Scale AI Debuts Muse Spark 1.2: A Multi-Tasking Multimodal Agent
Scale AI has announced the release of Muse Spark 1.2, a powerful multimodal model designed for complex agentic workflows. The model excels in visual coding, robotics planning, and audio-visual understanding. Its architecture is optimized to integrate with various agentic tools, allowing it to move beyond simple chat interactions into functional physical and digital planning.
Evaluation data shows that Spark 1.2 performs strongly across diverse benchmarks, particularly in tasks requiring cross-modal reasoning. Scale AI CEO Alexandr Wang noted that the model's ability to ground visual information into executable plans marks a significant step forward for the Muse ecosystem.
The 'Post-Training Scaling Law': Z.ai CEO Discusses GLM 5.3
Jie Tang, CEO of Z.ai (Zhipu AI), has introduced GLM 5.3 alongside a new perspective on model development: the 'Post-training Scaling Law.' The discussion suggests that the industry is moving away from purely increasing parameter counts ('Death of Params') and toward more intensive post-training optimization and reasoning-focused scaling. This shift reflects a broader trend where compute is increasingly allocated to inference-time reasoning and sophisticated fine-tuning rather than just raw pre-training scale.
OpenAI Introduces Zero Data Retention for Frontier Models
OpenAI has announced a major update to its data privacy policies, offering Zero Data Retention (ZDR) for eligible API customers using its frontier models. This move aims to appease enterprise clients with strict regulatory requirements by ensuring that data sent through the API is not stored on OpenAI's servers after processing. Additionally, the company is previewing Private Safety Processing, a new feature intended to maintain safety standards without compromising user data privacy.
Asana Leverages LLMs to Complete Years-Long Migration in Two Weeks
Asana has successfully utilized AI to complete a massive testing framework migration that was previously expected to take years. By using LLMs to handle the heavy lifting of refactoring and code conversion, the team was able to compress the timeline into just two weeks. This case study highlights the growing trend of 'AI-led migrations,' where large-scale technical debt and infrastructure transitions are automated, potentially disrupting the traditional relevance of legacy consultancy and research firms like Gartner.
OmniScientist Framework Proposes End-to-End AI Scientific Discovery
A new research paper introduces OmniScientist, an 'omni-modal' AI agent designed to perform multidisciplinary scientific research directly from heterogeneous raw evidence. Unlike existing models that rely on processed text, OmniScientist uses lifecycle-wide perception to interpret diverse scientific modalities. The framework aims to automate the discovery process by grounding its findings in raw experimental data across various scientific disciplines, significantly improving the accuracy and depth of evidence-grounded research.
Zetta ζ: Self-Evolving Closed-Loop Harness for Robotics
Researchers have developed Zetta, a closed-loop embodied harness that allows physical agents to self-evolve through online runtime critics and recovery skills. By governing physical execution at high action frequencies, Zetta enables robots to scale self-exploration and improve their success rates on standard benchmarks. The system features faster inference times and a robust mechanism for recovering from execution failures, addressing a critical bottleneck in the deployment of autonomous physical intelligence.
On-Device 125M Transformer Enables Real-Time Piano Autocomplete
An independent developer has released a 125M-parameter transformer model capable of autocompleting piano performances in real-time on consumer hardware. Operating at approximately 108 notes per second on an iPhone 15, the model functions similarly to GitHub Copilot but for MIDI input. This project demonstrates the increasing feasibility of highly specialized, low-latency generative models running entirely on-device, providing a unique tool for musicians to expand on their improvisations.
Meta AI Desktop App Gains Traction with Seamless Dictation Features
Meta's new AI desktop application is receiving positive early feedback, specifically for its integrated dictation capabilities. Users are reporting that the dictation tool provides a 'game-changing' UX for interacting with LLMs locally on desktop environments. This release signifies Meta's continued push to integrate its Llama models into daily productivity workflows, competing directly with the desktop experiences offered by OpenAI and Perplexity.
SemaPLC: Verification-Gated Agent Harness for Industrial Code
SemaPLC introduces an agent harness specifically for the generation of Programmable Logic Controller (PLC) code, a staple of industrial automation. The system uses a 'verification-gated' approach, where generated logic is validated through external compilation and live runtime execution before being finalized. This method achieves significantly higher verified pass rates than standard LLM generation, potentially bringing reliable AI assistance to specialized manufacturing and engineering sectors.
SPADE: Scaling Reasoning via Self-Play in Adaptive Environments
SPADE is a new reinforcement learning framework where a language model is tasked with designing its own adaptive, executable training environments. By using a regret-based targeting mechanism, the model learns to solve the very environments it creates, leading to improved performance in reasoning and tool-use tasks. This research emphasizes the value of synthetic environments and self-play for scaling AI capabilities without relying on static human-labeled datasets.