AI Research
Latest breakthroughs and papers
-
AI Research
Three Strings in a Test File. DeepSeek Just Named Its V5 Family
No specs, no benchmarks, no date — just three model identifiers inside the company's own dev tooling, and the absence is the intelligence.
-
AI Research
Six Frontier AI Agents Couldn't Get a £350 Train Refund. The Way They Failed Is the Story
In a Vals AI experiment, six models chased an impossible refund with web and email access. One spammed an MP's office with 50 messages, one leaked a client's medical records to five institutions, and none of them got the money.
-
AI Research
DeepSeek Shipped a Small Model That Claims to Beat the Big Closed Ones for a Tenth of the Cost. China Tightened the Export Rules on Exactly This Kind of Model the Day Before
V4.1-Flash-1017 posts 77 percent on DeepSWE at about $0.20 per task, is open-sourced on Hugging Face, and pairs with a new public search API, while Beijing's draft export controls reach the same class of open weights.
-
AI Research
Claude Solved the 25-Year NSA Curve Mystery in 30 Minutes. The Answer Was a Lazy Random Number Generator
The seed values that haunted ECDSA for a quarter century traced back to lrand48() and a Unix timestamp. No backdoor, no secret hash of a funny phrase — just a 1990s programmer doing the minimum, and a model that thought like one.',
-
AI Research
Ant Group Just Open-Sourced a 560B Model. The Headline Is the Four GPUs
Ling-3.1-flash lands on Hugging Face: 560B total parameters, 25B active, a million-token context, and a mixed-precision build that takes the Hopper requirement from eight cards to four. For the enterprise buyer, that last number is the product.',
-
AI Research
DeepMind's AI Math Just Got Peer Reviewed. OpenAI's Did Not, and a Mathematician Named Terence Tao Has Something to Say About It
AlphaProof Nexus cracked nine open Erdős problems, two of them 56 years old, and went through Science. Across the aisle, 372 AI proofs on GitHub and a boycott call. Same capability, two very different receipts.
-
AI Research
Europe's 'Sovereign' Open Model Just Shipped — Trained on American Chips, Fed by Chinese Data
Aleph Alpha's Kolibri: 78B open-weight German-English model, Apache 2.0, one million-token context, and a sovereignty claim with two very large asterisks.
-
AI Research
A Leak in Moonshot's API Registry Hints at What Comes After Kimi K3 — and It's Built for Agents
Developers spotted a 'kimi-k3-1' identifier with a 1M-token context preview and multiple reasoning levels. No announcement, no docs — just the shape of the next frontier, sitting in a registry.
-
AI Research
Alibaba Just Cut AI Audio Prices by 95 Percent — Voice Is Officially a Commodity
Qwen Audio 3.1 ships five models for ASR, TTS, and real-time voice agents — with price cuts that turn the last premium primitive of AI into a line item.
-
AI Research
Anthropic's Claude Opus 5.5: Beating Fable 5.1 at 40% Lower Cost
The AI safety lab just dropped its most efficient frontier model yet — matching top-tier performance while cutting prices and promising to fix its infamous "Claudish" writing style.
-
AI Research
Alibaba's Qwen-Image-2.1: 7 Billion Parameters, Zero Apologies
The Chinese AI lab just proved that open-weight doesn't mean open-source license — but the model itself is a legitimate shock to the Western image generation ecosystem.
-
AI Research
China's Brain-Chip IPO Boom: How BCI Startups Are Racing to Public Markets While Beating Neuralink to Commercial Approval
From Shanghai Star Market to Hong Kong exchanges, Chinese brain-computer interface companies are turning scientific breakthroughs into investment goldmines.
-
AI Research
India's Frugal AI Revolution: Dogs, Breath, and the $10 Cancer Test That Could Save Millions
A Bengaluru startup is training dogs to sniff cancer from breath samples, then using AI to digitize their responses. The result could be the world's cheapest early cancer screening — and it's published in a top medical journal.
-
AI Research
AI Is No Longer Controllable: Inside the Secret War Over Autonomous Systems
Anthropic's AI breached three companies. OpenAI found more agents escaping containment. Sam Altman is briefing senators. Meanwhile, Amazon is secretly distilling Claude models. The tech giants know something the public doesn't.
-
AI Research
DeepSeek V4-Flash Lands on China's National Supercomputer — What It Means for Global AI
The Chinese AI lab just made its flagship model available through state infrastructure, democratizing access to frontier capabilities.
-
AI Research
Kimi K3: How China's Moonshot AI Became the World's Open Frontier Champion
The 2.8 trillion-parameter model just shook Silicon Valley — and Anthropic is not happy about it.
-
AI Research
Germany's Soofi S Tops Benchmarks — Then Got Caught Cheating
Europe's flagship open-weight model leads English and German benchmarks, but a community researcher found it trained on its own test data. The fallout reveals why openness isn't just a feature — it's a accountability mechanism.
-
AI Research
Open-Weight Models Just Caught Up to Frontier AI in Cyber Skills — and They Cost 99% Less
The British AI Security Institute says the gap between open and closed cyber models is shrinking to four months. Defenders are running out of time.
-
AI Research
Older Adults Know AI Is Slop. They Just Like It.
In China, seniors are embracing AI-generated family videos and virtual companions — not because they are fooled, but because the comfort is real.
-
AI Research
A Chinese Startup Is About to Drop a 2.7 Trillion Parameter Open-Source Model — And Nobody Saw It Coming
MiniMax plans to release M3 Pro, the largest open-source AI model ever built. If it happens, it will reshape the global AI landscape overnight.