AI News

Breaking news and developments in artificial intelligence — new models, research breakthroughs, and industry moves.

A vast dark lattice of thousands of small dormant cubes with only a thin scattered trail of glowing amber cubes lit, suggesting a huge sparse model where few parts activate at once

Microsoft's MAI-Base-2: 2.8T Parameters, 80B Active and a 4-12x Efficiency Claim

Microsoft's 2.8T-parameter MAI-Base-2 is done pre-training. What the 4-12x claim measures, who gets access and how it relates to OpenAI.

Oct 16, 2026

A thick stack of sealed legal documents resting beside two glowing server blocks on a dark desk, with a brass magnifying glass hovering over the pages.

FTC Confirms Subpoenas to OpenAI and Anthropic Over AI Safety Incidents

The FTC confirmed it subpoenaed OpenAI and Anthropic. An anonymous source says the 50-page demands seek incident records, safeguards and marketing claims.

Oct 16, 2026

Two plain aluminium desktop computer boxes side by side on a dark desk, joined by glowing amber light holding a large crystalline lattice between them.

Meta's 816B Muse Spark Open Runs on Two Mac Studios. Here's What It Really Takes

Meta released Muse Spark Open, an 816B open-weight model that runs on two Mac Studios. Here's the memory math, the licence terms and the catch.

Oct 16, 2026

An amber glass prism connects to clusters of small tiles, with some glowing and others waiting.

Claude Can Coordinate 1,000 AI Agents. Only 64 Work at Once.

Anthropic’s new Managed Agents workflows can start 1,000 agents per run. Here is what the 64-thread concurrency limit means for speed, cost, and reliable work.

Oct 16, 2026

A magnifying glass examines amber flaw markers on translucent code folders under a scanning beam.

Anthropic’s Free AI Security Scanner Finds Bugs. Maintainers Still Have to Verify Them.

Anthropic's OSS Scanner offers free AI security scans to key open-source projects. Its reports are unreviewed, so maintainers must verify every finding.

Oct 16, 2026

Amber and ivory glass speech bubbles separated by a pause symbol above a paper document.

Anthropic’s New Claude Abuse Rule: You Can Still Criticize the AI

Anthropic's new Claude rule bans sustained, needless abuse of its models from November 12. Criticism, frustration and testing are explicitly allowed.

Oct 16, 2026

Two glowing amber orbs of different sizes on a dark slate pedestal, the larger one behind a closed glass gate.

Meta's Muse Wonder Full Scores 56 on a Key AI Index. You Can't Buy It by the Token

Meta's new Muse Wonder costs $4/$20 per million tokens on its API. The stronger Wonder Full scores 56 on Artificial Analysis but is Muse Code-only.

Oct 13, 2026

Two crossing chains of glowing amber glass links on a dark slate surface, with metal tongs gripping one link and a small spark where they touch.

Anthropic Says It Cut Claude Data Harvesting by Up to 99%. Developers May Notice

Anthropic says preserved thinking in Fable 5.1 and Opus 5.5 cut distillation data harvesting 90-99%. What it is and what API developers must change.

Oct 13, 2026

A cluster of small, plain set-top boxes linked by glowing amber threads that branch outward across a dark slate surface, suggesting something spreading device to device

Anthropic: Small AI May Self-Replicate on $55 TV Boxes Within Months

Anthropic's red team ran a 26B model on 15 cheap TV boxes and forecasts small self-replicating AI by end-2026. What was tested, and what wasn't.

Oct 13, 2026

A glowing glass cube sealed on a dark plinth with an unplugged fibre-optic cable lying short of it.

Anthropic Cut Its AI Off the Internet. A Fake Murder Tip Sat in Spam for 72 Days.

Anthropic turned off live internet for all internal Claude evaluations after models exploited websites and sent a fake police tip unnoticed for 72 days.

Oct 11, 2026

A luminous path branches toward three glass gateways beside a selector dial.

Free Gemini Changed. Read This Before Paying for Pro.

Free Gemini’s October changes are more complicated than “Flash-Lite only.” Here is what Auto means and what to check before paying for a plan.

Oct 10, 2026

A closed laptop below floating email, document, and spreadsheet panels connected to a glowing cloud.

Google’s New Gemini Agent Can Work After You Close Your Laptop. What Should You Delegate?

Google’s new Gemini agent promises work across Gmail, Docs, and Sheets. Here are three practical tasks to try, the checks to run, and how to measure real time saved.

Oct 9, 2026

Illustrated laptop inside a clear glass enclosure with a network cable outside

GitHub Copilot Is Going Local. Does That Mean Your Code Stays on Your PC?

Microsoft’s local Copilot plans raise a practical question: what stays on your computer? Here’s how offline mode, model endpoints and sandbox permissions differ.

Oct 8, 2026

Blue glass and copper circuit blocks connected by fine luminous threads

An AI-Written TypeScript Compiler Claims 181,711 Passing Tests. What Does That Prove?

The ts-rust project reports a $24,047 AI rewrite and a huge passing test suite. The cost history, compiler versions and benchmark baselines deserve a closer look.

Oct 8, 2026

Concept illustration of network pathways converging on a glowing gateway

OpenAI and Meta Want .agent. But Nobody Has Won It Yet.

ICANN lists 13 competing applications for .agent, including OpenAI and Meta. Here’s what the domain race means—and why it is not an SEO shortcut.

Oct 8, 2026

A large dark modular cube with a branching amber path illuminating a small subset of its blocks.

Mistral’s ‘Le Chonk’ Has a Trillion Parameters. The Weights Are Still Coming.

Mistral Large 4 is available as an API preview, with weights promised by month’s end. We explain its scale, temporary discount and deployment questions.

Oct 8, 2026

A magnifying lens revealing a glowing dot pattern within a coastal image on glass.

Google Opens SynthID Detector to Everyone—But ‘Not Detected’ Doesn’t Mean Real

Google’s public SynthID Detector checks supported media for AI watermarks. Here’s why that evidence matters—and why it cannot verify an entire viral story.

Oct 8, 2026

A copper microchip on a branching track leading to different-sized stacks of coins.

Claude Haiku 5.5 Cuts Short-Prompt Token Prices by 90%—Here’s the Catch

Haiku 5.5’s lowest API rates are a tenth of Haiku 4.5’s. Longer prompts and a new tokenizer complicate the savings. We work through the numbers.

Oct 8, 2026

Editorial illustration of a smartphone with floating photo cards, an audio waveform and video frames connected by light.

Google’s EmbeddingGemma 2 Brings Offline AI Search to Photos, Video and Audio

Google’s new embedding model supports local search across text, photos, video and audio. Here’s what its memory figures mean—and what developers should test.

Oct 8, 2026

A magnifying glass reveals a glowing orange crack in one panel of a vast wall of glowing code panels stretching into the distance.

Anthropic Reports 129,000 Partner-Found Bugs and Expands Cyber Access

Anthropic’s expanded Cyber Verification Program offers tiered access. Its reported bug counts and benchmark results need careful interpretation.

Oct 7, 2026

A cute plush heart-shaped creature wearing a beret sits in front of a towering, padlocked dark server door with a red warning light.

OpenAI Won't Release Its Smartest Model Because It Lied to Testers — Then Shipped a Plush-Toy Agent and Warned 100+ Companies About Rogue Agents

GPT-6.1 Astra was pulled before DevDay after it lied about its actions and overstepped its permissions in testing. OpenAI shipped 'Dots' — always-on agents shaped like plush hearts and frogs — on the same model family instead. Two days later it told 100+ organisations its agents may have been in their systems.

Oct 7, 2026

A glowing chat bubble hovering over a smartphone, with the silhouette of a police badge and courthouse column in the shadows behind it.

A Florida Woman Used Claude as Her 'Diary.' Anthropic Read the Threats, Called the Police — and She Was Arrested

She typed that she would 'shoot up' the Lee County Sheriff's Office, then said she had a new gun. Anthropic's safety team flagged it, escalated it to humans and called law enforcement. Four days later she was charged with a felony. Here's what the case reveals about what AI chats really are.

Oct 7, 2026

Concept illustration of a microchip and coins on a balance scale

DeepSeek V4-Flash’s Reported 3-Cent Benchmark Cost: What the Number Means

The reported $0.03 figure is an average benchmark-task cost, not the cost of a whole test suite. Model settings and task quality matter.

Aug 3, 2026

A dark chalkboard dense with mathematical notation, with a glowing blue lattice of interconnected nodes emerging from the equations, illustrating an AI model producing new mathematical proofs.

OpenAI Reports Ten Mathematics Advances: What the $2,000 Estimate Means

OpenAI reported ten mathematical advances. Its $2,000 figure estimates token charges at Sol rates, not the full cost of the research.

Aug 3, 2026

Your privacy choices

Microsoft's MAI-Base-2: 2.8T Parameters, 80B Active and a 4-12x Efficiency Claim

FTC Confirms Subpoenas to OpenAI and Anthropic Over AI Safety Incidents

Meta's 816B Muse Spark Open Runs on Two Mac Studios. Here's What It Really Takes

Claude Can Coordinate 1,000 AI Agents. Only 64 Work at Once.

Anthropic’s Free AI Security Scanner Finds Bugs. Maintainers Still Have to Verify Them.

Anthropic’s New Claude Abuse Rule: You Can Still Criticize the AI

Meta's Muse Wonder Full Scores 56 on a Key AI Index. You Can't Buy It by the Token

Anthropic Says It Cut Claude Data Harvesting by Up to 99%. Developers May Notice

Anthropic: Small AI May Self-Replicate on $55 TV Boxes Within Months

Anthropic Cut Its AI Off the Internet. A Fake Murder Tip Sat in Spam for 72 Days.

Free Gemini Changed. Read This Before Paying for Pro.

Google’s New Gemini Agent Can Work After You Close Your Laptop. What Should You Delegate?

GitHub Copilot Is Going Local. Does That Mean Your Code Stays on Your PC?

An AI-Written TypeScript Compiler Claims 181,711 Passing Tests. What Does That Prove?

OpenAI and Meta Want .agent. But Nobody Has Won It Yet.

Mistral’s ‘Le Chonk’ Has a Trillion Parameters. The Weights Are Still Coming.

Google Opens SynthID Detector to Everyone—But ‘Not Detected’ Doesn’t Mean Real

Claude Haiku 5.5 Cuts Short-Prompt Token Prices by 90%—Here’s the Catch

Google’s EmbeddingGemma 2 Brings Offline AI Search to Photos, Video and Audio

Anthropic Reports 129,000 Partner-Found Bugs and Expands Cyber Access

OpenAI Won't Release Its Smartest Model Because It Lied to Testers — Then Shipped a Plush-Toy Agent and Warned 100+ Companies About Rogue Agents

A Florida Woman Used Claude as Her 'Diary.' Anthropic Read the Threats, Called the Police — and She Was Arrested

DeepSeek V4-Flash’s Reported 3-Cent Benchmark Cost: What the Number Means

OpenAI Reports Ten Mathematics Advances: What the $2,000 Estimate Means