State Dept picks OpenAI. Plus: who verifies AI-written code? And new kernels surpass expert-written ones.
The US State Department yesterday announced it will phase out Anthropic in favor of OpenAI for internal use. This follows weeks of public friction between Anthropic and the Pentagon, signaling that government agencies are making concrete choices about their AI vendors. Meanwhile, a new analysis raises a critical question for builders: who verifies the software when AI is the author?
An investigation reveals data annotators in Kenya are exposed to sensitive personal videos and images from Meta's AI smart glasses, including intimate moments. This happens due to inadequate anonymization and automatic data processing mandatory for AI functionality. The report highlights a gap between Meta's marketing claims of user control and the reality of data sharing.
Read full story→A new paper introduces Evo 2, a large biological foundation model trained on 9 trillion DNA base pairs with a 1 million token context window. Evo 2 predicts genetic variation impacts and generates new genomic sequences across all domains of life, with model parameters and training dataset released open-source.
As AI speeds up software development, traditional code review and testing can't keep pace with AI-generated errors. Leo de Moura argues that formal verification, powered by AI and tools like Lean, is the answer, shifting verification from manual friction to mathematical proof.
Sean Goedecke argues that giving LLMs a 'personality' is a fundamental engineering requirement, not a marketing gimmick. He states base models are unmanageable, needing engineers to define a behavioral 'region' for usefulness and safety. This persona makes LLMs controllable tools.
This analysis argues the current AI boom is an 'information war' by companies like OpenAI and Anthropic. The analysis claims these firms use media leaks to inflate financial prospects while downplaying significant costs and unprofitability. It also criticizes their involvement with the US military, suggesting ethical stances are performative.
Unsloth now supports fine-tuning for the Qwen3.5 model family, covering both text and vision models. It trains 1.5x faster and uses 50% less VRAM than standard setups, with options to preserve reasoning ability and export to GGUF/vLLM.