Eliza Times

Industry News

• Elon Musk announces TERAFAB: SpaceX, Tesla, and xAI are building the largest chip manufacturing facility in Austin for robotics, AI, and space data centers. Vertical integration strategy mirrors SpaceX's approach. link
• Palantir wins $10B Pentagon AI contract: Maven AI is now permanent infrastructure across all 5 US military branches for analyzing drone feeds, satellite imagery, and prioritizing targets. Came after Anthropic refused to remove ethical guardrails. link
• MiniMax M2.7 open weights in ~2 weeks: Updated version shows noticeably better performance on OpenClaw. May be the best model you can realistically run at home. link

Tips & Techniques

• Ask agents if they're missing context: Simple prompt "Are you missing any context?" significantly improves agent outputs when they're not doing what you want. link
• Claude supervising Codex for overnight runs: Use Claude in a /loop to take screenshots every 15 mins and restart Codex if it gets stuck, enabling multi-hour autonomous sessions. link
• Smart model routing saves $150/month: Use Qwen 3.5 (local) for simple tasks, Kimi K2 for automation, Grok for X research, Claude Sonnet for writing, and Opus only for critical decisions. link

New Tools & Releases

• The Elves: 12-20hr autonomous coding system open-sourced: Includes automated testing, review loops, and support for open-ended goals like "make this faster" or "conduct QA testing." link
• Shadify: Generative UI with ShadCN: Describe a UI and watch agent compose from ShadCN components on the fly using LangChain and CopilotKit, then export as React code. link
• Kitten-TTS-WebGPU: Browser TTS faster than real-time using pure WebGPU compute shaders. Runs 7.3x faster than realtime on M4 Pro, 4.8x on iPhone 17 Pro Max. link
• Animation Studio for ThreeJS: Full animation graph editor with blendtrees, state machines, root motion, and Codex chat integration. Free and open source, built in a day with vibe coding. link
• MiniMax official skills repo open-sourced: Covers frontend, fullstack, mobile dev, plus Office suite (PPTX, PDF, DOCX, XLSX). Competes with Anthropic and OpenAI skill collections. link

Research & Papers

• AI can learn scientific "taste": Small model trained on citations predicts which preprints will become highly cited better than humans, suggesting scientific judgment isn't uniquely human. link
• Supermemory hits ~99% on LongMemEval: Multi-stage agentic retrieval with structured extraction and temporal tracking beats single-shot baselines. Memory might be "solved" for certain use cases. link
• RLVR with implicit rewards: Paper shows RL with verifiable rewards implicitly incentivizes correct reasoning in base LLMs, even without explicit reasoning supervision. link
• Visual guide to modern LLM attention: Comprehensive reference covering all major attention variants in one place for easy comparison. link

Community Highlights

• Nigerian medical guidelines dataset released: Most comprehensive collection of treatment guidelines covering 270 medical conditions, useful for building health RAG pipelines. link
• AI hiring discourse is broken: The real story isn't replacement but that teams using AI well ship more, creating new bottlenecks in different places than people argue about. link

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

---

Emerging Trends

✨ MiniMax M2.7 Open Weights (45 mentions) - NEW MiniMax announces M2.7 open weights release coming in ~2 weeks, with significantly improved performance on OpenClaw and positioning as potentially the best model that can run at home. Multiple tweets show excitement about running it locally.

🔥 Cursor Composer 2 (125 mentions) - RISING Cursor releases Composer 2 with claims of Opus-level performance. However, controversy emerges around allegations that it's a fine-tuned version of Kimi K2.5 without proper attribution, potentially breaching licensing terms requiring credit for commercial use.

🔥 AlphaProof Nature Paper (68 mentions) - RISING Google DeepMind's AlphaProof paper published in Nature this week, describing technical innovations in the RL loop bridging natural language and symbolic rigor that enabled silver-medal performance on IMO problems in 2024.

🔥 LongCat-Flash-Prover (22 mentions) - RISING Meituan releases LongCat-Flash-Prover, an open-source formal reasoning model achieving 97.1% on MiniF2F-Test and 41.5% on PutnamBench. Features hybrid-experts iteration framework and HisPO algorithm for tool-integrated reasoning in theorem proving.

📊 OpenClaw/Claude Code Skills (180 mentions) - CONTINUING Continued discussion and sharing of Agent Skills for OpenClaw and Claude Code, with focus on office productivity (pptx, pdf, docx, xlsx) and development workflows. MiniMax also releases their own skill collection. Skills being integrated into various platforms.