Eliza Times

Industry News

• Microsoft launches MAI model family: MAI-Transcribe-1 achieves #1 on FLEURS benchmark across 25 languages, MAI-Voice-1 for natural speech, and MAI-Image-2 in top 3 on Arena—all shipped in months by the Microsoft AI team. link

• Alibaba releases Qwen3.6-Plus: New flagship model with 1M context window, stronger agentic coding capabilities, repository-level task handling, and improved multimodal reasoning—pushing open model frontier toward adaptive compute. link

• GPU shortage intensifies: All 8x H200/B200 configurations sold out across Prime Intellect, RunPod, Lium, Verda, Lambda—rental prices up 40% according to SemiAnalysis H100 index. link

• x402 Foundation launches: Coinbase, Solana, Circle, Amazon, Amex, Cloudflare, Figma, Visa join as founding members to shape command-line commerce standards. link

Tips & Techniques

• Set up AI for daily reports: Configure your AI to create a morning report of what it needs from you, referencing previous reports for historical context, and send it via email for easy review. link

• Treat 200-day SSL certs as quantum prep: Shorter TLS certificate cycles (dropping to 100 days, then 47 by 2029) force automation and crypto-agility—exactly the infrastructure needed for post-quantum migration. link

• Avoid eval leakage in auto-research loops: If agents can see the evals they're optimizing for, they'll overfit and rationalize it—keep hold-out sets, review changes manually, and limit updatable components per turn. link

New Tools & Releases

• Chief of Staff skill for Claude Code: Turns Claude into an executive assistant that manages email, tracks to-dos across sessions, gathers intelligence from Slack/email, and delivers formatted briefs automatically using persistent plugin storage. link

• Runable hits $2M ARR in 3 weeks: One of the fastest companies to reach $2M ARR after launching version 2.0, demonstrating explosive growth in AI dev tools. link

• Mastra adds metrics and logs: New observability features in Studio with DuckDB backend for fast local analytics, ClickHouse support coming next. link

• Flowith Canvas introduces cowork skills: Agent collaboration workspace enabling shared context and coordinated task execution between AI agents. link

• MLX + TurboQuant enables local superpowers: Pre-fill 256k KV cache with local documents or codebase for private, high-performance local AI workflows. link

Research & Papers

• SAEs fail out-of-distribution due to linear inaccessibility: Features in superposition are linearly representable but not linearly accessible—requires per-sample sparse solvers like FISTA instead of learned encoders. link

• Diff3r: Feed-forward 3D Gaussian Splatting with optimization layer: Predicts optimization-ready 3DGS initialization end-to-end using implicit differentiation and Gauss-Newton approximation for fast, stable results. link

• AEC-Bench: First multimodal agent benchmark for construction: 196 tasks across real construction workflows to evaluate AI agents in architecture/engineering. link

• Point tracks as tokens for world models: Forecasting Motion in the Wild proposes using point track forecasting instead of pixels or 3D for generating complex non-rigid motion and animal behavior. link

--- *Curated from 600+ tweets across AI research and development communities*

---

Emerging Trends

✨ Microsoft MAI Model Family Launch (85 mentions) - NEW Microsoft released MAI-Transcribe-1 (most accurate transcription across 25 languages), MAI-Voice-1 (natural speech generation), and MAI-Image-2 through Foundry, marking a major multi-model release from Microsoft AI.

🔥 Claude Code Source Leak (320 mentions) - RISING Claude Code's entire source code was accidentally leaked via a map file in their npm registry, leading to widespread discussion, analysis, and even removal of safety systems. The leak has spawned numerous forks and reverse engineering efforts.

🔥 Agent-First GTM Strategy (58 mentions) - RISING Discussion about AI agents making purchasing decisions autonomously, with companies like Resend and Stripe becoming default choices for coding agents. Focus on API-first products, clean documentation, and machine-readable content as the new distribution channel.

📊 Hermes Agent Launch (42 mentions) - CONTINUING Hermes Agent continues gaining traction as an alternative to OpenClaw, with users praising its ability to handle multiple agent profiles and automate tasks like model monitoring and benchmarking.

📊 Pazi AI Team Launch (28 mentions) - CONTINUING Pazi launched as an AI team operating 24/7 in Slack and other platforms, powered by OpenClaw, executing across entire tech stacks including Sentry, Linear, and GitHub with one-click integration.