Eliza Times

Industry News

• Anthropic designated "supply chain risk" by Pentagon: After refusing to remove safety guardrails for mass surveillance and autonomous weapons, Anthropic faces government contract cancellation while OpenAI accepts similar terms. link
• OpenAI-Pentagon deal includes different terms than Anthropic's: Despite initial reports of identical contracts, fine print reveals OpenAI's agreement differs on weapons autonomy and surveillance scope. link
• Block lays off 40% citing AI: Mass layoffs hit fintech as companies accelerate AI replacement of entry-level roles. link

Tips & Techniques

• AGENTS.md files don't scale beyond modest codebases: New paper shows three-tier memory architecture (hot constitution, domain experts, cold knowledge base) enabled 108k-line C# system across 283 sessions with 6 autonomous turns per human prompt. link
• Mistral OCR + batch is unbeatable for invoice extraction: Simple combination delivers industrial-grade document processing. link
• How to properly set up Claude Cowork: Detailed guide on context files, global instructions, and plugin configuration for getting AI to work like a true coworker while you're away. link

New Tools & Releases

• Claude Code /batch and /simplify commands: New parallel agent features enable simultaneous PRs and auto code cleanup across entire codebases. link
• Graph-RAG framework for multi-hop reasoning: New architecture enables complex reasoning chains across knowledge graphs. link
• Perplexity releases 4 open-weights multilingual embedding models: State-of-the-art retrieval models designed for production use. link

Research & Papers

• Codified Context paper: Documents production three-tier agent memory system that enabled 108k-line codebase development with knowledge-to-code ratio of 24.2%. link
• SymTorch enables symbolic analysis of trained models: New framework reveals what deep learning models actually learned, not just their outputs. link
• Self-Sustaining Rewards (SSR) improves without human data: Models achieve +10.4 on SWE-bench through self-play, consistently outperforming human-data baseline despite evaluation on natural language issues absent from training. link

Contrarian Takes

• "AI slop" blames the wrong thing: The problem isn't AI output quality—it's that humans prompt for garbage, train on garbage, then blame the tool for producing what was requested. link
• Karpathy: CLIs are the killer app for AI agents: Legacy command-line tools become powerful when agents can natively combine them; building for agents means CLI-first design. link
• Anthropic's stance may attract top talent: Despite government pressure, taking a principled position on autonomous weapons could become a major recruiting advantage. link

--- *Curated from 500+ tweets across 15 tech professional lists*

---

Emerging Trends

✨ Codex Desktop App Launch (142 mentions) - NEW OpenAI's Codex Desktop app is receiving significant praise as a powerful coding interface, with many developers switching from terminal-based workflows and comparing it favorably to Claude Code.

✨ Iran Conflict Escalation (89 mentions) - NEW Major geopolitical crisis as Iran launched missile strikes on UAE and other Middle East targets, with US military operations intensifying and concerns about impacts on AI infrastructure in the region.

🔥 OpenAI $110B Funding Round (195 mentions) - RISING OpenAI announced a $110 billion funding round from Amazon, NVIDIA, and SoftBank, garnering massive attention and discussion about AI company valuations and the future of the industry.

🔥 Anthropic DoD Dispute (168 mentions) - RISING Anthropic's conflict with the Department of Defense over AI use in autonomous weapons and surveillance escalated, with the Pentagon designating them as a supply chain risk and threatening use of the Defense Production Act.

📊 Nano Banana 2 Release (67 mentions) - CONTINUING Google's Nano Banana 2 image generation model launched with comparisons to Grok Imagine and other models, though mixed reviews on quality and world knowledge compared to competitors.