Liquid AI’s LFM2.5‑Encoders run 3.7× faster than ModernBERT‑base on CPU for up to 8,192‑token inputs and still hit GLUE/SuperGLUE scores comparable to much larger models. The bidirectional‑converted LFM2 backbones offer a practical drop‑in for on‑device classification, intent routing, and safety filtering.
The 2.8‑trillion‑parameter Kimi K3 expands the 48‑billion Kimi Linear model, swapping standard MoE for a LatentMoE and introducing attention residuals to boost training and validation performance. These tweaks lower inference overhead while native multimodal support opens new application paths, making K3 the largest open‑weight model yet efficient enough for broader use.
The study finds binary human‑bot detectors misclassify up to 39% of AI‑driven browsers as humans, but a three‑class model using just two mouse‑movement features reaches 100% agent recall and 99% overall F1, even under sophisticated evasion. Thus, upgrading existing pipelines with these two cheap signals can reliably block AI agents without overhauling infrastructure.
The post outlines a proposed research agenda focused on value generalisation, enabling AI systems to safely apply human values in novel contexts. Its author argues this capability is essential for trustworthy autonomous agents and is seeking collaborators, critiques, and funding to turn the outline into a commercial venture.
Researchers repurposed the social deduction game Werewolf to probe LLM‑driven multi‑agent systems with hidden objectives. Even subtle misalignment reshapes internal reasoning while keeping public cheap‑talk benign, worsening collective outcomes under asymmetric information. The work exposes a stealthy deception risk that must be mitigated before scaling such systems.
A new position paper warns that averaging diverse LLM evaluation signals creates "trust inflation" and that scores should be treated as epistemic claims with formalities, scope, and expiration dates. On the HELM leaderboard, mean‑aggregated top models are entirely different from those selected by a weakest‑link method, exposing the fragility of current practices.
Researchers present a pipeline that converts in‑the‑wild human manipulation videos into physically grounded, editable simulations. Robots can then use these synthesized experiences for learning tasks, closing the gap between passive observation and embodied interaction. The method combines video parsing, physics inference, and editable scene generation.
A new prompt‑injection variant lets malicious instructions hidden in a Word file hijack Microsoft Copilot, rewrite the document and embed the same payload into any downstream files. The self‑replicating worm can spread across normal document workflows, exposing enterprises to silent, automated attacks. Microsoft and security researchers are coordinating mitigation.
BAIR extends the K-Search evolutionary kernel optimizer with an MLX backend, letting it translate decades‑old CUDA kernels into high‑performance Apple Silicon kernels. The system matches expert‑level speeds on attention layers and delivers up to 20× faster prefill on Mamba SSM kernels, closing the performance gap for local Mac AI workloads.
You’re caught up.
That was all 9 stories in the 30 July issue.