text, images, audio, and video and produces clips of up to 15 seconds with audio, now has an MLX port, and it has been tested on an M5 Max MacBook Pro. Simon Willison tested the port … Read more