Microsoft token metric, OpenAI GPT-Rosalind, NVIDIA Nemotron safety

Meta has again delayed the release of the API for its new Muse Spark model, leaving developers without a launch date. The company says testing continues with early partners while it aims to compete with OpenAI and other rivals.

Microsoft now shows average token usage on model release cards, letting users compare performance against the cost of achieving that intelligence. The new dual‑benchmark forces AI providers to compete on efficiency as well as raw capability, targeting enterprise buyers who care about dollars per outcome.

OpenAI upgraded GPT‑Rosalind, merging GPT‑3.5's agentic coding tools with deeper expertise in medicinal chemistry, genomics, and experimental workflows. The model now outperforms prior versions on LifeSciBench, a new benchmark covering evidence handling, design, reasoning, and validation across life‑science tasks. It is available in research preview via trusted‑access deployment.

NVIDIA's Nemotron 3.5 Content Safety model adds multimodal evaluation, 12‑language coverage and custom policy enforcement, letting enterprises tailor safety rules for text, images, and responses. It also offers optional reasoning traces (THINK mode) for auditable decisions, aiding compliance across global AI deployments.

At CVPR 2026 NVIDIA Research presented three new foundation models built on billions of simulated samples. GraspGen‑X enables zero‑shot grasp planning for any robot gripper, LCDrive uses compact latent representations to speed autonomous‑vehicle reasoning on embedded hardware, and NitroGen scales embodied‑agent training across massive virtual environments.

The authors experimentally measure how architectural symmetry (equivariance) reduces sample complexity. On a controlled C_n‑symmetric task they find a symmetry‑data exchange rate β≈1.28, close to the theoretical value of 1, and show that wrong‑group constraints hurt performance while augmentation with orbit averaging matches equivariant models.

Google researchers present a two‑stage 'Sleep' framework for large language models that distills short‑term in‑context knowledge into stable long‑term parameters via upward distillation and a reinforcement‑learning‑driven 'Dreaming' phase. Experiments show improved continual‑learning, knowledge incorporation, and few‑shot generalization.

Direct Preference Optimization (DPO) applied as a second‑stage training after supervised fine‑tuning slashes text‑degeneration in OCR models, cutting average failure loops by 59% and up to 87.6% across model families. The approach uses the model's own rejected outputs as binary preference signals, proving DPO works beyond chatbot alignment.

China's flagship AI startup DeepSeek is raising about 50 billion yuan ($7.4 billion) in its first external funding round, led by Tencent and CATL. The capital could value the company at $52‑$59 billion, marking one of China’s biggest private tech financings and signaling a shift from its previous self‑funded model.

Morgan Stanley will open its ShareWorks and Equity Edge stock‑plan platforms to autonomous AI agents from thousands of corporate clients, letting these tools pull data and execute workflows without human login. The rollout targets its 3,400 administration clients by next year, aiming to scale wealth‑management services and cut internal costs.

OpenAI’s new ‘Dreaming’ feature gives ChatGPT a scalable memory layer that retains user preferences, projects and constraints across sessions, reducing staleness and improving relevance. The rollout begins with Plus and Pro users in the US and will expand to free users worldwide in the coming weeks.

OpenAI released a strategic action plan that leverages its new GPT‑Rosalind model to strengthen biological security. The initiative aims to equip trusted developers with advanced AI tools for early pathogen detection, rapid countermeasure development, and coordinated pandemic response, while establishing safeguards and governance.

OpenAI’s Sam Altman, Anthropic’s Dario Amodei, DeepMind’s Demis Hassabis and other AI leaders signed a public letter urging Congress to require companies that sell synthetic DNA and RNA to screen customers and orders. They argue AI lowers barriers to designing biological weapons, so mandatory screening is needed to protect biosecurity.

Magistrate Judge Maritza Braswell reports a sharp rise in pro se filings drafted with AI, with a study showing AI‑generated pleadings growing from 1% to 18% of cases. While AI helps articulate arguments, judges worry about hallucinations and the broader legal duties of chatbots.

Hugging Face revamped its hf command‑line tool to detect and adapt to coding agents like Claude Code and Codex. In agent mode the CLI streams full, non‑colored output, eliminating prompts and reducing token consumption by up to six‑fold on multi‑step tasks, making the Hub more efficient for automated workflows.