Eliza Times

Tips & Techniques

• Agent organizational structure matters: Approaching agents as organizations with delegation, coordination, and decision rights yields better results than treating them as simple executors. High-ability work is expensive, hand-offs have costs, and strategic delegation is key. link
• LLM Wiki workflow for research: Karpathy's approach lets you skip writing but not reading/thinking. Process documents by reading them, their summaries, and the LLM's analysis of how they fit into your knowledge base—flexible and effective for personal research. link
• Security review prompt for client deliveries: Before every delivery, drop this comprehensive 10-point security checklist into Cursor with @codebase to catch exposed secrets, missing auth, SQL injection, CORS issues, and more. Non-negotiable pre-delivery ritual. link
• Overtraining models for test-time scaling: New T² (Train-to-Test) scaling research shows that radical overtraining combined with repeated sampling at inference time can outperform Chinchilla-optimal models. Both task-specific and general scaling approaches converge on the same recommendation. link

Industry News

• Anthropic limits intensify: Claude Code usage limits have become significantly more restrictive, with users reporting faster token exhaustion even on Max subscriptions. The crackdown on third-party harnesses continues to erode developer goodwill. link
• Gemma 4 ecosystem launch: Google's Gemma 4 release demonstrates unprecedented ecosystem coordination, with pre-launch collaboration across HuggingFace, vLLM, llama.cpp, Ollama, NVIDIA, and others. Partners enabled same-day integration across the stack. link
• OpenAI policy proposal: OpenAI published "Industrial Policy for the Intelligence Age" proposing public wealth funds, robot taxes replacing payroll taxes, 4-day work weeks, "Right to AI" access, and automated safety nets triggered by economic thresholds. link
• Goldman Sachs on AI job displacement: Analysis suggests AI substitution has already reduced monthly payrolls, with agents stealing jobs from humans at an accelerating rate. The impact is becoming quantifiable in economic data. link

New Tools & Releases

• Morphic Workflows: Launch of 72 pre-built workflows for filmmaking, social media, animation, and marketing. No prompting required—select assets and options, minimal nodes. Makes repeatable creative tasks instantly reusable. link
• dstack 0.20.16 performance boost: Major control plane update delivers 2-10x faster GPU provisioning. Provisioning 200 instances drops from 12 to 4 minutes, starting 200-node tasks from 25+ to 4 minutes. link
• ORBIT dataset for search agents: 20K reasoning-intensive web training dataset generated without paid APIs. Small (<4B) search agents trained with ORBIT outperform others by up to 9.0 EM accuracy on Wikipedia QA tasks. link
• Devin daily e2e automation: Using Devin to run daily automated tests that sign up, complete onboarding, test chat functionality and tools, then send morning reports with video. Provides reassuring confirmation of core functionality. link

Research & Papers

• Self-training for LLM reasoning (SePT): Model samples its own responses at low temperature, learns via standard SFT, and repeats—no reward, no verifier. Qwen2.5-Math-7B improves mean Pass@1 from 22.7 to 39.5 across 6 math benchmarks. link
• METR study: AI makes developers slower: Experienced developers working on mature codebases were 19% slower when using AI coding tools. Not faster—slower. Evidence of a vibe ceiling that most hit earlier than expected. link
• 25,000-task agent coordination experiment: MIPT ran the largest AI agent coordination study ever with 8 models and 4-256 agents. Results demonstrate scaling behavior and coordination bottlenecks in multi-agent systems. link
• High-dimensional Gaussian concentration: In high dimensions, Gaussian density concentrates along a thin spherical shell—the center is essentially empty. Counter-intuitive behavior that matters for understanding embedding spaces. link
• IV Co-Scientist for causal discovery: LLM framework that acts as thought partner for discovering instrumental variables in causal relationships. Successfully proposes novel, plausible IVs across socio-economic datasets that are statistically meaningful. link

--- *Curated from 400+ tweets across 5 lists*

---

Emerging Trends

✨ Morphic Workflows Launch (18 mentions) - NEW Morphic launched "Workflows" - a feature that allows users to capture and reuse repeatable creative tasks without starting from scratch or complex prompting. 72 workflows available at launch covering filmmaking, social media, animation, fashion, and marketing.

🔥 Karpathy's LLM Wiki (156 mentions) - RISING Andrej Karpathy shared a detailed approach to using LLMs to build and maintain personal knowledge bases as markdown wikis in Obsidian. The system includes data ingest, Q&A, output generation, and automated health checks. This evolved from his earlier "Idea Files" concept and generated massive engagement.

🔥 Vibe Coding Security Issues (41 mentions) - RISING Multiple detailed threads about common security vulnerabilities in AI-generated code, particularly around authentication, JWT storage, session management, and deployment practices. Emphasis on security audits before production deployment of vibe-coded applications.

📊 Gemma 4 Release (198 mentions) - CONTINUING Google released Gemma 4, the #3 open model in the world, with significant discussion about running it locally with OpenClaw and other frameworks. Strong ecosystem collaboration with HuggingFace, VLLM, Ollama, and others. Many quantized versions and device integrations (Android Studio, iPhone) announced.

📊 Anthropic Claude Harness Restrictions (87 mentions) - CONTINUING Continued controversy over Anthropic's restrictions on third-party harnesses and API usage. Discussion of session limits, token consumption issues, and developers feeling frustrated with usage policies. Some developers reporting workarounds or switching to alternative providers.