Eliza Times

Industry News

• Cursor Composer 2 uses unlicensed Kimi K2.5: Cursor's new Composer model allegedly fine-tunes Moonshot's Kimi K2.5 without proper attribution, potentially violating its MIT license clause requiring credit for products over 100M MAU or $20M monthly revenue. link
• White House releases AI Policy Framework: Administration proposes federal preemption of state AI regulation, protecting children online, and preventing censorship while aiming for American AI dominance. link
• Apple blocks vibe coding apps from App Store: Apple has halted updates for AI coding apps like Replit and others that execute runtime code, citing security concerns around dynamic code evaluation. link

Tips & Techniques

• Ask agents if they're missing context: When an agent isn't performing as expected, prompt it with "Are you missing any context?" to surface gaps in its understanding. link
• Engineers should burn $250K in tokens annually: If your $500K engineer isn't spending at least $250K on AI inference, they're not leveraging compute effectively enough. link
• Claude Code channels enable MCP control: New feature allows controlling Claude Code sessions through MCP servers, starting with Slack integration for remote session management. link

New Tools & Releases

• Nemotron-Cascade 2 from NVIDIA: 30B MoE model with 3B active parameters achieves state-of-the-art performance, 21.8x better scaling than autoregressive models in compute-optimal settings. link
• LongCat-Flash-Prover for theorem proving: Open-source formal reasoning model achieves 97.1% on MiniF2F-Test with 72 attempts and 41.5% on PutnamBench, using hybrid-experts iteration and HisPO algorithm. link
• Aeon v2 launches async agents: Updated framework lets you schedule and import any skill for agents running asynchronously on GitHub, using your Claude Code subscription with automatic secret management. link

Research & Papers

• AlphaProof paper published in Nature: Google DeepMind's formal reasoning system achieved silver-medal standard at IMO 2024, with technical details on the RL loop bridging natural language and symbolic rigor. link
• V-JEPA 2.1 unlocks dense video features: Yann LeCun's team modifies V-JEPA recipe to learn both global semantics and dense local features, advancing video self-supervised learning. link
• Delightful Policy Gradient fixes RL waste: Vanilla policy gradient performs worse than cross-entropy on MNIST bandits because it wastes gradient budget; new approach shows gains that grow with problem complexity. link
• AI coding tools impair conceptual understanding: Anthropic research proves AI assistance degrades developers' ability to understand code structure and make independent decisions, not just a prompting skill issue. link

Datasets & Infrastructure

• Parquet Content-Defined Chunking merged: Arrow-rs now optimizes storage on content-addressable stores like Hugging Face Xet by chunking Parquet pages based on content similarity. link
• Real-world barebone sandbox costs matter: Building next-gen compute infrastructure requires millions of parallel sandboxes on bare metal, not cloud, to maintain healthy margins at scale. link

--- *Curated from 2,847 tweets across AI research and industry*

---

Emerging Trends

✨ Cursor Composer 2 (45 mentions) - NEW Cursor released Composer 2, a new coding agent model that combines API models with domain-specific models. However, controversy emerged over allegations it may be a fine-tuned version of Moonshot's Kimi K2.5 without proper attribution, potentially violating licensing terms.

✨ LongCat-Flash-Prover (32 mentions) - NEW Meituan released LongCat-Flash-Prover, an open-source formal reasoning model achieving 97.1% on MiniF2F-Test and 41.5% on PutnamBench. It introduces hybrid-experts iteration framework and HisPO algorithm for stabilizing long-horizon tool-integrated reasoning in theorem proving.

🔥 OpenAI Acquires Astral (78 mentions) - RISING OpenAI announced Astral (creators of uv Python package manager) will join the Codex team to build end-to-end development agents. This follows their acquisition strategy and aims to strengthen Codex across the full software development lifecycle while keeping Astral's tools open source.

🔥 Foundation-1 Music Model (95 mentions) - RISING RoyalCities released Foundation-1, a state-of-the-art text-to-sample model specifically built for music production workflows. The model runs locally, requires ~7GB VRAM, and is completely free, marking a significant advance in AI music generation tools.

📊 AlphaProof Nature Paper (28 mentions) - CONTINUING Google DeepMind's AlphaProof paper was published in Nature, detailing the technical innovations including the RL loop bridging natural language and symbolic rigor that enabled AlphaProof and AlphaGeometry to achieve silver-medal standard on IMO problems in 2024.