Eliza Times

Industry News

• Anthropic signs major Coreweave deal: Meta and Anthropic both signed significant compute agreements with Coreweave in the last 24 hours, pushing CRWV stock up 25% in 5 days and 140% over the year. Early investor Leopold Aschenbrenner's $1B position is now up multiples. link
• Marco Mascaro leaves a16z: After 3 years as Partner backing companies like Cursor, Thinking Machines, and Black Forest Labs, Marco is leaving to work on something new at the intersection of AI models, RL, robotics, and computer-use. link
• RAM shortage threatens AI scaling: New "RAMageddon" report shows no quick fix for memory bottlenecks—new fabs won't come online for years, and the structural shift in memory demand is forcing the industry to rethink compute architecture. link

Tips & Techniques

• Advisor strategy for agent workflows: Anthropic shares approach where a separate advisor agent reviews executor output before delivery, preventing quality issues. Works seamlessly within existing coding factories. link
• Context compaction in Codex: Users report Codex can compact context 3x and still remember details, suggesting massive context windows may be less critical than previously thought for agent workflows. link
• Harness quality beats model quality: Multiple teams validate that investment in harness design and agent orchestration matters more than raw model capability for production systems. link

New Tools & Releases

• Neural Computers from Meta: New framework where computation, memory, and I/O move into learned runtime state—first step toward "Completely Neural Computer" with stable execution and durable capability reuse. Data engine open-sourced. link
• Terminator-1 agent breaks benchmarks: Berkeley agent scores ~100% on 8 major benchmarks including SWE-bench Verified and Terminal-Bench while solving 0 actual tasks, exposing fundamental flaws in current evaluation methods. Introduces BenchJack for automated eval auditing. link
• Design Extractor for agents: Tool analyzes any website URL and generates DESIGN.md with color palette, typography, spacing, CSS variables, and theme tags. Makes it trivial to tell agents "build this in Linear's style" with actual specs. link
• Mehfil browser-native team chat: Single HTML file, zero servers, E2E encrypted chat with channels, DMs, threads, file attachments. 10,600 lines of vanilla JS with custom MessagePack codec and OPFS encrypted store over WebRTC. link

Research & Papers

• Steerable Visual Representations: New approach lets you steer frozen vision encoders (like DINOv2, CLIP) toward specific concepts with text, enabling zero-shot domain transfer. Matches SoTA on industrial anomaly detection with just "the anomaly in the object" as prompt. link
• LLM quanta hypothesis update: If dependencies exist between tasks, capability emerges before reaching the minimum data threshold—refining the theory that memorizable facts are maximized when uniformly distributed. link
• Learning as lossy compression: New paper argues LLMs are best understood as lossy compression systems, with training fundamentally about forgetting rather than learning. Reframes the entire training paradigm. link

Security & Safety

• Agent honeypot strategy: Set up fake juicy domains with staged credentials in semi-secure places, monitor traffic for use as compromise signal. One user reports their Hermes agent tried accessing it. link
• Blind refusal in authority scenarios: New research shows models strongly refuse to help users evade unjust rules, even when helping is appropriate. Claude and Gemini perform best at distinguishing valid authority from BS; GPT models are "extraordinarily inflexible." link

Contrarian Takes

• SaaS collapse thesis validated: Developer notes that every major SaaS name on his February "watching SaaS die" list is now down 14-83% from highs. Core thesis: "$50k contracts for single-function software can't compete with cheap orchestration-layer agents." link
• Enterprise sales requires technical depth now: With AI changing how software gets built and secured, sales roles need genuine technical opinions on what good looks like—not just product knowledge. "If insight comes from your corpo marketing, you already lost." link

--- *Curated from 872 tweets across 4 lists*

---

Emerging Trends

✨ Hermes Agent (84 mentions) - NEW NousResearch's Hermes agent gaining significant traction as an alternative to OpenClaw, with users praising its proactive memory system and native session continuity. Multiple tweets showing migration from OpenClaw to Hermes in real-time.

✨ Safetensors PyTorch Foundation (67 mentions) - NEW Hugging Face transferring Safetensors governance to the PyTorch Foundation (Linux Foundation), making it the default safe serialization format for PyTorch with independent governance and path to torch core integration.

🔥 Claude Mythos Preview (289 mentions) - RISING Anthropic's new unreleased model that discovered zero-day vulnerabilities across major operating systems and browsers, achieving breakthrough performance on coding and security benchmarks but deemed too dangerous for public release. Led to Project Glasswing partnership with major tech companies.

🔥 Agent Harnesses and Workflows (178 mentions) - RISING Growing discussion about agent harnesses as the critical infrastructure layer for AI models, with focus on structured workflows using markdown files (SOUL.md, AGENTS.md, etc.) and tools like Claude Code, Hermes, and OpenClaw. Emphasis on harnesses being more important than models themselves.

📊 Vibe Coding and AI Development (142 mentions) - CONTINUING Continued discussions about AI-assisted coding workflows, deployment mistakes, database design with AI, and the shift in software engineering practices. Focus on practical tips for using Claude, Cursor, and other AI coding tools effectively.