AI digest: Speech models and code tooling hit production
Google ships real-time translation across 70 languages whilst developers get proper GPU programming tools and video game generation.
Real-time speech translation finally works properly, and the tooling for building AI systems is getting serious.
Google’s Gemini 3.5 Live Translate goes properly multilingual
Google shipped streaming speech-to-speech translation across 70+ languages through Meet, Translate, and the Live API. The model stays just a few seconds behind the speaker whilst generating continuous audio output. This feels like the first proper implementation of real-time translation that might actually work for business calls.
Anthropic’s Fable 5 generates playable games instantly
Anthropic’s new Claude Fable 5 can build working video games from simple prompts. It’s the first public version of their Mythos-class model, though with guardrails blocking high-risk outputs around cybersecurity and biology. The gaming angle is clever marketing, but the real story is how capable these models are becoming at generating functional code.
NVIDIA’s cuTile brings proper GPU programming to Python
NVIDIA released cuTile Python, a tile-based interface for writing CUDA-style kernels directly in Python. The tutorial shows vector addition, matrix operations, and proper benchmarking against PyTorch. This matters because it makes GPU programming accessible to the Python crowd without dropping into C++.
Lovable claims $500M revenue from AI code generation
Lovable says it’s hit $500M annualised revenue with users creating a million new projects weekly. Their AI builds complete applications and internal tools. The numbers seem ambitious, but if even half true, it shows demand for proper code generation tools is massive.