Eliza Times

Industry News

• NVIDIA eyes $20B OpenAI investment: NVIDIA nears deal to invest $20B in OpenAI as part of a $100B+ funding round, replacing their previous $100B long-term commitment. New capital will largely flow back to NVIDIA hardware. link
• Anthropic Claude Code Security preview: Anthropic launches limited research preview of Claude Code Security that scans codebases for vulnerabilities and suggests targeted fixes, triggering $21B wipeout in cybersecurity stocks. link
• Taalas ships 17k tokens/sec inference chip: Taalas launches specialized AI inference chip achieving 17,000 tokens per second on Llama-3.1-8B, claiming significant energy and cost savings over standard GPUs. link

Tips & Techniques

• 30-minute time caps for parallel agents: When running 20+ parallel agents, restricting each task to 30 minutes prevents scope creep and context-switching problems while maintaining high PR output. link
• Skills as one-time execution: Configure agents using markdown "skills" that execute once to set up infrastructure rather than traditional config files—agents fork the codebase into your desired setup via natural language. link
• "Claw" emerges as term of art: "Claw" is becoming the standard noun for OpenClaw-like agent systems that run on personal hardware, communicate via messaging protocols, and can both act on instructions and schedule tasks. link

New Tools & Releases

• Claude Code git worktree support: Anthropic ships built-in git worktree support for Claude Code, allowing agents to run in parallel without interfering with each other's work. link
• GPT-5.3-Codex-Spark 30% faster: OpenAI's Codex now serves at over 1,200 tokens per second after 30% speed improvement, closing the gap with Claude Code. link
• NanoClaw and alternatives emerge: Multiple lightweight OpenClaw alternatives launching including NanoClaw (~4000 lines, runs in containers by default) and others like nanobot, zeroclaw, ironclaw, picoclaw. link
• dmux for running Codex/Claude swarms: Open-sourced tool for running Claude Code and Codex swarms using tmux + worktrees, enabling parallel agent execution. link

Research & Papers

• METR time horizon misunderstanding: Common confusion about METR benchmarks—they measure task complexity that can be completed within a time limit, not how long a model can run continuously. link
• Bytedance frames reasoning as chemistry: New paper treats chain-of-thought as molecular structure with three bond types (deep reasoning=covalent, self-reflection=hydrogen, exploration=van der Waals) to improve RL training stability. link
• Self-distillation overview gains traction: Survey showing self-distillation as powerful technique for continual learning and model improvement without external teacher models. link

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

---

Emerging Trends

🔥 NanoClaw and Claw Alternatives (145 mentions) - RISING Developers are exploring lighter, more secure alternatives to OpenClaw like NanoClaw (4000 lines vs 400K), ZeroClaw, IronClaw, and PicoClaw, driven by security concerns about OpenClaw including RCE vulnerabilities, supply chain poisoning, and malicious skills in the registry.

🔥 OpenClaw Security Crisis (320 mentions) - RISING Major security concerns emerging around OpenClaw including exposed instances, RCE vulnerabilities, supply chain attacks, and malicious skills (1,184 malicious packages discovered, one attacker uploaded 677). Users describing it as a "security nightmare" with the codebase being "400K lines of vibe coded monster."

🔥 Gemini 3.1 Pro Release (185 mentions) - RISING Gemini 3.1 Pro Preview launched in AI Studio and Vertex AI with significant improvements, particularly for image-to-code tasks, 1M token context window, and described as having "unprecedented depth and nuance." Early users reporting strong performance on complex tasks.

🔥 Llama.cpp Team Joins Hugging Face (95 mentions) - RISING The llama.cpp team is joining Hugging Face in what's being described as a "match made in heaven" and "huge news," consolidating two major forces in open-source AI infrastructure and local model deployment.

📊 Seedance 2.0 Video Generation (82 mentions) - CONTINUING Developers building wrappers and integrations for Seedance 2.0 outside of China, with features including image, video, audio, and first-last frame inputs. Combined discussion with video upscaling tools like Magnific Video Upscaler from Freepik.