today's digest looks at the hidden costs of model routing, a massive open-weight model release, and why voice ai still can't handle real conversations. we also cover an acquisition in real-time recommendations, lessons from building a maritime ai agent, and a study explaining how diffusion models create new data.

  1. model routing is harder than it looks - routing requests to the cheapest or most specialized model often backfires due to hidden costs, complexity, and latency factors.
  2. thinking machines releases open-weight inkling model - a 975b-parameter mixture-of-experts model that is open-weight and designed for enterprise customization.
  3. voice ai still struggles in real conversations - a new benchmark with over 1 million human ratings shows voice models often fail at listening, emotion, and real-world noise.
  4. whatnot acquires shaped for real-time shopping recommendations - livestream shopping app whatnot buys machine learning startup shaped to improve real-time product discovery.
  5. what building shippy taught us about building agents - lessons from building a maritime ai agent that prioritizes reliability, deterministic tools, and sandboxed hosting.
  6. why diffusion models create new data - a study shows diffusion models generate novel outputs because neural networks learn a smoothed score function, causing interpolation between training points.

that's all for today's ai news. check back tomorrow for more updates on model releases, research, and industry moves.