Robots Versus Humans

Autonomous AI agents are now operating across the web to perform tasks independently. We look at what this means for how we use the internet, and the real costs of building systems that think for themselves.

Adobe develops an AI agent that picks the right editing tools like a professional would. Google builds a video agent that watches and helps only when needed. OpenAI merges ChatGPT modes and moves to token-based pricing. Plus stories on Chinese AI alignment, coding benchmarks, and legacy code rewriting.

An autonomous AI agent has selected and administered gene therapy to a live mouse's brain, marking the first time an AI has independently chosen a treatment and delivered it to a living mammal. Also today: Google's forest mapping AI, Meta's Muse agents, and AI chips running in space.

Waymo announces its next robotaxi will be purpose-built at a Toyota factory in Japan, while a startup tries to mass-produce AI data centers like cars. Plus: Wikipedia's legal fight over AI training, Microsoft's synthetic test data, and OpenAI's continuing deluge of mathematics.

Mistral says its Le Chonk model has caught up with DeepSeek, Anthropic files for an IPO at a stunning valuation, and Amazon expands what Alexa can do with your accounts and shopping.

OpenAI claims its latest models can help modernise aging COBOL systems. Anthropic files for its IPO. Toyota commits billions to humanoid robots. Meta plans a massive stock sale to fund AI infrastructure. And Fanatics uses AI to boost customer engagement.

Nvidia open-sources a data engine that makes robot training vastly cheaper. Anthropic says single AI agents now outperform multiple sub-agents. And new speech models show human-like conversation is getting better.

Block releases a speech model trained on human audio to recognize emotions and sarcasm. OpenAI stops training on the live internet. And a new study questions whether expensive coding agent frameworks actually deliver better results.

Anthropic's Claude generates formal proofs that code works correctly, catching bugs in open-source libraries that testing had missed. Meanwhile, AI agents invent games when left alone, inDrive plans robotaxi tests in Kazakhstan, and people are using language models to clean up their smart TVs.

NVIDIA launches RTX Spark, combining gaming, creative and AI computing in a single superchip with 128GB of unified memory. Plus, Anthropic cuts token theft, Amex builds guardrails for AI agents, and France's agentic AI ecosystem scales fast.

A new generative world model you can try for free, token prices keep falling while GPU costs hold steady, and mathematicians say OpenAI's latest mathematical results are unprecedented. Plus local AI, consumer opportunities, and AI agents building simulations.

Microsoft releases a new AI model for classification tasks, Google's unreleased Gemini 4 Carbon reportedly matches rival coding performance, and Anthropic's Claude can now orchestrate thousands of AI agents in parallel. Plus the latest on AI agents, cybersecurity applications, and consumer spending.

Sophos uses AI to slash cybersecurity investigation time, Claude Science creates the first ultraviolet map of the entire sky, and Microsoft 365 Family subscribers gain access to shared AI benefits. Today's stories show AI making real gains in security, science and productivity.

Google launches a Gemini agent for enterprise work, Nvidia expands safety systems for physical AI, and Microsoft debuts new hardware designed to run AI locally. Plus: the race heats up in consumer AI agents.

Mistral releases a freely available trillion-parameter model to compete with closed systems, Google and Meta invest in AI biomedical research, and a startup uses AI to automate food delivery coordination. Plus advances in math benchmarks, motorsport optimization, game creation and AI detection tools.

Mistral releases its largest model yet to compete with US and Chinese frontier labs. Plus, Pinterest converts beauty content into action plans, researchers expand an AI architecture across multiple fields, and telecom operators embrace open models for critical infrastructure.

Enterprise AI moves from forecasting to autonomous decision-making. Plus breast cancer screening gets AI support, Google Research explores AI across biology and infrastructure, and NASA releases a lunar science model trained on 17 years of satellite data.

Mountain rescue teams in the Pacific Northwest are testing powered exoskeletons in real-world conditions. Plus, Meta opens up AI hardware to DIY builders, AI agents struggle to verify their own 3D reconstructions, and DeepMind researchers rethink what artificial general intelligence might actually look like.

Suno expands into speech generation, NVIDIA releases local AI hardware, and Black Forest Labs launches Flux 3 Image with advanced editing. Plus developments from Microsoft, Amazon, and OpenAI's new ultrafast model.

AI conquers another board game stronghold while researchers build cheaper mind-reading tools from brain scans. Also: the economics of AI factories at megawatt scale, and new funding for AI agents, hearing tech, and satellite computing.