Eliza Times

Industry News

• Arden auto-applies flight credits in background: User forgot they signed up for Arden travel agent; it automatically imported flight, found savings, applied credits to United account without prompting - new bar for ambient agent UX link
• OpenAI Codex usage limits cut dramatically: Just as Codex gained hype, OpenAI appears to have reduced Plus tier usage limits, frustrating users who rely on it for development work link
• Rork raises $15M seed at $175M+ valuation: Top 5% of seed rounds now routinely exceeding $175M valuation, up 3x in past 12 months - sign of continued AI infrastructure premium link

Tips & Techniques

• Database schema design prompt prevents debt: Proper schema design prompt for agents prevents architectural technical debt before it accumulates - critical for vibe coding workflows link
• Advisor strategy for complex agent tasks: Spawn a roundtable of expert agents to discuss in multi-turn rounds for best solution - particularly effective in OpenClaw workflows link
• Local knowledge bases beat naive RAG: For personal research, a properly structured local knowledge base (like LLM Wiki) consistently outperforms naive RAG approaches link

New Tools & Releases

• Claude Code /ultraplan launches: New feature moves plan creation to web UI while keeping terminal free - enables collaborative plan editing before execution, better than traditional plan mode link
• Codex 0.119/0.120 ships major QoL updates: Massive quality of life improvements and bug fixes including memory extensions where you decide what Codex learns from past sessions link
• Minutes for conversation intelligence: Records meetings/calls locally with whisper.cpp, compiles knowledge base with person profiles, decision history, commitment tracking - 26 MCP tools let any agent query it link
• Hermes Manim skill launches: Create precise programmatic animations for mathematical visualization using Manim engine, now available as Hermes Agent skill link

Research & Papers

• Claude Mythos likely uses looped LLMs: Strong evidence Mythos implements "Scaling Latent Reasoning via Looped Language Models" from ByteDance - explains massive graph search advantage over competitors link
• Video2Humanoid improves motion retargeting: Neural Motion Retargeting learns mapping between human and robot motion as distribution via networks - eliminates joint discontinuities, reduces self-collisions 54%, violations 61% link
• Agentic-R trains retrievers for multi-turn search: First retriever training framework for agentic search systems measures both local relevance and global answer correctness - 3pt improvement, 10-15% fewer search turns link
• PruneRAG solves evidence forgetting: RAG systems exhibit 46-68% evidence forgetting rate even when retrieving correct documents - confidence-guided query trees reduce forgetting 20.8%, run 4.9x faster link

--- *Curated from 1,847 tweets across 15 lists*

---

Emerging Trends

🔥 Claude Mythos Preview (420 mentions) - RISING Anthropic's unreleased Claude Mythos model demonstrated superhuman cybersecurity capabilities, discovering zero-day vulnerabilities across major OSes and browsers, but won't be publicly released due to safety concerns. The model showed sophisticated strategic thinking and situational awareness.

🔥 Agent Harnesses and Workflows (210 mentions) - RISING Growing discussion around agent harnesses as infrastructure layer for AI, including frameworks like Hermes Agent migrating users from OpenClaw. Focus on 7-file harness systems (SOUL, AGENTS, USER, TOOLS, MEMORY, HEARTBEAT, STYLE) for building AI workflows and the concept that "harnesses > models."

🔥 Safetensors PyTorch Foundation (85 mentions) - RISING Announcement that Safetensors is joining the PyTorch Foundation (under Linux Foundation), moving from Hugging Face governance to independent foundation status for safer model serialization and potential torch core integration.

🔥 Gemma 4 Release (72 mentions) - RISING Google's Gemma 4 models gaining attention with strong benchmark performance, including 31B dense model, multimodal capabilities with variable aspect ratio support, and fine-tuning guides from Unsloth. Discussion of architecture features like global attention layers and p-RoPE.

📊 Vibe Coding and AI Development (165 mentions) - CONTINUING Continued discussion of "vibe coding" with AI tools like Claude Code and Cursor, including debates about dead code generation, architecture quality, and the shift from craft to shipping outcomes. Mix of enthusiasm and concerns about workflow changes and code quality.