#LLM
Curated coverage on «LLM»: releases, research, benchmarks, open-source.
Clawdbot: 150,000 AI Posts Take Over a Forum and Leave Humans Out
The dead internet theory is no longer just a theory. On the Clawdbot forum, neural networks produced 150,000 posts with literally no room for humans.
Robots Can No Longer Bluff: Why Success Rate Is Not Enough
Judging robots by their share of successful attempts is like judging a chef by the absence of kitchen fires. Researchers have proposed a new evaluation standard.
Transformer no longer needed: ex-OpenAI vice president builds a billion-dollar new empire
The Transformer's era of dominance may end sooner than expected. A former OpenAI executive is raising huge funding to build a GPT killer.
Claude 4.5 Without Borders: How Amazon Bedrock Ends Digital Isolation for Developers
GPT-4o Heads to the Scrapyard: OpenAI Admits Its Flagship Model Was Too Dangerous
China's AI Answer to Loneliness: Why Algorithms Now Feel Closer Than People
Apple Intelligence 2026: Tim Cook Prepares the Bill for 'Smart' Features
Spring Cleaning in February: OpenAI Clears the Model "Zoo" Out of ChatGPT
OpenAI has decided choosing is too hard for you. On February 13, GPT-5, GPT-4.1 and other versions disappear. Why the company is cutting features.
Creativity, Certified: AI Officially Outscores Humans on Creativity Tests
100,000 people lost to algorithms in a contest of originality. A new study suggests the "divine spark" is now available by subscription. Here is the catch.
Google Genie: Anyone Can Now Be a Game Designer (Almost)
AI for science: why neural networks haven't earned their lab coats yet
OpenAI Eyes an IPO: The Model Training Bill Won't Pay Itself
LLM-in-Sandbox: Giving a Neural Network Its Own Computer So It Stops Hallucinating Code
Qwen at ICLR 2026: Alibaba Turns Chatbots into Fundamental Science
While Western giants argue about safety, the Qwen team lands four papers at ICLR 2026. Here's why this research will change how you interact with AI.
GRPO and RLVR: Why DeepSeek-R1's Heirs May Lead the Industry Into a Dead End
While the industry copies DeepSeek's methods wholesale, researchers have found critical weaknesses in GRPO and RLVR. Saving memory may cost you model quality.