AI & design briefing for Thursday, July 16, 2026

tldr

• Thinking Machines, OpenAI, and Weco each shipped milestones advancing open-weights multimodal AI, automated red-teaming, and recursive self-improvement.
• PrismML crammed a 27B model onto a phone, Apple's lawsuit against OpenAI intensified, and a Suno hack exposed how the music AI scraped training data.
• Grok Build went open-source, Nvidia released a 34-language embedding model, and NameThatUI helps you prompt coding agents with precise UI component names.

Top Stories

• Thinking Machines releases Inkling. Mira Murati's lab shipped its first open-weight multimodal model handling text, audio, and images natively. 975B params (41B active), 1M-token context, controllable reasoning effort, Apache 2.0. ^[1]
• OpenAI's GPT-Red automates red-teaming. A self-play AI that attacks OpenAI's own models found 84% of attack scenarios vs 13% for humans, cutting prompt injection failures 6x on GPT-5.6. ^[2]
• OpenAI's hardware push: Codex Micro + screenless speaker. A $230 Work Louder keyboard with agent-control keys shipped now; Bloomberg reports a screenless, movable ChatGPT speaker with cameras and sensors is coming next. ^{[3] [4]}
• Weco claims first evidence of recursive self-improvement. An agent (AIDE²) redesigned its own research process over 8 days, beating a version humans tuned for 2 years on all benchmarks. ^[5]

Deeper Cuts

• PrismML Bonsai 27B fits on a phone. A 27B model compressed to 3.9GB (1-bit) retains 90% of full-precision performance, runs at ~10.8 tok/s on iPhone. Apache 2.0 with llama.cpp and MLX support. ^[6]
• Ant trains 1T-parameter reasoner with zero human labels. Ring-2.5-1T-Zero learns to reason without any human annotations. ^[7]
• Apple vs OpenAI lawsuit gets messy. Apple's lawyer reportedly mixed up two OpenAI employees while pressing trade-secret theft claims over hardware. ^[8]
• Anthropic launches Ode with Blackstone. A new enterprise AI services firm embedding Claude engineers into large enterprises. ^[9]
• Suno hack reveals training-data scraping. A breach exposed source code showing the AI music generator scraped songs from YouTube, Genius, and Deezer. ^[10]

Tools and Launches

• Grok Build. xAI/SpaceXAI open-sourced its coding-agent harness and CLI, deleting all retained coding data. ^[11]
• Nvidia Nemotron-3-Embed-1B-NVFP4. Quantized 1B embedding model supporting 34 languages for RAG. ^[12]
• HyperFrames. Open-source tool that generates videos from HTML, CSS, and JS. ^[13]
• Oak. Identity-control plane for managing humans, apps, and AI agents; raised $60M. ^[14]
• NameThatUI. Identify any UI element, learn its real name and API symbol, and copy a precise prompt for your coding agent. ^[15]
• RoboTTT. Test-time training for robot policies, stretching context to 8K timesteps without added inference latency. ^[16]

Prompt of the day