AI Daily | Google Tests Orbital TPUs, Meta Muse Open-Sources Hardware SDK, and Supabase Hits $10.65B on Agentic Growth
Model Releases & Updates
FrogNano-4B-2609 — Microsoft — Microsoft
- Bottom line：Microsoft has released FrogNano-4B-2609, a 4B parameter agentic model derived from Qwen3.5-4B optimized for repository-scale software engineering.
- Hybrid Architecture：Inherits Qwen3.5-4B’s 32-layer hybrid Gated DeltaNet and gated-attention architecture with post-training focused strictly on repository navigation and debugging.
- Reinforcement Learning Recipe：Trained using the five-tool Leaf harness across approximately 1,500 synthetic SWE environments generated via TaskPilot with executable test-based trajectory rewards.
- Local Execution Target：Engineered specifically to deliver long-horizon codebase navigation, code editing, and patch execution on consumer-grade local hardware.
- Source：Hugging Face Model
- Source：LocalLLaMA Thread
AstaBrief 8B — Allen Institute for AI — Allen Institute for AI (Ai2)
- Bottom line：Ai2 has open-sourced AstaBrief 8B, a dedicated scientific report generation model that turns research inquiries and retrieved literature into structured cited reports.
- Base & Architecture：Built upon Qwen3-8B and specialized in synthesizing complex multi-document scientific literature into coherent cited summaries.
- Production Deployment：Serves as the default Fast Mode engine for the Generate a Report feature within Ai2’s Asta research platform.
- Open Artifacts：Both the model weights and complete post-training datasets have been released for public download.
- Source：Ai2 Blog Post
- Source：Hugging Face Release
Qwen3.8-27B-Humanlike-Chat 2.0 — Community Release — LocalLLaMA Community
- Bottom line：Community developers released version 2.0 of Qwen3.8-27B-Humanlike-Chat, incorporating working tool execution and strong instruction following into its natural texting style.
- Tone Alignment：Removes rigid assistant formatting in favor of realistic human texting cadence while maintaining flexibility with user-defined character cards.
- Tool Calling Support：Replaces earlier non-functional tool implementations with full agentic tool dispatch capabilities across multi-turn workflows.
- Source：LocalLLaMA Discussion
Product Releases & Updates
Finances in ChatGPT — OpenAI — OpenAI
- Bottom line：OpenAI has rolled out ChatGPT Finances, enabling users to securely link bank accounts and credit data for conversational money management and portfolio analysis.
- Data Integrations：Integrates with Plaid and Experian to retrieve real-time account balances, credit ratings, and transaction histories.
- Automated Insights：Detects forgotten recurring subscriptions, identifies anomalous charges, tracks bill increases, and generates personalized budgets based on real spending.
- Voice Advisory：Allows users to discuss financial scenarios like career changes or analyze multi-account investment concentrations interactively using ChatGPT Voice.
- Source：
- Source：ChatGPT Finances Portal
Muse Gadgets & Muse Home Link — Meta — Meta
- Bottom line：Meta open-sourced its Muse Gadgets SDK and unveiled the Muse Home Link, allowing developers to build custom physical peripherals powered by the Muse AI agent.
- Open Source SDK：Provides Apache 2.0-licensed ESP32 firmware and a Linux SDK to interface displays, buttons, and environmental sensors with Muse.
- Hardware Hub：Manufactured a pilot run of 5,000 Muse Home Link devices that execute community skills such as room lighting automation, television control, and remote printing.
- Source：The Verge Article
- Source：Meta Research Post
DGX Spark 64GB Unified Memory Configuration — NVIDIA — NVIDIA
- Bottom line：NVIDIA announced a 64GB unified memory configuration for DGX Spark, enabling desktop and edge systems to run models up to 100 billion parameters locally starting at $4,999.
- Hardware Specs：Delivers 64GB unified memory architecture to execute large language model inference and fine-tuning on local desktop stations.
- OEM Availability：Slated to ship on October 23 through manufacturers including Acer, ASUS, Dell, Gigabyte, HP, and MSI.
- Source：NVIDIA Blog
Spanner Queues GA — Google Cloud — Google Cloud
- Bottom line：Google Cloud has made Spanner Queues generally available, introducing native transactional messaging directly inside Spanner for reliable agent execution.
- Atomic Agent Operations：Ensures that database state changes and downstream action queue dispatch commit together atomically, eliminating split-state execution errors.
- Agentic Workflows：Engineered for multi-step agent actions such as automated refund processing, inventory reconciliation, and sub-agent task routing.
- Source：Google Cloud Announcement
Web Search API via AI Gateway — Cloudflare — Cloudflare
- Bottom line：Cloudflare has introduced a native Web Search API within AI Gateway to provide real-time web grounding for AI agents.
- Launch Partners：Partnered with search and extraction providers Ceramic.ai, Exa, and Linkup to deliver structured search index results directly to agent harnesses.
- Eliminates URL Guessing：Replaces blind URL guessing and broken curl tool calls with grounded web search endpoints directly accessible through the gateway.
- Source：Cloudflare Blog
Industry News
Supabase Raises $150M at $10.65B Valuation Driven by Agentic Adoption — Supabase — Supabase / Y Combinator
- Bottom line：Supabase closed a $150 million funding round valuing the backend platform at $10.65 billion, propelled by rapid growth in AI agent database creation.
- Developer Scale：Serves over 13 million developers, adding more than 1 million active users and 4 million new databases every month.
- Agent Creation Share：Reports that 70% of all new databases created on the platform are now provisioned directly by autonomous agents and AI development tools.
- Source：New York Times Coverage
- Source：
Sean Parker Rebuilds Stability AI Around Professional Music Tools — Stability AI — TechCrunch
- Bottom line：Executive Chairman Sean Parker and CEO Prem Akkaraju are steering Stability AI into an AI music technology company backed by major music labels.
- Record Label Investment：Secured $76 million in backing from Sony Music, Warner Music Group, and Universal Music Group following a prior $80 million rescue financing.
- Licensed Training Data：All three major recording labels have formally licensed catalog access to train Stability AI’s next-generation professional audio models.
- Source：TechCrunch Article
ElevenLabs Attains FedRAMP 20x Class A Certification — ElevenLabs — ElevenLabs
- Bottom line：ElevenLabs has achieved FedRAMP 20x Class A certification, allowing US federal agencies to procure and deploy its conversational agents and voice synthesis APIs.
- Certified Offerings：Covers ElevenAgents, Text to Speech, and Speech to Text APIs operating in Zero Retention Mode with strict US data residency.
- Federal Marketplace：Officially listed on the FedRAMP Marketplace to streamline enterprise security and compliance reviews for government organizations.
- Source：ElevenLabs Blog
- Source：
NYC Council Introduces Mandatory AI Model Validation Bill — NYC Council — New York City Council
- Bottom line：NYC Council Speaker Julie Menin has introduced first-in-the-nation municipal legislation mandating independent third-party audits of frontier AI models.
- Independent Validation Mandate：Requires third-party validation of foundation model safety and risk profiles rather than relying exclusively on company self-auditing.
- Hearing & Subpoena：OpenAI, Anthropic, Google, and Meta have agreed to testify at the upcoming council hearing, while a subpoena was issued to SpaceX/xAI to compel participation.
- Source：
Research Papers
Toward Provably Private Learning from Federated Data via TEEs — Google Research — Google Research
- Bottom line：Google deployed a next-generation federated learning system powered by Trusted Execution Environments to train Gboard prediction models with verifiable privacy.
- Auditable Privacy：Publishes data access policies to Rekor public transparency logs, ensuring the execution binary is fully verifiable and built reproducibly from open-source code.
- Production Performance：Integrated into Gboard for English and Japanese next-word prediction, slashing model training turnaround from 1-2 months down to days while improving accuracy.
- Source：Google Research Blog
Solving Open Mathematical Research Problems with Muse Spark — Meta AI Research — Meta AI Research
- Bottom line：Meta researchers released six papers co-authored with Muse Spark 1.1 and 1.2 Thinking Mode, resolving open problems across multiple branches of pure mathematics.
- Disciplinary Scope：Directly answered previously unsolved questions in non-associative algebra, probability theory, differential equations, optimization, arithmetic physics, and group theory.
- Standard Interface：Formulated solutions within the standard meta.ai chat interface without bespoke agent scaffolding, followed by independent human mathematician peer review.
- Source：Meta Research Publication
Multi-Harness Reinforcement Learning for Coding Agents — Hugging Face & Liquid AI — Hugging Face
- Bottom line：Hugging Face published an open methodology for multi-harness reinforcement learning, enabling coding models to train across divergent agent frameworks simultaneously.
- Unified Proxy Design：Intercepts agent-model communication via an intelligent proxy supporting OpenAI, Anthropic, and Gemini message formats without altering underlying harness source code.
- Empirical Gains：Trained across four harnesses concurrently, Liquid AI’s LFM2.5-2.6B improved benchmark performance from 42% to 54% while reducing total tool calls by 31%.
- Source：
Project Suncatcher: Orbital TPU Satellite Prototype Launched — Google AI — Google AI
- Bottom line：Google launched a prototype satellite carrying TPU hardware into orbit on SpaceX Transporter-18 to explore space-hosted machine learning compute clusters.
- Hardware Validation：Measures TPU performance, physical stress, and radiation resilience in low Earth orbit in partnership with satellite builder Planet.
- Solar Energy Density：Leverages continuous orbital sunlight to capture up to 8 times more solar power than terrestrial data centers for machine learning workloads.
- Source：
Other Highlights
Geoffrey Hinton and CASP Warn of Accelerating Intelligence Explosion Risks — Centre for the Analysis of Strategic Policy (CASP)
- Bottom line：Geoffrey Hinton and CASP published an analysis warning that automated AI R&D and recursive self-improvement could trigger an intelligence explosion sooner than expected.
- Recursive Acceleration：Argues that self-improving AI research loops may rapidly outpace conventional governance and monitoring checkpoints.
- Research Call：Urges the AI research community to develop auditable chain-of-thought verification and robust containment standards for agentic swarms.
- Source：CASP Report
- Source：
Underdog AI Launches Private Personal AI Stack Backed by a16z — Underdog AI / a16z
- Bottom line：Underdog AI announced venture backing led by a16z to develop a decentralized personal AI system that runs locally on user-owned devices.
- Local Intelligence Stack：Enables users and enterprises to retain complete data sovereignty by storing memory, skills, and model execution directly on local hardware.
- Syndicate Backing：Backed by a16z, Khosla Ventures, and angel investors from Anthropic, Vercel, and Hugging Face.
- Source：a16z Announcement
- Source：
