Today's Hallucination HQThe Gap Is Closing, and Silicon Valley Is Pretending Not to NoticeMoonshot AI's Kimi K3 and Alibaba's latest Qwen model have arrived, both claiming to match OpenAI and Anthropic's finest — at considerably lower cost. China's top labs are releasing frontier-grade models at a pace that suggests "American AI dominance" may soon require quotation marks. Whether the benchmark claims hold up to scrutiny is another matter entirely, but the pattern is hard to dismiss: the world's most expensive AI race now has very fast company. Source: The Verge
YouTube Draws a Line Between "Content" and Whatever That Other Stuff IsYouTube has updated its monetisation rules to formally define the categories of AI-generated and low-quality video that will no longer earn ad revenue. In practice, this means the endless conveyor belt of AI-voiced listicles, recycled footage compilations, and suspiciously identical "relaxing" videos may finally stop paying quite so well. YouTube calls it a policy clarification. Everyone else calls it a Tuesday. Source: TechCrunch
Turns Out AI Didn't Need Our Help to Be PrejudicedNew research suggests AI hiring tools don't just inherit human biases from training data — they develop fresh ones entirely on their own. Researchers found that large language models can form novel prejudices with no human template to blame, which rather undermines the industry's favourite defence. Your CV may now be judged by a system that is both statistically unreliable and, in a technical sense, self-taught in the art of unfairness. Encouraging stuff. Source: MIT Technology Review
MCP Gets a Quiet Upgrade That Actually MattersThe Model Context Protocol — the standard that lets AI agents connect to external tools and data sources — has been updated to use a "stateless" approach to session management, meaning servers no longer need to track individual sessions continuously. It works more like a normal website now, which makes it significantly easier to build and deploy. Unglamorous? Absolutely. But plumbing upgrades matter more than anyone admits until the plumbing stops working. Source: TechCrunch
Someone Found the Weakness in AI Agent Teams, and It's the ManagerResearchers have identified a new attack method called PlanFlip, targeting multi-agent AI systems at their most vulnerable point: the Planner — the model that breaks big goals into smaller tasks for other agents to execute. Inject something malicious at the planning stage, and every agent downstream follows corrupted instructions without question. It is, in essence, the world's most obedient middle management. The paper is a useful reminder that "multiple AI agents working together" is only as trustworthy as whichever one is in charge. Source: ArXiv AI
Until tomorrow — may your prompts be specific and your planners uncompromised.
|