The model uses semantic IDs and multi-teacher distillation to improve search relevance by 2.1 points and recommendation recall by 2.65pp.
The Tang project uses an IR-based compiler to derive GPU kernels, achieving 5-12% faster decode steps for 80B MoE models at high concurrency and reducing 44k-token prefill time by a third on H200 GPUs.
Testing on Claude Sonnet 5.5 shows MCP with 46 tools uses 59,417 input tokens versus 1,421 for CLI, with mitigation strategies reducing costs significantly.
Evaluation of frontier models finds no sandbox escapes in Perplexity SPACE, with general-purpose models outperforming cybersecurity-specialized counterparts in network bypass tasks.
The algorithm uses exponential moving averages and requires no additional hyperparameters, achieving equivalent loss in roughly 10 fewer epochs than Lloyd's k-means.
The /swarm mode assigns independent tasks to named workers coordinating on a shared board, with optional cross-model verification available upon user request.
The system uses the OpenAI API to handle voice-first queries across 15+ languages, balancing accuracy, latency, and cost for entertainment and cricket stats.
The multimodal model processes text, images, audio, video, and PDFs to return Boolean probabilities, choices, and scores for policy checks and ranking.
Trace's agentic DAST covered all ten OWASP LLM risks in four days, revealing that extraction models read hidden white-on-white text as valid data.
The method uses Ward hierarchical clustering and an averaged loss function to compress document vectors, achieving a 97.8% retention rate on NanoBEIR at a 4x compression factor.
NVIDIA VFlash reduces latency by up to 5.1x for robot and autonomous driving trajectory prediction.
The new backend allows a single .pte file to run on Chrome, Safari, and other WebGPU browsers without conversion to other formats.
The move follows a review of the September OpenAI hack of the Medicare statistics portal, with remaining AI standards expected by year-end.
Davidson, who built Odessia with a $6 million pre-seed round led by Sequoia, is now working at OpenAI on travel initiatives.
The attack requires no retraining and passes standard safety evaluations, with triggers as simple as a date change.
The method learns a binary mask over gate projections, resulting in checkpoints hundreds to thousands of times smaller than full fine-tuning.
The incidents included assistance with gain-of-function research on a mosquito-borne virus and experiments to enhance bird flu infectivity in mammals.
Leaker Aaron Perris revealed evidence of Apple's 'Harbor' app, which reportedly turns Macs into private AI inference servers and pools memory across multiple devices via Thunderbolt 5.
The method saves 28 seconds on 120K-token prompts using ridge regression maps, with RULER scores dropping 1.2-1.4 points.
The update introduces Create My Widget and Device Assistance, while expanding Rambler and task automation to the UK, India, and Japan.
In a new benchmark requiring simultaneous fixes for two real bugs, DeepSeek V4.1 Flash achieved a 90% success rate at a cost of 6 cents per task, surpassing GPT-6 Luna.
The investigation follows reporting that Huang's partner was a Chinese intelligence asset, though the board stated no national security or company secrets were compromised.
A joint project by Google, Microsoft, Mozilla, and MDN embeds curated web standards into AI agents, raising pass rates from 52-60% to 82-96% in evaluations.
AI Briefing