1. NVIDIA Releases Nemotron 3 Super 120B Hybrid Model
NVIDIA released Nemotron 3 Super, a 120B parameter hybrid model combining state-space models and transformers with only 12B active parameters. It features a 1M-token context window and delivers 5x higher throughput than previous versions, specifically optimized for high-volume multi-agent systems.
2. Google Launches Multimodal Gemini Embedding 2
Google launched Gemini Embedding 2, a multimodal model that unifies text, images, video, and audio into a single numerical space. It supports Matryoshka Representation Learning for customizable output dimensions and is designed to optimize RAG pipelines for production-grade retrieval.
3. Amazon Wins Injunction Against Perplexity Shopping Agent
A U.S. judge issued a preliminary injunction barring Perplexity's Comet browser from making purchases on Amazon on behalf of users. The ruling requires Perplexity to stop using its agents to bypass Amazon's site restrictions and to destroy copies of scraped data, citing risks to customer data and advertising.
4. METR Analysis Finds Half of SWE-bench PRs Unmergeable
Research from METR indicates that approximately half of the PRs that pass the SWE-bench Verified benchmark would not be merged by maintainers in real-world scenarios. The study suggests that current benchmarks may overstate agent readiness by failing to account for the human-like iteration and feedback loops required in production.
5. Perplexity Debuts 'Personal Computer' AI Operating System
Perplexity introduced 'Personal Computer,' an AI-native operating system proxy that runs locally on a Mac mini. The system provides an always-on agent with access to local files and apps, allowing it to execute objectives autonomously across sessions while requiring user approval for sensitive actions.
6. Microsoft Releases bitnet.cpp for 1-Bit LLM Inference
Microsoft released bitnet.cpp, an official inference framework for 1-bit LLMs like BitNet b1.58. The framework enables lossless inference on standard CPUs, achieving speedups of up to 5x on ARM architectures and making large-scale models more accessible for local deployment.
7. Claude Code Updates: Auto-Memory and Cross-App Context
Anthropic updated Claude Code with auto-memory via persistent markdown files and expanded context sharing across Microsoft Excel and PowerPoint. The tool also introduced a multi-agent 'Code Review' system for automated pull request analysis and bug hunting.
8. Google Research: Agents Learn Cooperation via Diverse Opponents
Google researchers found that LLM agents can learn to cooperate in multi-agent systems when trained against a diverse pool of unpredictable opponents. This approach avoids hardcoded coordination rules, offering a more scalable and computationally efficient blueprint for enterprise agent deployments.
9. Industrial Robotics Funding: Mind Robotics and Rhoda AI
Mind Robotics, a Rivian spin-out, raised $500M to develop industrial AI robots, while Rhoda AI secured $450M for models trained on public internet videos. Both startups aim to deploy advanced autonomous systems in manufacturing and logistics environments to automate complex physical tasks.
10. ElevenLabs Launches ElevenCreative Multimodal Platform
ElevenLabs launched ElevenCreative, a browser-based platform for generating and localizing audio and video. The system integrates voice cloning, text-to-speech, and AI video generation with support for over 70 languages in a single unified interface.
11. Fish Audio S2-Pro Achieves Sub-150ms TTS Latency
Fish Audio released S2-Pro, a Large Audio Model (LAM) capable of expressive speech synthesis with sub-150ms latency. The model supports zero-shot voice cloning and granular emotion tagging, representing a shift toward integrated audio architectures.
12. Context-Aware Permission Guard for Claude Code Released
The 'nah' tool was released as a context-aware permission guard for Claude Code, moving beyond simple allow-or-deny tool permissions. It allows developers to define granular rules for sensitive actions like file deletions or git checkouts to prevent autonomous agents from causing catastrophic system changes.
13. AI Networking Infrastructure: Nexthop AI and Eridu
Nexthop AI raised $500M for specialized switches designed to reduce power consumption and latency in hyperscale data centers. Simultaneously, Eridu emerged from stealth with $200M to build high-performance AI networking equipment for large-scale GPU clusters.
14. LLMfit Utility Ranks Models by Hardware Compatibility
LLMfit is a new utility that scans local hardware to rank LLMs based on their compatibility with specific CPU, RAM, and GPU configurations. The tool helps developers select the most efficient models for on-device execution by analyzing memory and compute constraints.
15. Dify Secures $30M for Open-Source Agentic Workflow Platform
Dify raised $30M for its open-source platform designed to build and operate agentic workflows. The platform provides a structured environment for deploying AI applications with integrated memory, tool management, and security controls.
16. Expo Agent Compiles Native Apps from Natural Language
Expo Agent is a new tool that generates and compiles native iOS and Android applications from natural language prompts. It supports React Native, SwiftUI, and Jetpack Compose, allowing for browser-based mobile app deployment and testing.
17. NotebookLM-py Provides Python API for Research Tool
NotebookLM-py provides a Python API and CLI for Google's NotebookLM, enabling programmatic access to features not exposed in the web interface. This allows researchers to integrate NotebookLM's summarization and retrieval capabilities into automated data processing pipelines.
18. Anthropic Launches Research Institute for AI Policy
Anthropic co-founder Jack Clark is leading a new research institute to study the impact of AI on economies, law, and governance. The institute will have direct access to frontier model data to inform policy and regulatory discussions regarding AI's societal role.
19. WiFi-DensePose Reconstructs Body Position via WiFi Signals
Researchers developed WiFi-DensePose, a system that uses standard WiFi signals to reconstruct full-body positions in real time. The technology maps body segments through walls without cameras or wearables by analyzing channel state information already present in standard hardware.
20. Temporal API Implementation Fixes JavaScript Date Handling
Bloomberg engineers detailed a nine-year effort to implement the Temporal API in JavaScript to fix long-standing issues with the Date object. The new API provides a more robust and type-safe way to handle dates, times, and time zones, improving reliability for global software systems.
21. Mozilla Outlines WebAssembly First-Class Language Roadmap
Mozilla outlined plans to make WebAssembly a first-class language on the web, expanding its capabilities for high-level languages. Recent updates include support for shared memories, SIMD, and exception handling to improve performance for complex, compute-heavy web applications.
22. Replit Hits $9B Valuation in AI Infrastructure Boom
Replit reached a $9B valuation following a $400M funding round, targeting $1B in ARR by the end of the year. The company continues to expand its AI-powered development environment, positioning itself as a primary platform for collaborative and agentic coding.
23. Oracle Reports 44% Cloud Revenue Growth Driven by AI
Oracle reported a 44% surge in cloud revenue to $8.9B, driven by demand for AI infrastructure. The company's total Q3 revenue reached $17.19B, reflecting its growing role in hosting large-scale AI workloads and providing the compute required for frontier model training.
24. Zendesk Acquires Forethought for Agentic Customer Service
Zendesk acquired Forethought, an agentic customer service startup, to bolster its AI-driven support capabilities. The acquisition integrates Forethought's autonomous resolution technology into Zendesk's service platform to automate complex customer interactions.
25. WordPress Launches my.WordPress.net Private Workspace
WordPress launched my.WordPress.net, a browser-based private workspace that allows users to create sites without hosting or signing up. The service is designed as a personal environment for writing, research, and integrating AI tools directly within the browser.