[AINews] not much happened today
Summary
The AINews brief for July 13-14, 2026, highlights significant advancements across AI. OpenAI's agent products, including Codex + ChatGPT Work and GPT-5.6 Sol, are experiencing "insane" demand, with usage growing 2.5x in a week and Codex usage jumping 1M in one day to 6M active users. The industry is also focusing on harness quality and observability for agents, with tools like LangChain adding tracing for various code agents. In open models, aggressive compression techniques like Ternary Bonsai 27B (5.9 GB / 1.71 effective bits) and 1-bit Bonsai 27B (3.9 GB / 1.125 effective bits) are enabling frontier-adjacent models on consumer devices and single GPUs. Multimodal systems are evolving towards real-time continuous perception (MOSS-VL-Realtime) and active evidence search for long videos (OmniAgent). New benchmarks like Perplexity's WANDR (500-task) are emerging for agentic research, alongside progress in physical AI, such as Sakana AI's self-repairing "Smart Cellular Bricks" and autonomous micro-drones.
Key takeaway
For AI Scientists and Machine Learning Engineers evaluating deployment strategies, you should prioritize robust agent harness development and explore aggressive quantization techniques. OpenAI's agent products show immense demand, but model quality alone is insufficient; focus on observability and evaluation environments. Consider 1-bit or NVFP4 quantization to deploy frontier-adjacent models like Bonsai 27B or Hy3 on consumer devices or single GPUs, expanding your local inference capabilities for agentic workflows.
Key insights
AI is rapidly advancing across agentic systems, local inference, multimodal perception, and physical autonomy.
Principles
- Agent harness quality is a key differentiator.
- Aggressive quantization enables local frontier models.
- Motion is a novel data type for world models.
Method
OmniAgent uses an Observation–Thought–Action loop to request only necessary frames/audio for long-video understanding, trained with 58K agentic trajectories and entropy-weighted RL via TAURA.
In practice
- Implement tracing for agent tool calls and subagents.
- Explore 1-bit or NVFP4 quantization for local deployment.
- Design evaluations for agent degradation over sequential tasks.
Topics
- AI Agents
- Model Quantization
- Local Inference
- Multimodal AI
- World Models
- AI Benchmarking
- Physical AI
Best for: CTO, VP of Engineering/Data, Executive, AI Scientist, Machine Learning Engineer, Director of AI/ML
Related on AIssential
See Counsel's argued verdicts on the open AI decisions leaders are weighing →
Editorial summary, takeaway, and curation by AIssential. Original article published by Latent.Space - Www.latent.space.