AI digest: model wars, government pressure, and agents everywhere
The White House slows OpenAI's newest model, Gemini gets computer control, an open-source coding model learns its own RL scaffolds, and video games are now serious AI training data.
A busy Thursday. Governments are getting hands-on with model releases, agents are eating more infrastructure investment, and the open-source coding scene just got a serious new entry.
White House tells OpenAI to hold back GPT-5.6
The Trump administration reportedly asked OpenAI to limit the release of its newest model to a small group of partners rather than the general public, citing safety concerns. TechCrunch has the story. This is notable because it is the same administration that gutted the previous AI safety executive order. The fact that they are now pumping the brakes on a model release suggests something in GPT-5.6 genuinely spooked someone.
Google bakes computer control into Gemini 3.5 Flash
Google has put “Computer Use” directly into Gemini 3.5 Flash, meaning the model can now see and operate screens, browsers, and mobile devices natively through the API. It scores 78.4 on OSWorld, which puts it level with GPT-5.5. Full details at The Decoder. This matters for agent builders who want screen control without bolting on a separate tool.
Ornith-1.0 learns its own RL scaffolds
DeepReinforce dropped Ornith-1.0, an open-source coding model family built on Gemma 4 and Qwen 3.5. The interesting bit is that instead of using a fixed reinforcement learning harness, the model learns its own scaffold during training. The 397B flagship hits 82.4 on SWE-Bench Verified and everything ships under MIT. MarkTechPost covers it here. That self-scaffolding approach is worth watching closely.
$2.3B bet on video games as AI training ground
General Intuition raised $320 million as part of a $2.3B thesis that millions of hours of gameplay can train agents to develop something closer to real-world intuition. TechCrunch has the breakdown. Using action-dense game data to ground agents is a smart idea in theory. Whether it transfers to messy real-world tasks is the question that $2.3B is trying to answer.