AI Daily — May 5, 2026
2026-05-05
openai.com
How OpenAI delivers low-latency voice AI at scale
OpenAI published a technical deep-dive on how it rebuilt its WebRTC stack to support real-time voice AI with low end-to-end latency and global scale. The post covers architectural decisions around turn-taking, connection management, and infrastructure needed to handle production voice workloads. This is a useful engineering reference for teams building on the Realtime API.
www.technologyreview.com
Week one of the Musk v. Altman trial: What it was like in the room
The Musk v. OpenAI trial entered its first week in Oakland, with Elon Musk alleging his early donations were misused as OpenAI converted to a for-profit structure. The proceedings involve direct testimony from Sam Altman and surface core disputes about OpenAI's founding mission and governance. The trial's outcome could have meaningful implications for how AI non-profits and mission-driven labs are legally structured.
rss.arxiv.org
Model Organisms Are Leaky: Perplexity Differencing Often Reveals Finetuning Objectives
Researchers demonstrate that a simple perplexity-gap method — comparing token-level perplexity between a reference model and a finetuned model on diverse random-prefix completions — can reliably surface the finetuning objective of so-called model organisms used in safety research. The technique exploits overgeneralization: finetuned models bleed their trained behaviors into off-distribution contexts, making the objective detectable without access to training data. This has direct implications for evaluating how well model organisms actually isolate specific behaviors for controlled safety experimentation.