XRlabs raises $8.4M for AI surgical platform

Compact native audio-language model from inclusionAI that takes audio plus text instructions and supports speech recognition, spoken question answering, audio understanding and audio reasoning.

Built by Gal at doubleAI, Maxwell sits between code and hardware and makes programs run faster as they run without changing a single line. It is described as a "Code Execution Model" that sees programs from the inside, with no humans in the loop.

TorchNeuron lets PyTorch apps run on Amazon's Trainium chip by changing one line of configuration, no code rewrite required. AWS says ~85% of AI apps are written with PyTorch, the toolkit Nvidia's CUDA has long been embedded in by default.

A Hacker News thread asks how to handle AI-generated code that ships without the author understanding it, citing a LinkedIn post from a Meta engineer who says the job now feels like a black box.

Princeton's Sayash Kapoor, Arvind Narayanan and co-authors synthesized trends across evaluation science, research agenda and ecosystem with Stanford, the UK AI Security Institute and Brookings. They argue evaluators must scrutinize full agent trajectories, not just outcomes, since agents can take shortcuts with tool access.

428B-parameter MoE with ~23B active and a 1M-token context, priced at $0.15/$0.60 per 1M input/output tokens. MiniMax says inference is nearly 3x faster and usage costs drop 70%; weights arrive Oct 23.