GPT-Bidi-1 and a $2B AI lab without a product

OpenAI is testing a bidirectional audio model called GPT‑Bidi‑1 for ChatGPT Voice, letting the assistant listen and speak simultaneously and handle interruptions without freezing. The upgrade adds tiered intelligence modes so users can trade reasoning depth for lower latency, signaling a push to close the gap with its newer text models.

OpenAI is testing a bidirectional audio model called GPT‑Bidi‑1 for ChatGPT Voice, letting the assistant listen and speak simultaneously and handle interruptions without freezing. The upgrade adds tiered intelligence modes so users can trade reasoning depth for lower latency, signaling a push to close the gap with its newer text models.

Microsoft Research introduces Next-Latent Prediction (NextLat), an auxiliary loss that trains transformers to forecast their own next hidden state. This forces the model to build compact belief-state representations, yielding better generalization, higher downstream accuracy, and up to 3.3x faster decoding on language tasks.

OpenAI has added a Developer mode toggle that lets Codex use the Chrome DevTools Protocol. With full CDP access, Codex can profile JavaScript performance and inspect network, console, and page state directly from the in‑app browser. The opt‑in feature gives developers deeper, real‑time debugging capabilities for web projects.

Lin Junyang, the former Qwen lead at Alibaba, is raising a new AI lab at a roughly $2 billion post‑money valuation, with Gaorong Capital and Sequoia China in talks. The round, still unfolding, underscores how a single engineer’s open‑source success can command a valuation unheard of for a Chinese startup with no product yet.

Android 17 adds AppFunctions and the Android MCP protocol, letting on‑device agents like Gemini discover and execute app‑specific tools with local state. The new Jetpack library makes exposing these orchestratable functions as simple as annotating a class, opening a path for tighter AI‑app integration.