Daily Digest — 2026-05-19
6 items · 4 research labs, 2 industry media
MarkTechPost: all feed URLs failed (last tried: https://www.marktechpost.com/feed/)AI News: all feed URLs failed (last tried: https://artificialintelligence-news.com/feed/)
🏛️ Research Labs (4)
OpenAI and Dell partner to bring Codex to hybrid and on-premise enterprise environments
OpenAI and Dell Technologies collaborate to deploy Codex in hybrid and on-premises enterprise environments, addressing data security and workflow integration. The partnership integrates Codex with Dell AI Data Platform and explores connections to Dell AI Factory, enabling enterprises to leverage Codex near internal data and systems. This facilitates use cases across software development (4M weekly developers) and knowledge work, including code review, test coverage, and agentic automation, while maintaining enterprise-grade controls.
codexhybrid deploymenton-premisesagentic aienterprise data platform
Fine-Tuning NVIDIA Cosmos Predict 2.5 with LoRA/DoRA for Robot Video Generation
The article presents a parameter-efficient fine-tuning method for NVIDIA Cosmos Predict 2.5, a 2B-parameter video generation model, using LoRA and DoRA adapters for robot manipulation tasks. LoRA/DoRA injects small trainable modules into the frozen base model, reducing memory requirements and enabling fine-tuning on a single GPU. Training employs rectified flow loss, AdamW optimizer, and linear learning rate scheduling. Results demonstrate successful generation of synthetic robot trajectories, with temporal and cross-view Sampson errors used for geometric consistency evaluation. Training achieves decent results in 100 epochs, taking 17 hours on a single H100 GPU.
parameter-efficient fine-tuninglora/dorarectified flowsampson errorrobot manipulation
PaddleOCR 3.5: Running OCR and Document Parsing Tasks with a Transformers Backend
PaddleOCR 3.5 introduces a Transformers backend for OCR and document parsing tasks, enabling seamless integration with Hugging Face-centered workflows. The update provides a flexible inference-engine interface, allowing developers to configure backend-specific options such as dtype, device placement, and attention implementation via engine_config. Supported models include PP-OCRv5 for OCR and PaddleOCR-VL 1.5 for document parsing, managed by PaddleOCR pipelines. This enhancement reduces integration friction for RAG, Document AI, and agent applications, facilitating structured data extraction from diverse document formats. The Transformers backend is particularly beneficial for teams already leveraging PyTorch/Transformers infrastructure.
ocrtransformersdocument parsingraginference-engine
The Open Agent Leaderboard
The Open Agent Leaderboard introduces a novel evaluation framework for general-purpose AI agents, assessing both performance and cost across six diverse benchmarks (SWE-Bench Verified, BrowseComp+, AppWorld, tau2-Bench Airline & Retail, tau2-Bench Telecom). The Exgentic framework standardizes agent-benchmark interactions via a unified protocol, enabling cross-environment comparisons. Results show general agents matching specialized systems (18-29% gap between open/closed models), with agent architecture significantly impacting outcomes (e.g., tool shortlisting improved all models). Cost analysis reveals failed runs incur 20-54% higher expenses than successful ones.
agent evaluationin-context learningbenchmark standardizationtool shortlistingcost-performance tradeoff
📜 arXiv Papers
No new items today.
📰 Industry Media (2)
What to expect from Google this week
Google's 2026 I/O conference highlights its strategic positioning in the foundation model race, particularly in AI coding and scientific applications. Despite trailing behind Anthropic's Claude Code and OpenAI's Codex in coding capabilities, Google DeepMind is focusing on scientific AI tools like AlphaEvolve and AI co-scientist. The company faces internal challenges, including employee protests over a DoD deal, while maintaining strengths in AI-for-science research. Key announcements may include updates to Antigravity agentic coding platform and AI-powered Health Coach, though transformative breakthroughs appear unlikely.
foundation modelsgemini 2.5 proalphafoldagentic codingllm-based health
Inside Anduril and Meta’s quest to make smart glasses for warfare
Anduril and Meta are developing augmented-reality smart glasses for military use, featuring AI-driven target recognition, voice/eye-tracking controls, and integration with Anduril's Lattice software. The system aims to reduce cognitive load by overlaying contextual data (maps, drone positions) and enabling multi-step drone coordination via LLMs (Gemini, Llama, Claude). Prototypes face challenges in ruggedization, local AI processing, and weight constraints. Two variants exist: a helmet-attached version for the Army's SBMC program (2028 target) and Anduril's self-funded EagleEye integrated helmet. Competing projects by Rivet and Elbit highlight the sector's growth, though past failures (Microsoft's $22B cancellation) underscore technical and procurement risks.
augmented-realitylatticeeye-trackingllmsprototyping
Generated automatically at 2026-05-18 20:25 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.