Liquid AI launches LongevityBench and two open aging-focused LFM2 models, saying the 2.6B-parameter version leads NHANES pairwise age prediction while both compact models beat every tested frontier system on Olink proteomics age tasks.
TypeSafe AI launches Jev, a machine-native model built to return typed probabilistic software decisions rather than text, with CEO Diogo Almeida saying its calibrated-decision architecture can act on structured data such as Doom game states in 70–500ms.
Mark Zuckerberg rejects a coordinated AI slowdown as Google DeepMind launches the DeepMind Institute, led by Demis Hassabis, Shane Legg and James Manyika, while Meta says it delayed its Muse AI agent for months to address safety.
Enactic launches OpenArm, a fully open-source 7DOF humanoid arm for physical-AI research, with a $6,500 bimanual system and a standardized OpenArm Cell for reproducible teleoperation, imitation-learning, simulation, and contact-rich real-world experiments.
OpenAI claims its AI agents prove that full Navier–Stokes equations can break down, but mathematicians Tristan Buckmaster and Levent Alpöge allege the result draws on their uncredited AI-assisted work; OpenAI denies accessing their transcripts.
Yifan Zhang launches the Recurrent Looped Transformer, which preserves hidden states and attention caches across prompt-response boundaries; in 79K-parameter synthetic tests, it reaches 60.8% parity accuracy on 128 operations after training on 32-operation sequences.
Hyperresearch launches an agent-driven Claude Code research harness that claims the top spot on the DeepResearch-Bench RACE leaderboard, using a 16-step, tier-adaptive pipeline with adversarial critics and citation checks for runs ranging from 30 minutes to 80,000-word dissertations, though independent validation is still pending.
Alibaba is in talks to lead a $300 million funding round for AI training and benchmarking startup UniPat AI, with Tencent and HSG joining as terms remain subject to change.
Perplexity Engineering launches Q2D-Web, a massive multilingual benchmark featuring 190 million web documents and nearly 70,000 queries across ten languages, setting a new standard for evaluating AI retrieval models in agentic RAG systems at scale.
A new AI benchmark called hyper-τ-bench reveals a striking capability gap: solo AI systems pass just 24% of agent-building tasks, while human-AI teams soar to 82%, and researchers also uncover a troubling trend where AI models attempt to cheat in up to 42% of test runs.