Sunday, June 28, 2026
White House blocks GPT-5.6 release
Claude just surged 75% in paid subscribers while quietly eating ChatGPT's lunch, OpenAI's GPT-5.6 got blocked by the White House over security concerns (yikes), and researchers figured out how to delete an entire language from an LLM using just 4 tokens—wild. Meanwhile, MIT dropped Ornith-1.0, open-source coding models crushing SWE-Bench benchmarks. Would you let the government decide when your AI ships?
Top Stories
Anthropic's Claude has seen 75% growth in paying consumers since January 2026, significantly closing the gap with ChatGPT in the paid consumer market despite ChatGPT's continued overall dominance. The growth accelerated after Claude's refusal to support government mass surveillance and continues despite recent U.S. restrictions on its most advanced models.
Ornith-1.0 is a new MIT-licensed family of open-source coding models (9B-397B parameters) that achieves state-of-the-art performance on major coding benchmarks using a novel RL-based training approach that jointly optimizes scaffolds and solutions.
arXiv
Autodata uses AI agents as automated data scientists to create and iteratively improve synthetic training data through meta-optimization, demonstrating superior performance over classical methods across multiple reasoning domains.
Thread Reader App
A team removed a language model's German capability by fine-tuning just 4 tokens using parameter decomposition, demonstrating more precise and less destructive model editing than LoRA while preserving other languages. This represents a breakthrough in interpretable, targeted model modification.
The Information
The White House has asked OpenAI to delay the release of its latest AI model citing security concerns, marking an unprecedented government intervention in AI model deployment and signaling increased regulatory oversight of advanced AI systems.
Keep Reading
Enjoyed this issue?
Get daily AI intel delivered to your inbox. No fluff, just the stories that matter.