Professional AI media and industry service platform
DeepSeek releases V4.1-Flash-1017, a 763B-parameter model that improves coding benchmarks by 35%, reduces per-task cost from ¥3.3 to ¥1.3, and enables 'wish-style programming' for building games and 3D scenes with minimal prompts, while launching a paid search API at ¥0.005 per request.
Stanford's Christopher Manning argues that while LLMs excel at linguistic form, they still falter at pragmatics, discourse, and dialogue — revealing linguistics' irreplaceable role in probing whether models truly understand meaning beyond pattern abstraction.
Researchers propose RAVEL, a training-free method to evaluate vision encoders for multimodal LLMs by measuring cross-modal nearest-neighbor consistency with stabilized PCA whitening and patch-level matching, achieving 0.85 Spearman correlation across 210 models, outperforming prior methods by 36-101%.
This deep-dive traces OpenAI's Codex from a failed cloud launch to becoming the foundation of ChatGPT's desktop app, revealing how Tibo's team pivoted to a local-first, editor-less agent architecture, leveraged reasoning models and bootstrapping, and achieved 40M+ weekly users in months.
STILL introduces a content-aware intra-layer hybrid attention mechanism that routes tokens to Softmax or linear attention based on self-saliency scores computed locally, preserving pre-trained norm structure via NP-Map, achieving 80% memory reduction and 83% prefill speedup with only 0.04B training tokens.
FlowMAS, a NeurIPS 2026 paper, uses Generative Flow Networks with curiosity-driven and information-theoretic rewards to automatically generate diverse, efficient multi-agent workflow topologies, outperforming baselines across six benchmarks with far lower computational cost.
Apple and CMU researchers introduce SSR, a framework that replaces free-form reasoning in multimodal search agents with selection from six predefined reasoning options, cutting per-step latency over 90% while matching task success rates across seven benchmarks.
Nanjing University spin-off SinapisAI introduces Collective Model Intelligence, a framework that coordinates diverse models via capability profiling and dynamic routing (Mio), achieving top average scores and lowest costs across nine benchmarks, positioning models as the fourth AI growth factor beyond compute, algorithms, and data.
Tadpole introduces an autoencoder-based foundation model for 3D partial differential equations, pretrained via reconstruction on massive online-generated simulation data (202 TB equivalent) without storing it, enabling zero-shot reconstruction, dynamics prediction via lightweight adaptation, and generation across diverse 3D physical systems and resolutions.
SpecTr-GBV unifies multi-draft generation and greedy block verification into a single speculative decoding framework using optimal transport, achieving optimal expected acceptance length under conditionally independent couplings and outperforming SpecTr and GBV across five benchmarks with higher block efficiency and speedup.
