Technology & AI
AI agent runs vertical farm, cuts grow cycle 35%
An arXiv preprint shows LLMs autonomously controlling farm lighting and actuators, cutting grow cycles 35% and finding energy strategies humans missed.
AI Models
165 stories
Technology & AI
Reinforcement learning cuts AI training power violations 89%
An RL controller for GPU power cut violations 89% and boosted energy efficiency 26% in LLM training, then failed at 72B scale before a rebuild fixed it.
Technology & AI
AI image models fingerprinted without watermarks
A new arXiv preprint exploits 'collapsed generation' to verify ownership of text-to-image diffusion models via API access, surviving fine-tuning.
Technology & AI
Liquid AI's 3B vision model jumps 54% on grounding
LFM2.5-VL-3B, released August 12, pairs a 400M vision encoder with a 2.6B text backbone for on-device AI, but speed and edge claims stay unproven.
Technology & AI
Self-evolving GUI agents improve click accuracy 7.4% after deployment
A new arXiv preprint proposes a framework letting deployed GUI agents improve click accuracy 7.4% without human labels, with stakes for automation teams.
Technology & AI
AI models lose over 90% of safety signal in African languages
AI safety alignment in four African languages retains under 10% of English refusal signal and leaves those speakers without model guardrails.
Technology & AI
NVIDIA Magpie TTS expands to 12 languages with open weights
NVIDIA's Magpie TTS v2607 adds Arabic, Korean and Brazilian Portuguese to its 364M-parameter open-weights model, targeting low-latency voice agents.
Technology & AI
KnowPlan AI agents plan degrees with 99.5% certified accuracy
KnowPlan separates curriculum extraction from degree-pathway optimization with a hard boundary, cutting source access 47% while certifying 99.5% of plans.
Technology & AI
Self-evolving AI agents stumble under real task streams
New AgentStream benchmark from Microsoft and Chinese Academy of Sciences finds self-evolving AI agents behave unpredictably in task sequences
Technology & AI
Google Cloud scanner catches AI safety tampering in 10 of 14 models
AMS, a new Google Cloud tool, flags 71% of safety-training modifications across Llama, Gemma, Qwen and Mistral, but behavioural fine-tuning evades it.
Technology & AI
LLM agent: code-only verification flips goal abandonment 100% to 0%
New arXiv preprint: a deterministic executive owns all agent belief, the LLM only files proposals, and zero ARC-AGI-3 completions are honestly disclosed.
Technology & AI
NVIDIA Cosmos 3 open model combines three physical AI skills
NVIDIA's open-weights Cosmos 3 model combines vision reasoning, world generation and action prediction for robotics and autonomous vehicles.