OpenAI Codex, Runway Agent 2.0, and Midjourney Updates Highlight AI Progress

Key Takeaways

1. OpenAI's Codex agent is being used internally for complex, long-term, cross-functional tasks, heralding a shift to AI-driven work automation.
2. Runway launches Agent 2.0 for automated marketing campaign generation from simple prompts.
3. Midjourney previews V8.2 aesthetics and speeds up draft mode by 24x; Hugging Face's TRL v1.7.0 boosts training speed by 1.25x and saves 16GB VRAM.
4. NVIDIA enables real-world AI impact: Revolut improves fraud detection accuracy by 2.3x, PYLER reduces unsafe brand content by 76.7%.
5. Open source models gain traction: Gemma 4 downloads surpass 200M in 2.5 months, and Yann LeCun notes enterprises demand on-premise models.

1. AI Agents and Workflow Automation

1. OpenAI revealed that its internal teams are using Codex agents for more complex, long-duration, and cross-functional tasks, signaling how agentic tools could reshape future work patterns. — via 1
2. Runway introduced Agent 2.0, which can generate complete marketing briefs and campaign assets from a simple prompt, analyze performance data for optimization, and expand across platforms, formats, and markets. The company aims to make it the most powerful autonomous agent. — via 1
3. Ethan Mollick reported an experiment where Opus 4.7 completed an end-to-end coding task that would take a human engineer 2–17 weeks in just 14 hours, at a cost of $251, demonstrating rapid AI coding capability progress. — via 1
4. Hamel Husain received a Laude grant for his DocWriter project, which aims to build a new writing tool and UI that supports asynchronous human-AI collaboration, guides long-running agents, and open-sources continuous evaluation frameworks. — via 1

2. Model Performance and New Features

1. Midjourney added a --preview parameter to preview V8.2 aesthetics and personalization effects, and updated its big-batch draft mode with --sref random support, accelerating style space exploration by 24x. — via 1
2. Hugging Face released TRL v1.7.0, featuring continuous batching acceleration that improves GRPO and RLOO training by 1.25x while saving 16GB VRAM, plus support for MoE post-training and a new GMPO trainer. Separately, the LFM2.5 230M model achieved 1,400 tok/s inference speed in the browser via agentic WebGPU optimization. — via 1 2
3. Demis Hassabis announced that the Gemma 4 model surpassed 200 million downloads in just 2.5 months, with community growth accelerating. — via 1
4. NVIDIA AI unveiled AdaVoMP, a method that generates physically realistic interactive virtual worlds from a few photos, supporting a spatial resolution of 1024³ and making objects interactive and deformable. — via 1

3. Enterprise AI and Open Source Dynamics

1. NVIDIA showcased two enterprise AI success stories: Revolut built a trading foundation model on NVIDIA's full stack, achieving 2.3x improvement in credit risk accuracy and up to 5x training throughput gain; PYLER used NVIDIA-accelerated AI to analyze 2 million videos daily, reducing brand unsafe content exposure by 76.7% and cutting operational costs by 10x. — via 1 2
2. Yann LeCun observed that in recent weeks, large enterprises are increasingly demanding on-premise model deployments and post-training based on GLM-5.2, indicating that the open-source model paradigm is winning. — via 1
3. Ethan Mollick highlighted concerns that the U.S. government might effectively ban open-weight models by prohibiting enterprises from using unapproved models, sparking debate on AI safety regulation and the future of open source. — via 1
4. NVIDIA, in collaboration with the Linux Foundation and industry partners, launched the Akrites cybersecurity project, aiming to protect critical open-source software from AI-powered cyber threats through open collaboration and standardized disclosure. — via 1
5. swyx highlighted that a company (Pim de Witte's) accidentally built a perfect world model data collection business, raising $320M in Series A at a $2.3B valuation, transforming attention economy into attention industry. — via 1

4. Industry Milestones and Infrastructure

1. Elon Musk reported that SpaceX completed 76 orbital launches in 2026, all successful, surpassing the rest of the world combined, and also performed a full-duration static fire test of a single engine on Starship, demonstrating industrial-grade operations and a perfect record. — via 1 2
2. Starlink provided free internet to earthquake-affected areas in Venezuela and rapidly deployed terminals. — via 1
3. The AI Engineer World's Fair 2026 will feature a 4x larger exhibition area, a music corner for the first time, and tickets are nearly sold out. — via 1