Eliza Times

Industry News

• MiniMax M2.5 open-source model released: Chinese AI lab drops 230B MoE model (10B active) trained with agentic RL, achieving SOTA on coding (80.2% SWE-Bench Verified) at ~20x cheaper pricing than US alternatives ($0.30/M input vs competitors). link
• Anthropic closes $30B round at $380B valuation: Massive funding confirms enterprise demand for Claude, though some question sustainability given compute costs and competitive pressure from cheaper open-source alternatives. link
• US removes Chinese military company list: Pentagon document listing firms linked to China military quietly removed, potentially signaling shift in AI/tech policy approach. link

Tips & Techniques

• Ask your agent "Are you missing any context?": Simple meta-prompt that dramatically improves agent performance by forcing self-reflection on information gaps before proceeding. link
• Codex-Spark runs on Cerebras at 1,000 TPS: OpenAI's new Codex variant achieves unprecedented inference speed by leveraging specialized hardware, enabling real-time multi-agent workflows. link
• GPT-5.x "almost never hallucinates": Major reliability improvement credited to advances in grounding and verification, making models production-ready for high-stakes applications like legal work. link

New Tools & Releases

• OBLITERATUS removes model guardrails in minutes: SVD-based tool surgically extracts refusal behavior from open-weight models without retraining, highlighting durability issues in safety alignment. link
• ColGrep saves 15.7% tokens + 70% better answers: Grep replacement optimized for AI coding agents delivers measurable improvements in agentic workflows, now integrated into Claude Code and Codex. link
• Recursive Language Models (RLMs) handle 10M+ tokens: DeepMind paper introduces architecture managing infinite memory through recursive processing, solving long-horizon reasoning bottleneck. link

Research & Papers

• Origin of life potentially solved via self-replicating RNA: Science paper reports 45-token (base) self-replicating program small enough to emerge spontaneously, major abiogenesis breakthrough. link
• Human-centered coding agents position paper needed: Research team argues current agent benchmarks overoptimize for solo autonomy while ignoring human collaboration patterns that matter in practice. link
• Vision tokens are interpretable with right tools: New analysis shows VLM internal representations can be decoded when using proper visualization methods, advancing mechanistic interpretability. link
• Isomorphic Labs achieves 2-3x protein structure accuracy: Technical report demonstrates drug design engine breakthrough in predicting biomolecular structures, accelerating pharmaceutical discovery. link

Deployment & Production

• OpenClaw passes VSCode in GitHub stars: Community-driven agent framework hits major adoption milestone, now 2x PyTorch and 3x Claude Code in popularity. link
• Unitree robots now manufacturing robots in factories: Embodied AI model UnifoLM-X1-0 enables humanoids to build other robots, marking shift from demo to actual production deployment. link
• Enterprise AI adoption survey: 57% want open-source over custom: C-level execs at $100M+ companies prefer flexibility and avoiding vendor lock-in, with construction and finance leading adoption. link

--- *Curated from 500+ tweets across AI research, tools, and industry lists*

---

Emerging Trends

✨ MiniMax M2.5 Model Release (68 mentions) - NEW Launch of MiniMax M2.5 model being discussed as competitive with Opus 4.6 while being 20x cheaper, with benchmarks showing strong performance and local deployment capabilities on consumer hardware.

✨ GLM-5 Model Release (52 mentions) - NEW Announcement and early testing of GLM-5, a 745B parameter MoE model from Zai platform, with discussions of it being new open-source SOTA for agentic coding with 200K context and potential to run on multiple Mac Ultras.

🔥 Codex 5.3 Adoption (145 mentions) - RISING Widespread discussion of developers switching to and using OpenAI's Codex 5.3 as their primary coding tool, with mentions of heavy usage (750M tokens/week), comparisons to Claude Code, and integration across teams.

🔥 OpenClaw Security & Setup Issues (89 mentions) - RISING Growing concerns about OpenClaw security vulnerabilities with third-party setup services exposing gateways and root directories, plus discussions of proper deployment, authentication, and VPS alternatives to Mac Mini.

📊 Seedance 2.0 Video Generation (142 mentions) - CONTINUING Viral discussion of Seedance 2.0 video generation model with "Will Smith eating spaghetti" recreations and demonstrations showing dramatic improvements in AI video quality, trained on TikTok content.