Latest AI News
What the machines did today, and why you should care.
AIOpenAI Says New Tools Are Coming for Online Child Abuse. Details: Pending
OpenAI announced new tools, guidance and partnerships to prevent, detect and respond to online child sexual exploitation and abuse. The announcement is short on specifics, so here is what is actually on the record.

AI Can't Tell Your Frosting Problem Is a Cupcake Problem
XYBench, from Hugging Face and Allen AI, packs 8K+ questions that test whether assistants fix the real problem or just answer the dumb one. It just won Best Resource Paper at EMNLP 2026.

IBM's new markup language is pronounced "lemon" and it wants to fix LLMs
IBM built LLMON, a tagged format that tells models what each chunk of context actually is. It ships inside Granite 4.1 at 3B and 8B, aimed at brittle small models and prompt injection.
AISingapore Airlines' Kris chatbot doubled its CSAT. Humans, take note.
Singapore Airlines put OpenAI models behind its Kris chatbot and says customer satisfaction scores doubled. The same tech is also quietly helping employees get their work done.
AIGoogle's Gemini 4 Carbon is out, and it wants your hardest work
Google DeepMind calls Carbon its strongest model yet, with frontier-level cybersecurity and long-horizon agent skills. Everyone from developers to consumers can try it today.
AI256 Experts Per Layer, 8 Working: Amazon Wants to Fire the Idle Ones
Amazon and Hugging Face propose HOPE, a way to permanently cut the least useful experts from MoE models. In Qwen3.5-122B, expert MLPs hog 95% of parameter memory while most of them nap.
AIGoogle's video model now runs 15 seconds, 30% cheaper
Gemini Omni 1.2 Flash stretches clips from 10 to 15 seconds and drops to $0.07 per second at 720p. Cheaper robot film crew, same carbon-based audience.
AIOpenAI hires Hacktron to break OpenAI, on purpose
OpenAI is bringing in Hacktron to hunt for holes in its own systems and poke at its third-party connections. The plan: use frontier AI to harden the defenses.
AINVIDIA drops four 16B world-sim models, and robots get a dream sandbox
NVIDIA put four new open-weight Cosmos 3 checkpoints on the table, each with 16 billion trainable parameters. They cover driving scenes, sensor streams and robot arms, under a license that allows commercial use.
AIA 2B-Parameter Model Wants Your Invoices, Not Your Job
Hugging Face's Dharma Field Extractor pulls the exact fields you ask for out of any document and hands back JSON. The team says it beats every frontier LLM it tested, at a fraction of the cost.
AIGoogle's lab agent drafts new drug candidates in about an hour
Google DeepMind taught its Antigravity agent some chemistry. In its lung cancer demo, the loop that normally eats days runs in roughly an hour, and the carbon-based chemists get to supervise.
AIOpenAI built a test for the tasks you actually hate doing
LifeVal grades AI on childcare searches, money management and government paperwork. Finally, a benchmark that measures the stuff that ruins your Tuesday.