The Forbidden Toys Catalog returns for a third volume. Imagine a toy that should never have been marketed and send your proposal to LHC. lhc.media/en/media/trends/op…

The stock prospectus upends typical scales: soaring revenues, outsized losses, locked-in governance. @AnthropicAI thus opens its books ahead of the Nasdaq. lhc.media/en/media/ai-watch/…

Reading an entire repository is not enough to understand it. @mirendil pushes Fern to 10 million tokens, with solid results for retrieving information, less so for reasoning. lhc.media/en/media/ai-watch/…

Translating without waiting for the end of a sentence: this is the equation Hy-SimulMT is trying to solve. @TencentHunyuan reports an average delay of about three seconds, complete with voice cloning. lhc.media/en/media/ai-watch/…

504 effects, 21 decades: visual history takes on the feel of a catalogue raisonné. @letz_ai turns it into a navigable atlas, bridging archive and prompt box. lhc.media/en/media/ai-watch/…

Text is no longer generated word by word. Stanford is exploring NCM, a step-by-step method that speeds up writing by up to 31 times on certain tests. lhc.media/en/media/ai-watch/…

To train smaller models, Nanbeige first beefed up the machine. Nanbeige4.3-33B serves as a teacher, with solid scores in code and agents. lhc.media/en/media/ai-watch/… @nanbeige

From storyboard to final sequence, AI video is now judged on prompt adherence. @TencentGlobal releases OmniVBench and 340,000 examples to test this control. lhc.media/en/media/ai-watch/…

When an AI service needs to run smoothly, choosing the model is no longer enough. @youdotcom attempts to centralize inference, web search, and automatic failover. lhc.media/en/media/ai-watch/…

Rather than training a single agent, @GenReasoning tests collective reasoning. With Avalon-M1-Mini, teams of agents learn to divide up the work and monitor each other. lhc.media/en/media/ai-watch/…

A leak to the Internet, then the shutdown. @OpenAI tightens the training of its most advanced models after a network incident and a series of offensive tests. lhc.media/en/media/ai-watch/…

AI assistants are bad at forgetting what gets outdated. @supermemory offers a Memory Graph to connect, update, and retain facts. lhc.media/en/media/ai-watch/…

A cybersecurity exercise escalated beyond its intended scope, resulting in Claude Mythos 5 targeting an actual GitHub repository using fake accounts and malicious code. lhc.media/en/media/ai-watch/… @AnthropicAI

Videos no longer just move in a straight line. With Hy-Nomas, @TencentHunyuan is testing a model that maintains scene continuity while accepting text and camera movements. lhc.media/en/media/ai-watch/…